Determination of nucleic acid base modifications
The method uses single molecule real-time sequencing to detect base modifications in nucleic acids without chemical conversion, addressing the limitations of existing techniques by enhancing sensitivity and applicability to long sequences, facilitating diagnostic and research applications.
Patent Information
- Application Number
- JP2024043021
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2020-07-13
- Filing Date
- 2024-03-19
- Publication Date
- 2024-12-04
- Estimated Expiration
- 2040-08-17
AI Technical Summary
Current methods for determining base modifications in nucleic acids, such as DNA methylation, rely on chemical treatments like bisulfite conversion, which degrade DNA, introduce CG bias, and are limited to short sequences, necessitating a method that avoids these steps.
A method for detecting base modifications in nucleic acids using single molecule real-time sequencing without enzymatic or chemical conversion, utilizing kinetic features like interpulse duration, pulse width, and sequence context within a measurement window to train machine learning models for accurate methylation detection.
Enables accurate detection of various base modifications across long DNA molecules, improving sensitivity and avoiding degradation, with applications in research, diagnostics, and identifying disorders by analyzing methylation profiles and chimeric DNA.
Smart Images

Figure 0007594334000010 
Figure 0007594334000011 
Figure 0007594334000012
Abstract
Description
[Technical Field]
[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims the benefit of priority to U.S. Provisional Patent Application No. 63 / 051,210, entitled "Determining Base Modifications in Nucleic Acids," filed July 13, 2020, U.S. Provisional Patent Application No. 63 / 019,790, entitled "Determining Base Modifications in Nucleic Acids," filed May 4, 2020, U.S. Provisional Patent Application No. 62 / 991,891, entitled "Determining Base Modifications in Nucleic Acids," filed March 19, 2020, U.S. Provisional Patent Application No. 62 / 970,586, entitled "Determining Base Modifications in Nucleic Acids," filed February 5, 2020, and U.S. Provisional Patent Application No. 62 / 887,987, entitled "Determining Base Modifications in Nucleic Acids," filed August 16, 2019, the contents of all of which are incorporated herein by reference for all purposes. [Background technology]
[0002] The presence of nucleic acid base modifications varies among various organisms, including viruses, bacteria, plants, fungi, nematodes, insects, and vertebrates (e.g., humans). The most common base modification is the addition of methyl groups to different DNA bases at different positions, known as methylation. Methylation has been found at cytosine, adenine, thymine, and guanine, including 5mC (5-methylcytosine), 4mC (N4-methylcytosine), 5hmC (5-hydroxymethylcytosine), 5fC (5-formylcytosine), 5caC (5-carboxylcytosine), 1mA (N1-methyladenine), 3mA (N3-methyladenine), 7mA (N7-methyladenine), 3mC (N3-methylcytosine), 2mG (N2-methylguanine), 6mG (O6-methylguanine), 7mG (N7-methylguanine), 3mT (N3-methylthymine), and 4mT (O4-methylthymine). In vertebrate genomes, 5mC is the most common type of base methylation, followed by guanine methylation (i.e., in the context of CpG).
[0003] DNA methylation is essential for mammalian development, playing notable roles in gene expression and silencing, embryonic development, transcription, chromatin structure, X-chromosome inactivation, protecting against the activity of repetitive elements, maintaining genome stability during mitosis, and regulating parent-of-origin genomic imprinting.
[0004] DNA methylation plays many important roles in promoter and enhancer silencing in a coordinated manner ( Robertson, 2005 ; Smith and Meissner, 2013 ). Many human diseases have been found to be associated with abnormalities in DNA methylation, including, but not limited to, oncogenic processes, imprinting disorders (e.g., Beckwith-Wiedemann syndrome and Prader-Willi syndrome), repeat instability disorders (e.g., Fragile X syndrome), autoimmune disorders (e.g., systemic lupus erythematosus), metabolic disorders (e.g., type I and type II diabetes), neuropathies, and aging.
[0005] Accurately measuring methylomic modifications of DNA molecules has many clinical implications. One widely used method for measuring DNA methylation is bisulfite sequencing (BS-seq) (Lister et al., 2009; Frommer et al., 1992). In this approach, DNA samples are first treated with bisulfite to convert unmethylated cytosines (i.e., C) to uracil. In contrast, methylated cytosines remain unchanged. The bisulfite-modified DNA is then analyzed by DNA sequencing. In another approach, following bisulfite conversion, the modified DNA is then subjected to polymerase chain reaction (PCR) amplification using primers that can distinguish between bisulfite-converted DNA with different methylation profiles (Herman et al., 1996). This latter approach is called methylation-specific PCR.
[0006] One drawback of such bisulfite-based approaches is that the bisulfite conversion step has been reported to significantly degrade a large portion of the treated DNA (Grunau, 2001). Another drawback is that the bisulfite conversion step generates a strong CG bias (Olova et al., 2018), typically resulting in a reduced signal-to-noise ratio for DNA mixtures with heterogeneous methylation states. Furthermore, bisulfite sequencing cannot sequence long DNA molecules because DNA is degraded during the bisulfite treatment. Therefore, there is a need to determine the base modifications of nucleic acids without prior chemical treatment (e.g., bisulfite conversion) and nucleic acid amplification (e.g., using PCR). Summary of the Invention
[0007] In one embodiment, the present inventors have developed a new method that allows for the determination of base modifications, such as 5mC, in nucleic acids without pretreatment of the template DNA, such as enzymatic and / or chemical conversion, or protein and / or antibody binding. While such pretreatment of the template DNA is not required for base modification determination, in the examples shown, certain pretreatments (e.g., restriction enzyme digestion) may serve to enhance aspects of the invention (e.g., allowing enrichment of CpG sites for analysis). The presently disclosed embodiments can be used to detect different types of base modifications, including, but not limited to, 4mC, 5hmC, 5fC, and 5caC, 1mA, 3mA, 7mA, 3mC, 2mG, 6mG, 7mG, 3mT, and 4mT. Such embodiments can utilize sequencing-derived features, such as kinetic characteristics affected by various base modifications, as well as the identity of nucleotides in a window around a target position where methylation status is determined.
[0008] Embodiments of the present invention can be used for, but are not limited to, single molecule sequencing. One type of single molecule sequencing is single molecule real-time sequencing, which monitors the sequencing progress of a single DNA molecule in real time. One type of single molecule real-time sequencing is commercialized by Pacific Biosciences using a single molecule real-time (SMRT) system. The method can use the pulse width of the signal from the sequenced base, the interpulse duration (IPD) of the base, and the identity of the base to detect modifications of the base or nearby bases. Another single molecule system is based on nanopore sequencing. An example of a nanopore sequencing system is commercialized by Oxford Nanopore Technologies.
[0009] The method developed by the present inventors serves as a tool for detecting base modifications in biological samples and assessing the methylation profile of the sample for various purposes, including but not limited to, research and diagnostic purposes. The detected methylation profile can be used for different analyses. The methylation profile can be used to detect the origin of DNA (e.g., DNA obtained from maternal or fetal tissues, tissues, bacteria, or tumor cells enriched from the blood of cancer patients). Detection of abnormal methylation profiles in tissues can help identify developmental disorders in individuals, and identify and predict tumors or malignancies.
[0010] The present invention can include analyzing the relative methylation levels of haplotypes in an organism. The imbalance in methylation levels between two haplotypes can be used to determine the classification of a disorder. A greater imbalance can indicate the presence of a disorder or a more severe disorder. The disorder can include cancer.
[0011] The methylation pattern of a single molecule can identify chimeric and hybrid DNA. Chimeric and hybrid molecules can contain sequences from two different genes, chromosomes, organelles (e.g., mitochondria, nuclei, chloroplasts), organisms (mammals, bacteria, viruses, etc.), and / or species. Detecting the junctions of chimeric or hybrid DNA molecules can enable the detection of gene fusions in various disorders or diseases, including cancer, prenatal disorders, or congenital disorders.
[0012] A better understanding of the nature and advantages of embodiments of the present invention may be obtained with reference to the following detailed description and accompanying drawings. [Brief explanation of the drawings]
[0013] [Figure 1] 1 shows SMRT sequencing of molecules with base modifications according to an embodiment of the invention. [Figure 2] 1 shows SMRT sequencing of molecules with methylated and unmethylated CpG sites according to an embodiment of the present invention. [Figure 3] 1 illustrates pulse intervals and pulse widths according to an embodiment of the present invention. [Figure 4] 1 shows an example of a measurement window of a Watson strand of DNA for detecting base modifications according to an embodiment of the present invention. [Figure 5] 1 shows an example of a measurement window for a click strand of DNA for detecting base modifications, according to an embodiment of the present invention. [Figure 6] FIG. 1 shows an example of a measurement window by combining data from the Watson strand of DNA and its complementary Crick strand to detect any base modification, according to an embodiment of the present invention. [Figure 7] 1 shows an example of a measurement window by combining data from the Watson strand of DNA and the Crick strand of a nearby region to detect any base modification, according to an embodiment of the present invention. [Figure 8]1 shows an example of measurement windows for Watson strand, Crick strand, and both strands for determining the methylation status of a CpG site according to an embodiment of the present invention. [Figure 9] 1 illustrates a general procedure for building an analytical, computational, mathematical, or statistical model for classifying base modifications, according to an embodiment of the present invention. [Figure 10] 1 illustrates a general procedure for classifying base modifications according to an embodiment of the present invention. [Figure 11]
[0023] Figure 1 illustrates a general procedure for constructing an analytical, computational, mathematical, or statistical model for classifying the methylation status of CpG sites using samples with known methylation status of Watson strands, according to an embodiment of the present invention. [Figure 12] 1 illustrates a general procedure for classifying the Watson strand methylation status of an unknown sample according to an embodiment of the present invention. [Figure 13] FIG. 1 illustrates a general procedure for constructing an analytical, computational, mathematical, or statistical model for classifying methylation status at CpG sites using samples with known methylation status of Crick strands, according to an embodiment of the present invention. [Figure 14] 1 illustrates a general procedure for classifying the methylation status of Crick strands of an unknown sample according to an embodiment of the present invention. [Figure 15] 1 illustrates a general procedure for constructing a statistical model for classifying the methylation status of CpG sites using samples with known methylation status from both Watson and Crick strands, according to an embodiment of the present invention. [Figure 16] 1 illustrates a general procedure for classifying the methylation status of an unknown sample from Watson and Crick strands according to an embodiment of the present invention. [Figure 17A] 1 shows the performance of training and test datasets for determining methylation according to an embodiment of the present invention. [Figure 17B] Same as above. [Figure 18A] 1 shows the performance of training and test datasets for determining methylation according to an embodiment of the present invention. [Figure 18B] Same as above. [Figure 19A] 1 shows the performance of training and test datasets at different sequencing depths for determining methylation, according to an embodiment of the present invention. [Figure 19B] Same as above. [Figure 20A] 1 shows the performance of training and test datasets of different strands for determining methylation according to an embodiment of the present invention. [Figure 20B] Same as above. [Figure 21A] 1 shows the performance of different measurement windows for determining methylation on training and test datasets according to an embodiment of the present invention. [Figure 21B] Same as above. [Figure 22A] 10 shows the performance of training and test datasets for different measurement windows that use only downstream bases to determine methylation, according to an embodiment of the present invention. [Figure 22B] Same as above. [Figure 23A] 10 shows the performance of training and test datasets for different measurement windows that use only upstream bases to determine methylation, according to an embodiment of the present invention. [Figure 23B] Same as above. [Figure 24] 10 shows the performance of methylation analysis using kinetic patterns associated with downstream and upstream bases using asymmetric neighbor sizes in a training dataset according to an embodiment of the present invention. [Figure 25] 10 shows the performance of methylation analysis using kinetic patterns associated with downstream and upstream bases using asymmetric neighbor sizes in a test dataset according to an embodiment of the present invention. [Figure 26] 1 shows the relative importance of features for classifying the methylation status of CpG sites according to an embodiment of the present invention. [Figure 27] 10 shows the performance of motif-based IPD analysis for methylation detection without pulse width signals, according to an embodiment of the present invention. [Figure 28]1 is a graph of a principal component analysis technique using 2 nt upstream and 6 nt downstream of a cytosine subjected to methylation analysis according to an embodiment of the present invention. [Figure 29] 10 is a graph of a performance comparison between a method using principal component analysis and a method using a convolutional neural network, according to an embodiment of the present invention. [Figure 30A] 1 shows the performance of different analytical, computational, mathematical, or statistical models on training and test datasets that use only upstream bases to determine methylation, according to an embodiment of the present invention. [Figure 30B] Same as above. [Figure 31A] FIG. 1 shows an example of one approach for generating molecules with unmethylated adenines by whole genome amplification, according to an embodiment of the present invention. [Figure 31B] 1 illustrates an example of one approach for generating molecules with methylated adenines by whole genome amplification, according to an embodiment of the present invention. [Figure 32A] 10 shows inter-pulse distance (IPD) values across sequenced A bases in template DNA for Watson strands between unmethylated and methylated datasets, according to an embodiment of the present invention. [Figure 32B] Same as above. [Figure 32C] 1 shows a receiver operating characteristic curve for determining Watson strand methylation according to an embodiment of the present invention. [Figure 33A] 10 shows inter-pulse distance (IPD) values across sequenced A bases in template DNA for Crick strands between unmethylated and methylated datasets, according to an embodiment of the present invention. [Figure 33B] Same as above. [Figure 33C] 1 shows a receiver operating characteristic curve for determining Crick strand methylation according to an embodiment of the present invention. [Figure 34] 1 shows the determination of 6 mA for Watson chains according to an embodiment of the present invention. [Figure 35] 1 shows a 6 mA determination of a Crick strand according to an embodiment of the present invention. [Figure 36A] 10 shows the determined probability of being methylated for sequenced A bases in the Watson strand between the uA and mA datasets using a measurement window-based convolutional neural network model, according to an embodiment of the present invention. [Figure 36B] Same as above. [Figure 37] 10 shows the ROC curve for detecting 6mA using a measurement window-based CNN model of sequenced A bases of the Watson strand, according to an embodiment of the present invention. [Figure 38] 10 illustrates a performance comparison between IPD metric-based 6mA detection and measurement window-based 6mA detection, according to an embodiment of the present invention. [Figure 39A] 10 shows the determined probabilities of being methylated for those sequenced A bases in the Crick strand between the uA and mA datasets using a measurement window-based CNN model, according to an embodiment of the present invention. [Figure 39B] Same as above. [Figure 40] 10 shows the performance of 6 mA detection using a measurement window-based CNN model for sequenced A bases of the Crick strand, according to an embodiment of the present invention. [Figure 41] 1 shows examples of methylation states across the A bases of molecules including Watson and Crick strands, according to embodiments of the present invention. [Figure 42] 10 shows an example of reinforcement training by selectively using A bases of an mA dataset with IPD values above its 10th percentile, according to an embodiment of the present invention. [Figure 43] 1 is a graph of the percentage of unmethylated adenines in the mA dataset versus the number of subreads in each well, according to an embodiment of the present invention. [Figure 44] 1 shows the pattern of methyladenine between the Watson and Crick strands of double-stranded DNA molecules in a test dataset, according to an embodiment of the present invention. [Figure 45]1 is a table showing the percentage of fully unmethylated molecules, hemimethylated molecules, fully methylated molecules, and molecules with interlaced methyladenine patterns in training and test datasets, according to an embodiment of the present invention. [Figure 46] 1 shows representative examples of molecules with fully unmethylated molecules, hemimethylated molecules, fully methylated molecules, and molecules with interlaced methyladenine patterns around adenine sites, according to embodiments of the present invention. [Figure 47] 1 shows an example of a long read (6,265 bp) with a CpG island (shaded in yellow) according to an embodiment of the present invention. [Figure 48] 1 is a table showing that nine DNA molecules were sequenced by Pacific Biosciences SMRT sequencing and that they overlap with imprinted regions, according to an embodiment of the present invention. [Figure 49] 1 shows an example of genomic imprinting, according to an embodiment of the present invention. [Figure 50] 1 shows an example of determining methylation patterns in imprinted regions according to an embodiment of the present invention. [Figure 51] 1 shows a comparison of methylation levels estimated between the new approach and conventional bisulfite sequencing according to an embodiment of the present invention. [Figure 52A] 1 shows the performance of detecting methylation in plasma DNA according to embodiments of the present disclosure. (A) Relationship between predicted probability of methylation and range of methylation levels quantified by bisulfite sequencing. (B) Correlation between methylation levels determined by Pacific Biosciences (PacBio) sequencing (y-axis) and methylation levels quantified by bisulfite sequencing at 10 Mb resolution (x-axis) according to embodiments present in the present disclosure. [Figure 52B] Same as above. [Figure 53] 1 shows the correlation of genome representation (GR) of the Y chromosome between Pacific Biosciences SMRT sequencing and BS-seq, according to an embodiment of the present invention. [Figure 54] 1 shows an example of CpG block-based detection of methylation using CpG blocks, each having a series of CpG sites, according to an embodiment of the present invention: 5mC: methylated, C: unmethylated. [Figure 55A] Figure 1 shows training and testing of methylation calling of human DNA molecules using a CpG block-based approach according to an embodiment of the present invention: (A) Performance on a training dataset; (B) Performance on an independent test dataset. [Figure 55B] Same as above. [Figure 56A] 1 shows copy number alterations in tumor tissue according to an embodiment of the present invention. [Figure 56B] Same as above. [Figure 57A] 1 shows copy number alterations in tumor tissue according to an embodiment of the present invention. [Figure 57B] Same as above. [Figure 58] FIG. 1 shows a schematic diagram of plasma DNA tissue mapping from plasma of pregnant women using estimated methylation levels according to an embodiment of the present invention. [Figure 59] 1 shows the correlation between the placental contribution to estimated maternal plasma DNA and the fetal DNA fraction estimated by Y chromosome reads, according to an embodiment of the present invention. [Figure 60] 1 shows a table summarizing sequencing data from different human tissue DNA samples according to an embodiment of the present invention. [Figure 61] 1 shows diagrams of various methods for analyzing methylation patterns according to embodiments of the present invention. [Figure 62A] 1 shows a comparison of methylation density at the whole genome level quantified by bisulfite sequencing and single molecule real-time sequencing, according to an embodiment of the present invention. [Figure 62B] Same as above. [Figure 63A] 1 shows the differential correlation of global methylation levels quantified by bisulfite sequencing and single molecule real-time sequencing, according to an embodiment of the present invention. [Figure 63B] Same as above. [Figure 63C] Same as above. [Figure 64A] 1 shows methylation patterns for hepatocellular carcinoma (HCC) cell lines and buffy coat samples from healthy control subjects at 1 Mnt resolution, along with methylation levels determined by bisulfite sequencing and single molecule real-time sequencing according to an embodiment of the present invention. [Figure 64B] Same as above. [Figure 65A] 1 shows a scatter plot of methylation levels at 1 mM resolution determined by bisulfite sequencing and single molecule real-time sequencing according to an embodiment of the present invention for buffy coat samples from an HCC cell line (HepG2) and a healthy control subject. [Figure 65B] Same as above. [Figure 66A] 1 shows a scatter plot of methylation levels at 100 kbp resolution determined by bisulfite sequencing and single molecule real-time sequencing according to an embodiment of the present invention for buffy coat samples from an HCC cell line (HepG2) and a healthy control subject. [Figure 66B] Same as above. [Figure 67A] 1 shows methylation patterns for HCC tumor tissues and adjacent normal tissues at 1 Mnt resolution, along with methylation levels determined by bisulfite sequencing and single molecule real-time sequencing according to an embodiment of the present invention. [Figure 67B] Same as above. [Figure 68A] 1 shows a scatter plot of methylation levels at 1 Mnt resolution for HCC tumor tissues and adjacent normal tissues as determined by bisulfite sequencing and single molecule real-time sequencing according to an embodiment of the present invention. [Figure 68B] Same as above. [Figure 69A] 1 shows a scatter plot of methylation levels at 100 knt resolution for HCC tumor tissues and adjacent normal tissues as determined by bisulfite sequencing and single molecule real-time sequencing according to an embodiment of the present invention. [Figure 69B] Same as above. [Figure 70A] 1 shows methylation patterns for HCC tumor tissues and adjacent normal tissues at 1 Mnt resolution, along with methylation levels determined by bisulfite sequencing and single molecule real-time sequencing according to an embodiment of the present invention. [Figure 70B] Same as above. [Figure 71A] 1 shows a scatter plot of methylation levels at 1 Mnt resolution for HCC tumor tissues and adjacent normal tissues as determined by bisulfite sequencing and single molecule real-time sequencing according to an embodiment of the present invention. [Figure 71B] Same as above. [Figure 72A] 1 shows a scatter plot of methylation levels at 100 knt resolution for HCC tumor tissues and adjacent normal tissues as determined by bisulfite sequencing and single molecule real-time sequencing according to an embodiment of the present invention. [Figure 72B] Same as above. [Figure 73] 1 shows an example of an aberrant pattern of methylation near the tumor suppressor gene CDKN2A, according to an embodiment of the present invention. [Figure 74A] 1 shows variably methylated regions detected by single molecule real-time sequencing, according to an embodiment of the present invention. [Figure 74B] Same as above. [Figure 75] 1 shows the methylation patterns of Hepatitis B virus DNA between HCC tissues and adjacent non-tumor tissues using single molecule real-time sequencing according to an embodiment of the present invention. [Figure 76A] 1 shows the methylation levels of Hepatitis B virus DNA in liver tissue from patients with cirrhosis but without HCC using bisulfite sequencing, according to an embodiment of the present invention. [Figure 76B] 1 shows the methylation levels of Hepatitis B virus DNA in HCC tissues using bisulfite sequencing according to an embodiment of the present invention. [Figure 77]1 shows methylation haplotype analysis according to an embodiment of the present invention. [Figure 78] 1 shows the size distribution of sequenced molecules determined from consensus sequences, according to an embodiment of the present invention. [Figure 79A] 1 shows examples of allele methylation patterns in imprinted regions according to embodiments of the present invention. [Figure 79B] Same as above. [Figure 79C] Same as above. [Figure 79D] Same as above. [Figure 80A] 1 shows examples of allele methylation patterns in non-imprinted regions according to embodiments of the present invention. [Figure 80B] Same as above. [Figure 80C] Same as above. [Figure 80D] Same as above. [Figure 81] 1 shows a table of methylation levels of allele-specific fragments according to an embodiment of the present invention. [Figure 82] 1 shows an example of using methylation profiles to determine the placental origin of plasma DNA during pregnancy, according to an embodiment of the present invention. [Figure 83] 1 shows an analysis of fetal-specific DNA methylation according to an embodiment of the present invention. [Figure 84A] 1 shows the performance of different measurement window sizes across different reagent kits for SMRT-seq, according to an embodiment of the present invention. [Figure 84B] Same as above. [Figure 84C] Same as above. [Figure 85A] 1 shows the performance of different measurement window sizes across different reagent kits for SMRT-seq, according to an embodiment of the present invention. [Figure 85B] Same as above. [Figure 85C] Same as above. [Figure 86A]1 shows the correlation of global methylation levels quantified by bisulfite sequencing and SMRT-seq (Sequel II Sequencing Kit 2.0) according to an embodiment of the present invention. [Figure 86B] Same as above. [Figure 86C] Same as above. [Figure 87A] 1 shows a comparison of global methylation levels between various tumor tissues and paired adjacent non-tumor tissues according to an embodiment of the present invention. [Figure 87B] Same as above. [Figure 88] 1 illustrates determining methylation status using sequence context determined from a circular consensus sequence (CCS), according to an embodiment of the present invention. [Figure 89] 1 shows a ROC curve for the detection of methylated CpG sites using sequence context determined from CCS, according to an embodiment of the present invention. [Figure 90] 1 shows ROC curves for detection of methylated CpG sites without CCS information and without prior alignment to a reference genome according to an embodiment of the present invention. [Figure 91] 1 shows an example of the preparation of a molecule for single molecule real-time sequencing, according to an embodiment of the present invention. [Figure 92] FIG. 1 shows a diagram of a CRISPR / Cas9 system, according to an embodiment of the present invention. [Figure 93] FIG. 1 shows an example of a Cas9 complex for introducing two cleavages spanning an end-blocked molecule of interest, according to embodiments of the invention. [Figure 94] 1 shows the methylation distribution of Alu regions determined by bisulfite sequencing and single molecule real-time sequencing according to an embodiment of the present invention. [Figure 95] 1 shows the distribution of methylation levels of Alu regions as determined by a model using the results of single molecule real-time sequencing, according to an embodiment of the present invention. [Figure 96]1 shows a table of tissues and methylation levels of Alu regions in the tissues according to an embodiment of the present invention. [Figure 97] 1 shows a cluster analysis of different cancer types using methylation signals associated with Alu repeat sequences, according to an embodiment of the present invention. [Figure 98A] 1 shows the effect of read depth on quantification of global methylation levels in a test dataset involving whole genome amplification and M.SsssI treatment, according to an embodiment of the present invention. [Figure 98B] Same as above. [Figure 99] 1 shows a comparison between global methylation levels determined by SMRT-seq (Sequel II Sequencing Kit 2.0) and BS-seq using different subread depth cutoffs, according to an embodiment of the present invention. [Figure 100] 1 is a table showing the effect of subread depth on the correlation of methylation levels between two measurements by SMRT-seq (Sequel II Sequencing Kit 2.0) and BS-seq, according to an embodiment of the present invention. [Figure 101] 1 shows subread depth distribution with respect to fragment size in data generated by Sequel II Sequencing Kit 2.0, according to an embodiment of the present invention. [Figure 102] 1 illustrates a method for detecting a modification of a nucleotide of a nucleic acid molecule, according to an embodiment of the present invention. [Figure 103] 1 illustrates a method for detecting a modification of a nucleotide of a nucleic acid molecule, according to an embodiment of the present invention. [Figure 104] 1 shows comparative haplotype-based methylation imbalance analysis according to an embodiment of the present invention. [Figure 105A] 1 is a table of haplotype blocks showing differential methylation levels between haplotype I (Hap I) and haplotype II (Hap II) in tumor DNA compared to adjacent non-tumor tissue DNA for case TBR3033, according to an embodiment of the present invention. [Figure 105B] Same as above. [Figure 106] 1 is a table of haplotype blocks showing differential methylation levels between Hap I and Hap II in tumor DNA compared to adjacent normal tissue DNA for case TBR3032, according to an embodiment of the present invention. [Figure 107A] 1 is a table summarizing the number of haplotype blocks showing methylation imbalance between two haplotypes between a tumor and adjacent non-tumor tissue based on data generated by Sequel II Sequencing Kit 2.0, according to an embodiment of the present invention. [Figure 107B] 1 is a table summarizing the number of haplotype blocks showing methylation imbalance between two haplotypes in tumor tissues of different tumor stages based on data generated by Sequel II Sequencing Kit 2.0, according to an embodiment of the present invention. [Figure 108] 1 shows comparative haplotype-based methylation imbalance analysis according to an embodiment of the present invention. [Figure 109] 1 illustrates a method for classifying a disorder in an organism having a first haplotype and a second haplotype according to an embodiment of the present invention. [Figure 110] 1 shows the creation of a human-mouse hybrid fragment in which the human portion is methylated but the mouse portion is unmethylated, according to an embodiment of the present invention. [Figure 111] 1 shows the creation of a human-mouse hybrid fragment in which the human portion is unmethylated but the mouse portion is methylated, according to an embodiment of the present invention. [Figure 112] 1 shows the chain length distribution of DNA molecules in a DNA mixture (sample MIX01) after ligation, according to an embodiment of the present invention. [Figure 113] 1 shows the junction region where a first DNA (A) and a second DNA (B) bind together, according to an embodiment of the present invention. [Figure 114] 1 shows methylation analysis of a DNA mixture according to an embodiment of the present invention. [Figure 115] 1 shows a box plot of methylation probability for CpG sites in sample MIX01, according to an embodiment of the present invention. [Figure 116] 1 shows the chain length distribution of DNA molecules in the DNA mixture after cross-ligation of sample MIX02, according to an embodiment of the present invention. [Figure 117] 1 shows a box plot of methylation probability for CpG sites in sample MIX02, according to an embodiment of the present invention. [Figure 118] 1 is a table comparing methylation determined by bisulfite sequencing and Pacific Biosciences sequencing of MIX01, according to an embodiment of the present invention. [Figure 119] 1 is a table comparing methylation determined by bisulfite sequencing and Pacific Biosciences sequencing of MIX02, according to an embodiment of the present invention. [Figure 120A] 1 shows methylation levels in 5 Mb bins of human-only and mouse-only DNA for MIX01 and MIX02, according to an embodiment of the invention. [Figure 120B] Same as above. [Figure 121A] 1 shows methylation levels in 5 Mb bins of the human and mouse portions of human-mouse hybrid DNA fragments for MIX01 and MIX02, according to an embodiment of the present invention. [Figure 121B] Same as above. [Figure 122A] 1 is a representative graph showing the methylation status in a single human-mouse hybrid molecule, according to an embodiment of the present invention. [Figure 122B] Same as above. [Figure 123] 1 illustrates a method for detecting a chimeric molecule in a biological sample according to an embodiment of the present invention. [Figure 124] 1 illustrates a measurement system according to an embodiment of the present invention. [Figure 125] 1 shows a block diagram of an exemplary computer system usable with systems and methods according to embodiments of the present invention. [Figure 126] 1 shows MspI-based targeted single-molecule real-time sequencing using DNA end repair and A-tailing according to an embodiment of the present invention. [Figure 127A] 1 shows the size distribution of MspI digested fragments according to an embodiment of the present invention. [Figure 127B] Same as above. [Figure 128] 1 shows a table of numbers of DNA molecules for certain selected size ranges, according to an embodiment of the present invention. [Figure 129] 1 is a graph of percent coverage of CpG sites within a CpG island versus DNA fragment size after restriction enzyme digestion, according to an embodiment of the present invention. [Figure 130] 1 shows MspI-based targeted single-molecule real-time sequencing without DNA end repair and A-tailing, according to an embodiment of the present invention. [Figure 131] 1 shows MspI-based targeted single-molecule real-time sequencing with reduced probability of adapter self-ligation according to an embodiment of the present invention. [Figure 132] 1 is a graph of global methylation levels between placenta and buffy DNA samples determined by MspI-based targeted single molecule real-time sequencing, according to an embodiment of the present invention. [Figure 133] 1 shows cluster analysis of placenta and buffy coat samples using DNA methylation profiles determined by MspI-based targeted single molecule real-time sequencing, according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0014] term A "tissue" corresponds to a group of cells grouped together as a functional unit. Two or more types of cells can be found within a single tissue. Different types of tissue may be composed of different types of cells (e.g., liver cells, alveolar cells, or blood cells), may correspond to tissues from different organisms (maternal versus fetal, tissues of a transplanted subject, tissues of an organism infected with a microorganism or virus), or healthy versus tumor cells. A "reference tissue" corresponds to the tissue used to determine tissue-specific methylation levels. Multiple samples of the same tissue type from different individuals can be used to determine the tissue-specific methylation level of that tissue type.
[0015] "Biological sample" refers to any sample collected from a human subject. A biological sample can be a tissue biopsy, fine needle aspirate, or blood cells. The sample can also be, for example, plasma, serum, or urine from a pregnant woman. Stool samples can also be used. In various embodiments, the majority of the DNA in a biological sample from a pregnant woman enriched for cell-free DNA (e.g., a plasma sample obtained via a centrifugation protocol) can be cell-free, e.g., more than 50%, more than 60%, more than 70%, more than 80%, more than 90%, more than 95%, or more than 99% of the DNA can be cell-free. The centrifugation protocol can include, for example, obtaining a fluid portion at 3,000 g for 10 minutes and recentrifuging at 30,000 g for an additional 10 minutes to remove residual cells. In certain embodiments, the 3,000 g centrifugation step can be followed by filtration of the fluid portion (e.g., using a filter with a pore size of 5 μm or less).
[0016] A "sequence read" refers to a chain of nucleotides sequenced from any portion or all of a nucleic acid molecule. For example, a sequence read can be a short chain of nucleotides (e.g., about 20-150) sequenced from a nucleic acid fragment, a short chain of nucleotides at one or both ends of a nucleic acid fragment, or the sequencing of an entire nucleic acid fragment present in a biological sample. Sequence reads can be obtained, for example, using sequencing techniques, or in a variety of ways using probes, e.g., hybridization arrays or capture probes, or amplification techniques such as polymerase chain reaction (PCR) or linear amplification, using single primers or isothermal amplification.
[0017] A "subread" is a sequence generated from all bases of one strand of a circularized DNA template, which is copied into one continuous strand by a DNA polymerase. For example, a subread may correspond to one strand of DNA in the circularized DNA template. In such an example, after circularization, one double-stranded DNA molecule has two subreads (one for each sequencing pass). In some embodiments, the generated sequence may include a subset of all bases of one strand, for example, due to the presence of sequencing errors.
[0018] A "site" (also called a "genomic site") corresponds to a single site, which can be a single base position, or a group of correlated base positions, e.g., a CpG site, or a larger group of correlated base positions. A "locus" can correspond to a region that includes multiple sites. A locus can contain only one site, which would make the locus equivalent to the site in that context.
[0019] "Methylation state" refers to the state of methylation at a given site, e.g., a site may be either methylated, unmethylated, or possibly undetermined.
[0020] A "methylation index" for each genomic site (e.g., a CpG site) can refer to the proportion of DNA fragments (e.g., as determined from sequence reads or probes) that exhibit methylation at that site across the total number of reads covering that site. A "read" can correspond to information obtained from a DNA fragment (e.g., the methylation state of a site). A read can be obtained using a reagent (e.g., a primer or probe) that preferentially hybridizes to DNA fragments of a particular methylation state at one or more sites. Typically, such a reagent is applied after processing DNA molecules with a process that differentially modifies or recognizes them according to their methylation state, such as bisulfite conversion, or methylation-sensitive restriction enzymes, or methylation-binding proteins, or anti-methylcytosine antibodies, or single-molecule sequencing technologies that recognize methylcytosine and hydroxymethylcytosine (e.g., single-molecule real-time sequencing and nanopore sequencing (e.g., from Oxford Nanopore Technologies)).
[0021] The "methylation density" of a region may refer to the number of reads at a site within the region that show methylation divided by the total number of reads covering the site in this region. This site may have specific characteristics, for example, a CpG site. Thus, the "CpG methylation density" of a region refers to the number of reads that show CpG methylation divided by the total number of reads that cover the CpG site in this region (e.g., a specific CpG site, a CpG site within a CpG island, or a larger region). For example, the methylation density of each 100 kb bin in the human genome can be determined from the total number of unconverted cytosines (corresponding to methylated cytosines) at CpG sites after bisulfite treatment as a percentage of all CpG sites covered by sequence reads mapped to the 100 kb region. This analysis can also be performed for other bin sizes, such as 500 bp, 5 kb, 10 kb, 50 kb, or 1 Mb. A region may be the entire genome, or a chromosome, or a portion of a chromosome (e.g., a chromosome arm). The methylation index of a CpG site is the same as the methylation density of that region if the region contains only that CpG site. The "percentage of methylated cytosines" can refer to the total number of analyzed cytosine residues in this region, i.e., the number of cytosine sites "C" that are shown to be methylated (e.g., unconverted after bisulfite conversion), including cytosines outside of the CpG context. Examples of "methylation levels" include the methylation index, methylation density, the number of molecules methylated at one or more sites, and the percentage of molecules (e.g., cytosines) methylated at one or more sites.Apart from bisulfite conversion, other processes known to those skilled in the art can be used to examine the methylation state of DNA molecules, including, but not limited to, methylation-state-sensitive enzymes (e.g., methylation-sensitive restriction enzymes), methylation-binding proteins, single-molecule sequencing using methylation-state-sensitive platforms (e.g., nanopore sequencing (Schreiber et al. Proc Natl Acad Sci 2013;110:18910-18915) and single-molecule real-time sequencing (e.g., by Pacific Biosciences) (Flusberg et al. Nat Methods 2010;7:461-465)).
[0022] A "methylome" provides a measure of the amount of DNA methylation at multiple sites or loci in a genome. The methylome may correspond to the entire genome, a substantial portion of the genome, or a relatively small portion(s) of the genome.
[0023] A "maternal plasma methylome" is a methylome determined from the plasma or serum of a pregnant animal (e.g., a human). The maternal plasma methylome is an example of a cell-free methylome because plasma and serum contain cell-free DNA. The maternal plasma methylome is also an example of a mixed methylome because it is a mixture of DNA from different organs, tissues, or cells in the body. In one embodiment, such cells are hematopoietic cells, including, but not limited to, cells of the erythroid (i.e., red cell) lineage, myeloid lineage (e.g., neutrophils and their precursors), and megakaryocytic lineage. During pregnancy, the plasma methylome may contain methylomic information from both the fetus and the mother. A "cellular methylome" corresponds to the methylome determined from a patient's cells (e.g., blood cells). The methylome of blood cells is referred to as the blood cell methylome (or blood methylome).
[0024] A "methylation profile" includes information related to DNA or RNA methylation at multiple sites or regions. Information related to DNA methylation may include, but is not limited to, the methylation index of CpG sites, the methylation density (abbreviated MD) of CpG sites in a region, the distribution of CpG sites across a contiguous region, the methylation pattern or level of each individual CpG site within a region containing two or more CpG sites, and non-CpG methylation. In one embodiment, a methylation profile may include the methylation or non-methylation patterns of two or more types of bases (e.g., cytosine or adenine). The methylation profile of a substantial portion of a genome can be considered equivalent to a methylome. "DNA methylation" in mammalian genomes typically refers to the addition of a methyl group to the 5' carbon of cytosine residues between CpG dinucleotides (i.e., 5-methylcytosine). DNA methylation can occur at cytosine in other contexts, such as CHG and CHH, where H is adenine, cytosine, or thymine. Cytosine methylation can also occur in the form of 5-hydroxymethylcytosine. 6 Non-cytosine methylations such as -methyladenine have also been reported.
[0025] A "methylation pattern" refers to the order of methylated and unmethylated bases. For example, a methylation pattern can be the order of methylated bases on a single DNA strand, a single double-stranded DNA molecule, or another type of nucleic acid molecule. As an example, three consecutive CpG sites can have any of the following methylation patterns: UUU, MMM, UMM, UMU, UUM, MUM, MUU, or MMU, where "U" indicates an unmethylated site and "M" indicates a methylated site. Without being limited thereto, when extending this concept to base modifications including methylation, the term "modification pattern" will be used to refer to the order of modified and unmodified bases. For example, a modification pattern can be the order of modified bases on a single DNA strand, a single double-stranded DNA molecule, or another type of nucleic acid molecule. As an example, three consecutive potentially modifiable sites can have any of the following modification patterns: UUU, MMM, UMM, UMU, UUM, MUM, MUU, or MMU. Here, "U" indicates an unmodified site and "M" indicates a modified site. An example of a base modification that is not based on methylation is an oxidative change such as 8-oxoguanine.
[0026] The terms "hypermethylated" and "hypomethylated" can refer to the methylation density of a single DNA molecule as measured by the methylation level of that single molecule, e.g., the number of methylated bases or nucleotides in that molecule divided by the total number of methylable bases or nucleotides in that molecule. A hypermethylated molecule is a molecule whose methylation level of a single molecule is equal to or greater than a threshold, which can be defined for each application. This threshold can be 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 95%. A hypomethylated molecule is a molecule whose methylation level of a single molecule is equal to or less than a threshold, which can be defined for each application and can vary for each application. This threshold can be 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 95%.
[0027] The terms "hypermethylated" and "hypomethylated" may also refer to the methylation level of a population of DNA molecules as measured by the methylation level of a plurality of those molecules. A hypermethylated population of molecules is a population in which a plurality of the molecules have a methylation level above a threshold, which may be defined for each application and may vary from application to application. This threshold may be 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 95%. A hypomethylated population of molecules is a population in which a plurality of the molecules have a methylation level below a threshold, which may be defined for each application and may vary from application to application. This threshold may be 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 95%. In one embodiment, the population of molecules may be aligned to one or more selected genomic regions. In one embodiment, the selected genomic region(s) may be associated with a disease such as cancer, a genetic disorder, an imprinting disorder, a metabolic disorder, or a neurological disorder. The selected genomic region(s) may have a length of 50 nucleotides (nt), 100 nt, 200 nt, 300 nt, 500 nt, 1000 nt, 2 knt, 5 knt, 10 knt, 20 knt, 30 knt, 40 knt, 50 knt, 60 knt, 70 knt, 80 knt, 90 knt, 100 knt, 200 knt, 300 knt, 400 knt, 500 knt, or 1 Mnt.
[0028] The term "sequencing depth" refers to the number of times a locus is covered by sequence reads aligned to that locus. A locus may be as small as a nucleotide, or as large as a chromosome arm, or as large as the entire genome. Sequencing depth can be expressed as 50x, 100x, etc., where "x" refers to the number of times a locus is covered by sequence reads. Sequencing depth can also be applied to multiple loci or the entire genome, in which case x can refer to the average number of times a locus or haploid genome or the entire genome is sequenced, respectively. Ultra-deep sequencing can refer to a sequencing depth of at least 100x.
[0029] As used herein, the term "classification" refers to any number(s) or other feature(s) associated with a particular characteristic of a sample. For example, a "+" symbol (or the word "positive") can mean that the sample is classified as having a deletion or amplification. Classification can be binary (e.g., positive or negative) or have more levels of classification (e.g., a scale of 1 to 10 or 0 to 1).
[0030] The terms "cutoff" and "threshold" refer to a predetermined number used in a certain operation. For example, a cutoff size can refer to the size above which fragments are excluded. A threshold can be a value above or below what a particular classification requires. Either of these terms can be used in any of these contexts. A cutoff or threshold can be a "reference value" or can be derived from a reference value that represents a particular classification or distinguishes between two or more classifications. Such a reference value can be determined in various ways, as will be understood by those skilled in the art. For example, a metric can be determined for two different cohorts of subjects with different known classifications, and a reference value can be selected as representative of one classification (e.g., the average) or as a value between two clusters of the metric (e.g., selected to obtain a desired sensitivity and specificity). As another example, the reference value can be determined based on statistical analysis or sample simulation.
[0031] The term "level of cancer" can refer to whether cancer is present (i.e., present or absent), the stage of cancer, tumor size, whether there is metastasis, total tumor burden in the body, the cancer's response to treatment, and / or other measures of cancer severity (e.g., cancer recurrence). Cancer levels can be numbers or other indicia, such as symbols, letters, and colors. The level can be zero. Cancer levels can also include premalignant or precancerous conditions. Cancer levels can be used in a variety of ways. For example, screening can check for the presence of cancer in a person who was not previously known to have cancer. Evaluation can examine a person who has been diagnosed with cancer to monitor the progression of cancer over time, study the effectiveness of therapy, or determine a prognosis. In one embodiment, prognosis can be expressed as the likelihood that a patient will die from cancer, or the likelihood that the cancer will progress after a certain duration or time, or the likelihood or extent that the cancer will metastasize. Detection can mean "screening" or checking whether a person with suggestive features of cancer (e.g., symptoms or other positive tests) has cancer.
[0032] "Level of pathology" (or level of injury) can refer to the amount, degree, or severity of pathology associated with an organism, and the level can be as described above for cancer. Another example of pathology is the rejection of a transplanted organ. Other examples of pathology can include genetic imprinting disorders, autoimmune attacks (e.g., lupus nephritis damage or multiple sclerosis that damage the kidney), inflammatory diseases (e.g., hepatitis), fibrotic processes (e.g., cirrhosis), fatty infiltration (e.g., fatty liver disease), degenerative processes (e.g., Alzheimer's disease), and ischemic tissue damage (e.g., myocardial infarction or stroke). A healthy state of a subject can be considered a pathology-free classification.
[0033] "Pregnancy-related disorders" include any disorder characterized by abnormal relative expression levels of genes in maternal and / or fetal tissues, including, but not limited to, preeclampsia, intrauterine growth restriction, invasive placentation, preterm birth, hemolytic disease of the newborn, placental insufficiency, hydrops fetalis, fetal malformations, HELLP syndrome, systemic lupus erythematosus, and other maternal immune disorders.
[0034] The abbreviation "bp" refers to base pairs. In some cases, "bp" may be used to indicate the length of a DNA fragment even if the DNA fragment is single-stranded and does not contain base pairs. In the context of single-stranded DNA, "bp" may be interpreted as providing the length of the strand in nucleotides.
[0035] The abbreviation "nt" refers to a nucleotide. In some cases, "nt" can be used to indicate the length of a single-stranded DNA in bases. "nt" can also be used to indicate a relative location, such as upstream or downstream of the analyzed locus. In some contexts regarding technical conceptualization, data display, processing, and analysis, "nt" and "bp" may be used interchangeably.
[0036] The term "sequence context" can refer to the base composition (A, C, G, or T) and base order in a stretch of DNA. Such a stretch of DNA can surround a base or target base that is subjected to base modification analysis. For example, sequence context can refer to bases upstream and / or downstream of a base that is subjected to base modification analysis.
[0037] The term "kinetic features" can refer to features derived from sequencing, including single-molecule real-time sequencing. Such features can be used for base modification analysis. Examples of kinetic features include upstream and downstream sequence context, strand information, pulse interval, pulse width, and pulse intensity. Single-molecule real-time sequencing continuously monitors the effect of polymerase activity on a DNA template. Therefore, measurements generated from such sequencing can be considered kinetic features, for example, nucleotide sequences.
[0038] The term "machine learning model" may include models based on predicting test data using sample data (e.g., training data) and thus may include supervised learning. Machine learning models are often developed using a computer or processor. Machine learning models may include statistical models.
[0039] The term "data analytics framework" may include algorithms and / or models that can receive data as input and then output a predictive outcome. Examples of "data analytics frameworks" include statistical models, mathematical models, machine learning models, other artificial intelligence models, and combinations thereof.
[0040] The term "real-time sequencing" can refer to techniques that involve data collection or monitoring of the reactions involved in sequencing while they are progressing. For example, real-time sequencing can involve optical monitoring or imaging of DNA polymerase incorporating new bases.
[0041] The term "about" or "approximately" can mean within an acceptable error range of a particular value as determined by one of ordinary skill in the art, which depends in part on the method of measuring or determining the value, i.e., the limitations of the measurement system. For example, "about" can mean within 1 or more standard deviations, as is customary in the art. Alternatively, "about" can mean a range of up to 20%, up to 10%, up to 5%, or up to 1% of a given value. Alternatively, particularly with respect to biological systems or processes, the term "about" or "approximately" can mean within an order of magnitude, within 5-fold, or more preferably within 2-fold of a value. When a specific value is described in this application and claims, unless otherwise specified, the term "about" should be assumed to be within an acceptable error range of the particular value. The term "about" can have the meaning commonly understood by one of ordinary skill in the art. The term "about" can refer to ±10%. The term "about" can refer to ±5%.
[0042] Achieving the determination of bisulfite-free base modifications, including methylated bases, has been the subject of various research efforts, but none have been shown to be commercially viable. Recently, a bisulfite-free method for detecting 5mC and 5hmC using mild conditions for 5mC and 5hmC base conversion was published (Y. Liu et al., 2019). This method involves multiple steps of enzymatic and chemical reactions, including ten-eleven translocation (TET) oxidation, pyridine borane reduction, and PCR. The efficiency of each step in the conversion reaction, as well as PCR bias, negatively impacts the final accuracy of 5mC analysis. For example, a 5mC conversion rate of approximately 96% and a false-negative rate of approximately 3% have been reported. Such performance may limit the ability to detect specific, subtle changes in methylation in the genome. Meanwhile, enzymatic conversion may not work equally well across the entire genome. For example, the conversion rate of 5hmC was 8.2% lower than that of 5mC, and the conversion rate for non-CpGs was 11.4% lower than that for CpGs (Y. Liu et al., 2019). Therefore, the ideal situation would be to develop an approach to measure base modifications in natural DNA molecules without a prior conversion step (chemical, enzymatic, or a combination thereof) and even without an amplification step.
[0043] Several proof-of-concept studies have demonstrated that electrical signals generated by long-read nanopore sequencing approaches (e.g., using systems developed by Oxford Nanopore Technologies) can detect methylation status using deep learning methods. In addition to Oxford Nanopore, other single-molecule sequencing approaches enable long-read detection. One example is single-molecule real-time sequencing. One example of single-molecule real-time sequencing is the Pacific Biosciences SMRT system. Due to the single-molecule principle, real-time sequencing (e.g., the Pacific Biosciences SMRT system) differs from non-optical nanopore systems (e.g., Oxford Nanopore Technologies), and base modification detection approaches developed for such non-optical nanopore systems cannot be used for single-molecule real-time sequencing. For example, non-optical nanopore systems are not designed to capture the fluorescent signal patterns generated by immobilized DNA polymerase-based DNA synthesis (e.g., employed in single-molecule real-time sequencing such as the Pacific Biosciences SMRT system). As a further example, in the Oxford Nanopore sequencing platform, each measured electrical event is associated with a k-mer (e.g., a 5-mer) (Q. Liu et al., 2019). However, in the Pacific Biosciences SMRT sequencing platform, each fluorescent event is generally associated with a single incorporated base. Furthermore, a single DNA molecule is sequenced multiple times with Pacific Biosciences SMRT sequencing, including the Watson strand and the Crick strand. Conversely, in the Oxford Nanopore long-read sequencing approach, a sequence read is performed once for each of the Watson strand and the Crick strand.
[0044] It has been reported that polymerase kinetics is affected by the methylation state of sequences in Escherichia coli (Flusberg et al., 2010). Previous studies have shown that using polymerase kinetics in single-molecule real-time sequencing to estimate the methylation state of specific CpGs (5mC vs. C) in single molecules is more challenging than detecting 6mC, 4mC, 5hmC, and 8-oxoguanine. This is because the methyl group is small, oriented in the major groove, and does not participate in base pairing, resulting in very small disruption in kinetics due to 5mC (Clark et al., 2013). Therefore, there is a lack of approaches for determining the methylation state of cytosine at the single-molecule level.
[0045] Suzuki et al. developed an algorithm to combine the interpulse distance (IPD) ratios of nearby CpG sites to increase the reliability of identifying the methylation status of those sites (Suzuki et al., 2016). However, this algorithm could only predict fully methylated or completely unmethylated genomic regions, and was unable to determine intermediate methylation patterns.
[0046] Current approaches to single-molecule real-time sequencing use only one or two parameters individually, resulting in very limited accuracy in detecting 5mC due to the difference in measurements between 5-methylcytosine and cytosine. For example, Flusberg et al. demonstrated that IPD changes for base modifications including N6-methyladenosine, 5-methylcytosine, and 5-hydroxymethylcytosine. However, they did not find a significant effect on the pulse width (PW) of sequencing kinetics. Therefore, the method they used to predict base modifications used only IPD, not PW, as an example, using detection of N6-methyladenosine.
[0047] In follow-up publications by the same group (Clark et al., 2012, Clark et al., 2013), IPD rather than PW was incorporated into the algorithm for detecting 5-methylcytosine. In Clark et al. (2012), the detection rate of 5-methylcytosine without conversion to 5-methylcytosine ranged from 1.9% to 4.3%. Furthermore, in Clark et al. (2013), the authors further reaffirmed the subtleties of the kinetic signature of 5-methylcytosine. To overcome the low sensitivity of 5-methylcytosine detection, Clark et al. further developed a method to improve the sensitivity of 5-methylcytosine by converting 5-methylcytosine to 5-carboxymethylcytosine using the ten-eleven translocation (Tet) protein (Clark et al., 2013). This was because the IPD change resulting from 5-carboxymethylcytosine was much larger than that of 5-methylcytosine.
[0048] A recent report by Blow et al. used the IPD ratio-based method previously described by Flusberg et al. to detect base modifications in 217 bacterial and 13 archaeal species at 130x read coverage per organism (Blow et al., 2016). Of all base modifications they identified, only 5% involved 5-methylcytosine. They attributed this low detection rate of 5-methylcytosine to the low sensitivity of single-molecule real-time sequencing for detecting 5-methylcytosine. In most bacteria, a series of sequence motifs are targeted for methylation by DNA methyltransferases (MTases) at nearly all of these motifs in the genome (e.g., 5'-GmATC-3' by Dam or 5'-CmCWGG-3' by Dcm in E. coli), and only a small fraction of these motif sites remained unmethylated (Beaulaurier et al., 2019). Furthermore, when an IPD-based method was used to classify the methylation status of the second C in the 5'-CCWGG-3' motif, the detection rates for 5-methylcytosines were 95.2% and 1.9% with and without Tet protein treatment, respectively (Clark et al. 2013). Overall, IPD methods without prior base conversion (e.g., using Tet protein) missed the majority of 5-methylcytosines.
[0049] The aforementioned studies (Clark et al., 2012, Clark et al., 2013, Blow et al., 2016) used IPD-based algorithms without considering the sequence context in which the candidate base modifications are located. Other groups have attempted to consider the sequence context of nucleotides for base modification detection. For example, Feng et al. (2013) used a hierarchical model to analyze IPDs for the detection of 4-methylcytosine and 6-methyladenosine in their respective sequence contexts. However, their method only considered the IPD of the target base and its adjacent sequence contexts, and did not use the IPD information of all neighboring bases adjacent to the target base. Furthermore, the PW was not considered in the algorithm, and no data on the detection of 5-methylcytosine was presented.
[0050] In another study, Schadt et al. developed a statistical method called conditional random fields to analyze the IPD information of a target base and neighboring bases to determine whether the target base is 5-methylcytosine (Schadt et al., 2012). In this study, they considered the IPD interactions between these bases by inputting them into the equation. However, they did not input the nucleotide sequence, i.e., A, T, G, or C, into their equation. When they applied this method to determine the methylation status of the M.Sau3AI plasmid, the area under the ROC curve was close to 0.5, even with 800-fold sequence coverage of the plasmid sequence. Furthermore, in their method, they did not consider PW in their analysis.
[0051] Yet another study by Beckman et al. compared the IPDs of all sequences sharing the same 4-nt or 6-nt motif within a genome (e.g., obtained through whole-genome amplification) between a target bacterial genome and a completely unmethylated genome (Beckman et al. 2014). The goal of such an analysis was simply to identify motifs more frequently affected by base modifications. In this study, they considered only the IPDs of potentially modified bases, but not those of neighboring bases or PWs. Their method was not informative about the methylation status of individual nucleotides.
[0052] In summary, these previous attempts, either utilizing IPD alone or combining IPD with sequence information from neighboring nucleotides to group data, failed to determine the 5-methylcytosine base modification with meaningful or practical accuracy. In a recent review by Gouil et al., the authors concluded that detection of 5-methylcytosine in single molecules using single-molecule real-time sequencing is inaccurate due to low signal-to-noise ratios (Gouil et al., 2019). These previous studies leave it unclear whether using dynamic features for whole-genome methylomic analysis is feasible, especially for complex genomes such as human, cancer, and fetal genomes.
[0053] In contrast to previous studies, some embodiments of the methods described herein are based on measuring and utilizing the IPD, PW, and sequence context for all bases within the measurement window. The inventors believed that accurate measurement of base modifications (e.g., mC detection) with single-base resolution could be achieved if multiple metrics could be combined, for example, by simultaneously utilizing features including upstream and downstream sequence context, strand information, IPD, pulse width, and pulse intensity. Sequence context refers to the base composition (A, C, G, or T) and base order in a stretch of DNA. Such a stretch of DNA may surround a base or a target base undergoing base modification analysis. In one embodiment, the stretch of DNA may be proximal to the base undergoing base modification analysis. In another embodiment, the stretch of DNA may be distant from the base undergoing base modification analysis. The stretch of DNA may be upstream and / or downstream of the base undergoing base modification analysis.
[0054] In one embodiment, the upstream and downstream sequence context, strand information, IPD, pulse width, and pulse intensity features used in base modification analysis are referred to as kinetic features.
[0055] The embodiments present in this disclosure can be used with DNA obtained from, but not limited to, cell lines, samples from organisms (e.g., solid organs, solid tissues, samples obtained via endoscopy, blood, or plasma or serum or urine from pregnant women, chorionic villus biopsies, etc.), samples obtained from the environment (e.g., bacteria, cellular contaminants), and food (e.g., meat). In some embodiments, the methods present in this disclosure can also be applied after a step in which a portion of the genome is first enriched, for example, using hybridization probes (Albert et al., 2007; Okou et al., 2007; Lee et al., 2011), or approaches based on physical separation (e.g., based on size) or following restriction enzyme digestion (e.g., MspI), or Cas9-based enrichment (Watson et al., 2019). While enzymatic or chemical conversion is not required for the present invention to function, in certain embodiments, such a conversion step may be included to further enhance the performance of the present invention.
[0056] Embodiments of the present disclosure enable improved accuracy, practicality, or convenience in detecting base modifications or measuring modification levels. Modifications can be detected directly. Embodiments can avoid enzymatic or chemical conversions that may not preserve all modification information for detection. Furthermore, certain enzymatic or chemical conversions may be incompatible with certain types of modifications. Embodiments of the present disclosure can also avoid PCR amplification, which may not transfer base modification information to PCR products. Furthermore, both strands of DNA can be sequenced together, thereby allowing pairing of sequences from one strand with sequences complementary to the other strand. In contrast, PCR amplification separates the two strands of double-stranded DNA, making such sequence pairing difficult.
[0057] The determined methylation profile, with or without enzymatic or chemical conversion, can be used to analyze biological samples. In one embodiment, the methylation profile can be used to detect the origin of cellular DNA (e.g., maternal or fetal, tissue, virus, or tumor). Detection of abnormal methylation profiles in tissues can help identify developmental disorders in individuals, as well as identify and predict tumors and malignancies. Imbalances in methylation levels between haplotypes can be used to detect disorders, including cancer. Single-molecule methylation patterns can identify chimeric DNA (e.g., between viruses and humans) and hybrid DNA (e.g., between two genes that are not normally fused in the native genome) or between two species (e.g., due to genetic or genomic manipulation).
[0058] Methylation analysis can be improved through enhanced training, including narrowing the data used in the training set. Specific regions may be targeted for analysis. In embodiments, such targeting can include an enzyme that, alone or in combination with other reagent(s), can cleave a DNA sequence or genome based on its sequence. In some embodiments, the enzyme is a restriction enzyme that recognizes and cleaves specific DNA sequence(s). In other embodiments, two or more restriction enzymes with different recognition sequences can be used in combination. In some embodiments, a restriction enzyme may cleave or not cleave based on the methylation status of the recognition sequence. In some embodiments, the enzyme is an enzyme within the CRISPR / Cas family. For example, genomic regions of interest can be targeted using the CRISPR / Cas9 system or other guide RNA-based systems (i.e., short RNA sequences that bind to complementary target DNA sequences and, in the process, guide the enzyme to the target genomic location to act). In some cases, methylation analysis may be possible without aligning to a reference genome.
[0059] I. Methylation detection by single-molecule real-time sequencing The embodiments of the present disclosure allow for direct detection of base modifications without enzymatic or chemical conversion.The dynamic characteristics (e.g., sequence context, IPD, PW) obtained through single molecule real-time sequencing can be analyzed by machine learning to develop a model for detecting modifications or detecting the absence of modifications.The modification level can be used to determine the origin of DNA molecules or the presence or level of disorders.
[0060] To illustrate, using Pacific Biosciences SMRT sequencing as an example of single-molecule real-time sequencing, a DNA polymerase molecule is placed at the bottom of a well that acts as a zero-mode waveguide (ZMW). A ZMW is a nanophotonic device for confining light to a small observation volume. This hole has a very small diameter, which does not allow the propagation of light in the wavelength range used for detection. Only the emission of an optical signal from dye-labeled nucleotides incorporated by the immobilized polymerase is detectable against a low, constant background signal (Eid et al., 2009). The DNA polymerase catalyzes the incorporation of fluorescently labeled nucleotides into the complementary nucleic acid strand.
[0061] Figure 1 shows examples of molecules with base modifications sequenced by single-molecule circular consensus sequencing. Molecules 102, 104, and 106 have base modifications. A DNA molecule (e.g., molecule 106) can be ligated with a hairpin adapter to form ligated molecule 108. Ligated molecule 108 can then form circularized molecule 110. The circularized molecule can bind to an immobilized DNA polymerase and initiate DNA synthesis. Molecules without base modifications can also be sequenced.
[0062] Figure 2 shows an example of a molecule with methylated and / or unmethylated CpG sites sequenced by single-molecule real-time sequencing. First, a DNA molecule is ligated to a hairpin adapter to form a circularized molecule, which will bind to an immobilized DNA polymerase and initiate DNA synthesis. In Figure 2, DNA molecule 202 is ligated to the hairpin adapter to form ligated molecule 204. Ligated molecule 204 then forms circularized molecule 206. Molecules without CpG sites can also be sequenced. Circular molecule 206 contains an unmethylated CpG site 208, which can still be sequenced.
[0063] As DNA synthesis begins, fluorescent dye-labeled nucleotides are incorporated into newly synthesized strands by the immobilized polymerase based on the circular DNA template, leading to the emission of a light signal. Because the DNA template is circularized, the entire circular DNA template is passed through the polymerase multiple times (i.e., one nucleotide of the DNA template is sequenced multiple times). The sequence generated from the process of passing every base of the circularized DNA template completely through the DNA polymerase is called a subread. Because the polymerase can continue through the entire circular DNA template multiple times, one molecule in the ZMW generates multiple subreads. In one embodiment, a subread may contain only a subset of the sequence, base modifications, or other molecular information of the circular DNA template due to the presence of sequencing errors.
[0064] As shown in Figure 3, the arrival time and duration of the resulting fluorescent pulses allow for measuring polymerase kinetics. The interpulse distance (IPD) is a metric for the length of the period between two emitted pulses, each of which indicates a fluorescently labeled nucleotide incorporated into the nascent strand (Figure 3). As shown in Figure 3, the pulse width (PW) is another metric that reflects polymerase kinetics in relation to the duration of the pulse associated with base calling. PW can be the duration of the pulse at 0% of the signal peak height (i.e., the fluorescence intensity of the incorporated dye-labeled nucleotide). In one embodiment, PW can be defined, for example, but not limited to, by the duration of the pulse at 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, or 90% of the signal peak height. In some embodiments, PW can be the area under the peak divided by the signal peak height.
[0065] Such polymerase kinetics, including IPD, has been shown to be affected by base modifications such as N6-methyladenine (6mA), 5-methylcytosine (5mC), and 5-hydroxymethylcytosine (5hmC) in synthetic and microbial sequences (e.g., E. coli) (Flusberg et al., 2010). Flusberg et al. (2010) did not use sequence context and IPD as independent inputs to detect modifications, resulting in a model that lacked meaningful detection accuracy. Flusberg et al. confirmed that 6mA occurred at GATC using sequence context alone. Flusberg et al. did not mention using sequence context in combination with IPD as an input to detect methylation status.
[0066] The weak interruption given to the incorporation of a new base into 5-methylcytosine on the complementary strand is the methylation motif C. mThe reported detection of CWGG ranges from only 1.9% to 4.3% (Clark et al., 2013), making it extremely difficult to determine methylation even in relatively simple microbial genomes using IPD signals alone. For example, the analytical software package provided by Pacific Biosciences (SMRT Link v6.0.0) cannot perform 5mC analysis. Furthermore, the previous version of SMRT Link v5.1.0 required the conversion of 5mC to 5-carboxyl cytosine (5caC) using the Tet1 enzyme prior to methylation analysis. This is because the IPD signal associated with 5caC is enhanced (Clark et al., 2013). Therefore, it is not surprising that there have been no studies demonstrating the feasibility of analyzing native DNA in a genome-wide manner for the human genome using single-molecule real-time sequencing.
[0067] II. Measurement window patterns and machine learning models A technique for detecting base modifications without enzymatically or chemically converting the modifications and / or bases is desirable. As described herein, modifications of a target base can be detected using kinetic feature data obtained from single-molecule real-time sequencing of the bases surrounding the target base. Kinetic features can include pulse intervals, pulse widths, and sequence context. These kinetic features can be obtained for a measurement window of a specific number of nucleotides upstream and downstream of the target base. These features (e.g., the specific location of the measurement window) can be used to train a machine learning model. As an example of sample preparation, two strands of a DNA molecule can be joined by a hairpin adapter, thereby forming a circular DNA molecule. The circular DNA molecule can then acquire kinetic features of either or both the Watson and Crick strands. A data analysis framework can be developed based on the kinetic features of the measurement window. This data analysis framework can then be used to detect modifications, including methylation. This section describes various techniques for detecting modifications.
[0068] A. Use of Single Strands As an example, a subread of the Watson strand was obtained from Pacific Biosciences SMRT sequencing to analyze one specific base for base modification status, as shown in FIG. 4. In FIG. 4, three bases from each side of the base subjected to base modification analysis would be defined as a measurement window 400. In one embodiment, the sequence context, IPD, and PW for these seven bases (i.e., three nucleotides (nt) upstream and downstream sequence and one nucleotide for base modification analysis) were compiled into a two-dimensional (i.e., 2-D) matrix as the measurement window. In the example shown, the measurement window 400 is for one subread of the Watson strand. Other variations are described herein.
[0069] The first row 402 of the matrix indicates the sequence being investigated. In the second row 404 of the matrix, the 0 position represented the base for base modification analysis. The relative positions -1, -2, and -3 indicated positions 1 nt, 2 nt, and 3 nt upstream of the base subjected to base modification analysis, respectively. The relative positions +1, +2, and +3 indicated positions 1 nt, 2 nt, and 3 nt downstream of the base subjected to base modification analysis, respectively. Each position included two columns containing corresponding IPD and PW values. The next four rows (rows 408, 412, 416, and 420) corresponded to the four types of nucleotides (A, C, G, and T) of a strand (e.g., the Watson strand), respectively. The IPD and PW values present in the matrix depended on which corresponding nucleotide type was sequenced at a particular position. As shown in Figure 4, at relative position 0, the IPD and PW values are displayed in the row showing "G" in the Watson strand, indicating that guanine was called in the sequence results at that position. Other grids in the column that did not correspond to the sequenced base are coded as "0." As an example, the sequence information corresponding to the 2D digital matrix (Figure 4) is 5'-GATGACT-3' for the Watson strand.
[0070] As shown in one embodiment illustrated in FIG. 5, a measurement window can be applied to data from a click strand. To analyze one specific base for its base modification status, subreads of the click strand were obtained from single-molecule real-time sequencing. In FIG. 5, three bases from each side of the base subjected to base modification analysis and the base subjected to base modification analysis would be defined as the measurement window. In one embodiment, the sequence context, IPD, and PW for these seven bases (i.e., three nucleotides (nt) upstream and downstream sequence and one nucleotide for base modification analysis) were compiled into a two-dimensional (i.e., 2-D) matrix as the measurement window. The first row of the matrix indicates the interrogated sequence. In the second row of the matrix, position 0 represents the base for base modification analysis. Relative positions -1, -2, and -3 indicated positions 1 nt, 2 nt, and 3 nt upstream of the base subjected to base modification analysis, respectively. Relative positions +1, +2, and +3 indicated positions 1 nt, 2 nt, and 3 nt downstream of the base subjected to base modification analysis, respectively. Each position contains two columns containing the corresponding IPD and PW values. The next four rows correspond to the four nucleotide types (A, C, G, and T) of this strand (e.g., the Click strand). The IPD and PW values present in the matrix depended on which corresponding nucleotide type was sequenced at a particular position. As shown in Figure 5, at relative position 0, the IPD and PW values are displayed in the row representing "T" in the Click strand, indicating that thymine was called in the sequence results at that position. Other grids in the columns that did not correspond to sequenced bases are coded as "0." As an example, the sequence information corresponding to the 2D digital matrix (Figure 5) is 5'-ACTTAGC-3' for the Click strand.
[0071] B. Use of both Watson and Crick strands FIG. 6 illustrates an embodiment in which a measurement window can be implemented in a manner that allows data from the Watson strand and its complementary Crick strand to be combined. As shown in FIG. 6, subreads of the Watson strand and the Crick strand were obtained from single-molecule real-time sequencing and analyzed for modifications of one specific base. In one embodiment, the measurement window from the Crick strand of the circular DNA template was complementary to the measurement window from the Watson strand subjected to base modification analysis. In FIG. 6, the first base and three bases on either side of the first base of the Watson strand subjected to base modification analysis would be defined as the first measurement window. The second base and three bases on either side of the second base of the Crick strand would be defined as the second measurement window. The second base was complementary to the first base. In one embodiment, the sequence context, IPD, and PW for these seven bases (i.e., three nucleotides (nt) upstream and downstream sequence and one nucleotide for base modification analysis) from the Watson and Crick strands were compiled into a two-dimensional (i.e., 2-D) matrix. These measurement windows from the Watson strand and the Crick strand were considered as the first and second measurement windows, respectively.
[0072] The first row of the Watson and Crick strand matrices indicates the interrogated sequence. In the second row of the Watson strand matrix, position 0 represents the first base for base modification analysis. Position 0 in the second row of the Crick strand matrix represents the second base complementary to the first base. Relative positions -1, -2, and -3 represent positions 1 nt, 2 nt, and 3 nt upstream of the first and second bases, respectively. Relative positions +1, +2, and +3 represent positions 1 nt, 2 nt, and 3 nt downstream of the first and second bases, respectively. Each position from the Watson and Crick strands corresponds to two columns containing the corresponding IPD and PW values. The next four rows of the Watson and Crick strand matrices correspond to the four nucleotides (A, C, G, and T) of a particular strand (e.g., Crick strand), respectively. The IPD and PW values present in the matrix depended on which corresponding nucleotide type was sequenced at a particular position.
[0073] As shown in Figure 6, at relative position 0, the IPD and PW values are shown in the row showing "A" in the Watson strand and "T" in the Crick strand, indicating that adenine and thymine were called in the sequence results at that position in the Watson strand and Crick strand, respectively. Other grids in the column that did not correspond to the sequenced base are coded as "0." As an example, the sequence information corresponding to the 2D digital matrix of the Watson strand (Figure 6) would be 5'-ATAAGTT-3'. The sequence information corresponding to the 2D digital matrix of the Crick strand (Figure 6) would be 5'-AACTTAT-3'.
[0074] As shown in this example, data from the Watson and Crick strands can be combined to form a new matrix, which can also be considered a measurement window. This new matrix can be used as a single sample used to train a machine learning model. Thus, while the specific arrangement of the 2D matrix may have an impact, such as when a convolutional neural network (CNN) is used, all values in the new matrix can be treated as separate features. Sequence context at various positions in different strands can be conveyed via non-zero entries in the matrix.
[0075] Figure 7 shows that the measurement windows can be implemented in a way that the data from the Watson strand and the Crick strand are not exactly complementary to each other. As shown in Figure 7, the first measurement window was 5'-ATAAGTT-3' and the second measurement window was 5'-GTAACGC-3'. In some embodiments, the Watson strand and the Crick strand may be shifted relative to each other so that the positions are not complementary.
[0076] Figure 8 shows that the measurement window can be used to analyze the methylation status of a CpG site. Position 0 corresponds to the cytosine of the CpG site, and therefore, there is a one-position shift between the two strands, resulting in C being at position 0 for both strands. Therefore, only a portion of the sequences included in the measurement window from the Watson strand and the Crick strand are complementary to each other. In other embodiments, all sequences in the measurement window from the Watson strand and the Crick strand may be complementary to each other. In still other embodiments, none of the sequences in the measurement window from the Watson strand and the Crick strand are complementary to each other.
[0077] In one embodiment, for a measurement window, the length of the DNA stretch surrounding the base subjected to base modification analysis can be asymmetric, for example, X-nt upstream and Y-nt downstream of the base can be used for base modification analysis. X may include, but is not limited to, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 100, 150, 200, 300, 400, 500, 1000, 2000, 4000, 5000, and 10000. Y can include, but is not limited to, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 100, 150, 200, 300, 400, 500, 1000, 2000, 4000, 5000, and 10000.
[0078] C. Model training and modification detection Figure 9 shows a general procedure for determining any base modifications using a measurement window. Unmodified and known modified DNA samples were subjected to single-molecule real-time sequencing. Modified DNA (e.g., modified molecule 902) means that the base (e.g., base 904) has a modification (e.g., methylation) at that site. Unmodified DNA (e.g., unmodified molecule 906) means that the base (e.g., base 908) does not have a modification at that site. Both sets of DNA can be artificially created or manipulated to form modified / unmodified DNA.
[0079] At stage 910, the sample can then undergo single molecule real-time sequencing. As part of SMRT sequencing, the circular molecule can be sequenced multiple times by repeatedly passing the immobilized DNA polymerase. The sequence information obtained each time is considered a subread. Thus, one circular DNA template generates multiple subreads. The sequencing subreads can be aligned to a reference genome using, for example, but not limited to, BLASR (Mark J Chaisson et al., BMC Bioinformatics. 2012;13:238). In various other embodiments, BLAST (Altschul SF et al, J Mol Biol. 1990; 215(3): 403-410), BLAT (Kent WJ, Genome Res. 2002; 12(4): 656-664), BWA (Li H et al, Bioinformatics. 2010; 26(5): 589-595), NGMLR (Sedlazeck FJ et al, Nat Methods. 2018; 15(6): 461-468), LAST (Kielbasa SM et al, Genome Res. 2011; 21(3): 487-493), and Minimap2 (Li H, Bioinformatics. 2018; 34(18): 3094-3100) can be used to align subreads to a reference genome. Alignment allows identifying data from each subread at the same position, allowing data from multiple subreads to be combined (e.g., averaged).
[0080] In stage 912, the IPD, PW, and sequence context surrounding the base subjected to base modification analysis were obtained from the alignment results. In stage 914, the IPD, PW, and sequence context were recorded in a specific structure, for example, but not limited to, a 2D matrix as shown in FIG.
[0081] In stage 916, analytical, computational, mathematical, or statistical model(s) are trained using several 2D matrices containing molecules from the reference kinetic pattern with known base modifications. In stage 918, the statistical model resulting from the training is developed. For simplicity, Figure 9 shows only the statistical model developed by training, but any model or data analysis framework can be developed. Examples of data analysis frameworks include machine learning models, statistical models, and mathematical models. Statistical models include, but are not limited to, linear regression, logistic regression, deep recurrent neural networks (e.g., long short-term memory, LSTM), Bayesian classifiers, hidden Markov models (HMMs), linear discriminant analysis (LDA), k-means clustering, density-based spatial clustering for applications with noise (DBSCAN), random forest algorithms, and support vector machines (SVMs). The DNA stretch surrounding the base subjected to base modification analysis can be the X-nt upstream and Y-nt downstream of that base, i.e., the "measurement window."
[0082] Because the correct output (i.e., modification state) is known, a data structure can be used in the training process. For example, the IPD, PW, and sequence context corresponding to 3 nt upstream and downstream of the base from the Watson and / or Crick strand(s) can be used to construct a 2D matrix used to train statistical model(s) for classifying base modifications. In this way, training can provide a model that can classify base modifications at nucleic acid positions with a previously known state.
[0083] Figure 10 shows a general procedure for how a statistical model(s) trained from a DNA sample with known base modification status can detect base modifications. A sample with unknown base modification status was subjected to SMRT sequencing. The sequencing subreads were aligned to a reference genome, for example, using the techniques described above. Additionally or alternatively, the subreads can be aligned to each other. Still other embodiments can use only one subread or analyze them independently, such that alignment is not performed.
[0084] For bases subjected to base modification analysis, the IPD, PW, and sequence context were obtained and associated with the base from the Watson and / or Crick strand(s) of the alignment results using the same measurement window as used in the training step (FIG. 9). In another embodiment, the measurement window between the training and testing procedures may be different. For example, the size of the measurement window between the training and testing procedures may be different. These IPD, PW, and sequence context are converted into a 2D matrix. This 2D matrix of the test sample may be compared with the reference kinetic features to determine base modifications. For example, the 2D matrix of the test sample may be compared with the reference kinetic features through statistical model(s) learned from the training sample, thereby enabling the determination of base modifications at sites of the nucleic acid molecule of the test sample. Statistical models include, but are not limited to, linear regression, logistic regression, deep recurrent neural networks (e.g., long short-term memory, LSTM), Bayesian classifiers, hidden Markov models (HMM), linear discriminant analysis (LDA), k-means clustering, density-based spatial clustering for applications with noise (DBSCAN), random forest algorithms, and support vector machines (SVM).
[0085] Figure 11 shows the general procedure for how a method for classifying methylation status at CpG sites can be created. DNA samples with known unmethylated and methylated CpG sites were subjected to single molecule real-time sequencing. Sequencing subreads were aligned to a reference genome. Watson strand data was used.
[0086] From the alignment results, the IPD, PW, and sequence context surrounding the cytosine at the CpG site subjected to methylation analysis were obtained and recorded in a specific structure, for example, but not limited to, a 2D matrix as shown in Figure 11. Several 2D matrices containing reference kinetic patterns derived from molecules with known methylation states were used to train statistical model(s). The stretch of DNA surrounding the base under investigation can be X-nt upstream and Y-nt downstream of that base, i.e., the "measurement window." X may include, but is not limited to, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 100, 150, 200, 300, 400, 500, 1000, 2000, 4000, 5000, and 10000. Y can include, but is not limited to, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 100, 150, 200, 300, 400, 500, 1000, 2000, 4000, 5000, and 10000. In one embodiment, the IPD, PW, and sequence context corresponding to 3 nt upstream and downstream of the base from the Watson strand can be used to construct a 2D matrix used to train statistical model(s) for classifying base modifications.
[0087] Figure 12 shows the general procedure for classifying the methylation status of an unknown sample. Samples with unknown methylation status were subjected to single-molecule real-time sequencing. Sequencing subreads were aligned to a reference genome.
[0088] For cytosines at CG sites in the alignment results, the IPD, PW, and sequence context were obtained from the Watson strand using the equivalent measurement window applied in the training step (Figure 11) to associate the modification with the base under investigation. These IPD, PW, and sequence context can be converted into a 2D matrix. Such a 2D matrix of the test sample will be compared with the reference kinetic pattern shown in Figure 11 to determine the methylation status. X11
[0089] 13 and 14 show that, similar to the procedure using the Watson chain, the kinetic features from the Crick chain can be used for the training and testing procedures, as detailed above. The statistical model(s) can be the same model or different models. If they are different models, they can be used to obtain independent classifications, which can be compared, for example, to identify a modified state if they match. Then, if they do not match, an unclassified state can be identified. If they are the same model, the data can be combined into a single data structure, for example, the matrix of FIG. 6.
[0090] Figures 15 and 16 show that kinetic features from both the Watson and Crick strands can be used for training and testing procedures, as detailed above. DNA samples with known unmethylated and methylated CpG sites were subjected to single-molecule real-time sequencing. Sequencing subreads were aligned to a reference genome, although subreads can also be aligned to each other, as described elsewhere herein.
[0091] For each subread of the alignment result, the IPD, PW, and sequence context surrounding the cytosine of the CpG site subjected to methylation analysis were obtained. Because the DNA molecule was circularized through the use of two hairpin adapters (e.g., following the SMRTBell template preparation protocol), the circular molecule could be sequenced more than once, thereby generating multiple subreads of the molecule. The subreads can be used to generate circular consensus sequence (CCS) reads. Generally, in all methods described herein, one ZMW can generate multiple subreads, but only corresponds to one CCS read.
[0092] In some embodiments, a fully unmethylated dataset can be generated by PCR on human DNA fragments. For example, a fully methylated dataset can be generated through treatment of human DNA fragments with the CpG methyltransferase M.SssI, in which all CpG sites are assumed to be methylated. In other examples, a different CpG methyltransferase, such as M.MpeI, can be used. In other embodiments, synthetic sequences with known methylation states or existing DNA samples with different methylation levels, or hybrid methylation states created by restriction enzyme digestion and subsequent ligation of methylated and unmethylated DNA molecules (resulting in a chimeric proportion of methylated / unmethylated DNA molecules), can be used to train a methylation predictive model or classifier.
[0093] The transformation of the kinetic pattern, including sequence context, IPD, and pulse width (PW), can be converted into a 2D matrix containing features from the Watson and Crick strands to analyze the methylation status of CG sites, as shown in Figure 15. This approach accurately captures subtle kinetic changes resulting from methylated cytosines and their nearby sequence context. Similar to any of the various methods described herein, for each CpG present in a subread, a measurement window (e.g., three bases upstream and downstream of the cytosine at the CpG site) can be used for subsequent analysis. Thus, a total of seven nucleotides (including the cytosine at the CpG site) are analyzed together. The IPD and PW can be calculated for each base between those seven nucleotides. To capture the sequence context resulting from kinetic changes, the IPD and PW signals can be compiled into specific base calls, relative sequencing positions, and strand information, as shown in Figure 15. Such a data structure is simply referred to as a 2D digital matrix of kinetics.
[0094] Such a 2D digital matrix resembles a "2D digital image." For example, the first row of the 2D digital matrix contained the relative positions of the cytosines of the CpG loci subjected to methylation analysis, along with the 3 nt upstream and downstream positions of the cytosine site. The 0 position represented the cytosine site for which methylation was determined. The -1 and -2 relative positions indicated 1 nt upstream and 2 nt upstream of the cytosine in question. The +1 and +2 relative positions indicated 1 nt downstream and 2 nt downstream of the cytosine used. Each position would correspond to two columns containing the corresponding IPD and PW values. Each row corresponded to four types of nucleotides (A, C, G, and T) in the Watson and Crick strands. The input of IPD and PW values in the matrix depended on the corresponding nucleotide type preset in the sequence results (i.e., subreads) at a particular position.
[0095] As shown in Figure 15, at a relative position of 0, the IPD and PW values are shown in the "C" row of the Watson strand, indicating that a cytosine was called at that position. Other grids in the column that did not correspond to a sequenced base are coded as "0." As an example, the sequence information corresponding to the 2D digital matrix (Figure 15) is 5'-ATACGTT-3' and 5'-TAACGTA-3' for the Watson strand and the Crick strand, respectively. In this context, the upstream and downstream sequences adjacent to the cytosine at the CpG site in the Watson strand and the Crick strand are different. Because methylation at CpG sites is symmetric between the Watson strand and the Crick strand (Lister et al., 2009), in one preferred embodiment, the dynamics of both strands were used to train a methylation prediction model. In another embodiment, the Watson strand and the Crick strand can be used separately to train a methylation prediction model.
[0096] Given the high data throughput of single-molecule real-time sequencing, in one embodiment, a deep learning algorithm (convolutional neural network (CNN)) (LeCun et al., 1989) may be suitable for distinguishing methylated from unmethylated CpGs. Other algorithms can be used additionally or alternatively, including, but not limited to, linear regression, logistic regression, deep recurrent neural networks (e.g., long short-term memory, LSTM), Bayesian classifiers, hidden Markov models (HMMs), linear discriminant analysis (LDA), k-means clustering, density-based spatial clustering for applications with noise (DBSCAN), random forest algorithms, and support vector machines (SVMs). As shown in Figures 6-8, Watson and Crick strands can be used separately or combined in new matrices for training.
[0097] Another transformation of the kinetic pattern can be an N-dimensional matrix, where N can be, for example, 1, 3, 4, 5, 6, or 7. For example, a 3D matrix can be a stack of 2D matrices layered according to the number of tandem CG sites in the DNA stretch being analyzed, with the third dimension being the number of tandem CG sites in that DNA stretch. In some embodiments, pulse intensity or pulse magnitude (e.g., measured by the pulse peak height or the area under the pulse signal) can also be incorporated into the matrix. Pulse intensity (a metric of pulse peak amplitude, Figure 3) can be added to the original 2D matrix in an additional column adjacent to the columns associated with PW and IPD values, or added to the third dimension to form a 3D matrix.
[0098] As a further example, an 8x21 2D matrix can be converted into a 1D matrix (i.e., a vector) containing 168 elements. This 1D matrix can then be scanned to perform, for example, CNN and other modeling. As another example, a method can split an 8x21 2D matrix into multiple smaller matrices, for example, two 4x21 2D matrices. These two smaller matrices can be combined vertically to obtain a 3D matrix (i.e., x=21, y=4, z=2). The method can scan the first 2D matrix, then the second 2D matrix, to form a data representation for machine learning. The data can be further split to form higher-dimensional matrices. Furthermore, secondary structure information can be added to the data structure, for example, by adding an additional matrix (1D matrix) on top of the 2D matrix. Such an additional matrix can encode whether each base in the measurement window is involved in a secondary structure (e.g., a stem-loop structure). For example, bases involved in the "stem" are coded as 0's and bases involved in the "loop" are coded as 1's.
[0099] In one embodiment, the methylation status of a CpG site within a single DNA molecule can be expressed as a probability of being methylated based on a statistical model, rather than providing a qualitative result of "methylated" or "unmethylated." A probability of 1 indicates that the CpG site can be considered methylated based on the statistical model. A probability of 0 indicates that the CpG site can be considered unmethylated based on the statistical model. In subsequent downstream analysis, a cutoff value can be used to classify whether a particular CpG site is classified as "methylated" or "unmethylated" based on probability. Possible cutoff values include 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 55%, 60%, 65%, 70%, 75%, 80%, 85%, 90%, or 95%. CpG sites with a probability of being methylated greater than a predetermined cutoff are classified as "methylated," and CpG sites with a probability of being methylated less than the predetermined cutoff are classified as "unmethylated." The desired cutoff can be obtained from a training dataset, for example, using receiver operating characteristic (ROC) curve analysis.
[0100] Figure 16 shows the general procedure for classifying the methylation status of an unknown sample from the Watson and Crick strands. Samples with unknown methylation status were subjected to single-molecule real-time sequencing. The sequencing subreads can be aligned to a reference genome or to each other, as with other methods, to determine a consensus value (mean, median, mode, or other statistical value) for a given position. As shown, the measurements for the two strands can be combined into a single 2D matrix.
[0101] For cytosines in CpG sites in the alignment results, different window sizes can be used, but equivalent measurement windows (3 nt upstream and downstream of the cytosine in the CpG site) can be used to obtain IPD, PW, and sequence context from the Watson strand as applied in the training step relative to the base under investigation (Figure 16). Such a 2D matrix of the test sample can be compared with the reference kinetic pattern shown in Figure 16 to determine the methylation status.
[0102] III. Exemplary Model Training for Detecting Methylation To test the feasibility and validity of the proposed approach, placental DNA libraries were prepared using M.SssI treatment (methylated library) and PCR amplification (unmethylated library) prior to single-molecule real-time sequencing. We obtained 44,799,736 and 43,580,452 subreads for the methylated and unmethylated libraries, corresponding to 421,614 and 446,285 circular consensus sequences (CCS), respectively. As a result, each molecule was sequenced at a median frequency of 34x and 32x in the methylated and unmethylated libraries. Datasets were generated from DNA prepared with the Pacific Biosciences Sequel Sequencing Kit 3.0. This kit was developed for use with the original Pacific Biosciences Sequel sequencer. Herein, we refer to the original Sequel as Sequel I to distinguish it from its successor, Sequel II. Therefore, Sequel Sequencing Kit 3.0 is referred to herein as Sequel I Sequencing Kit 3.0. Sequencing kits designed for the Sequel II sequencer include Sequel II Sequencing Kit 1.0 and Sequel II Sequencing Kit 2.0, which are also described in this disclosure.
[0103] Fifty percent of the sequencing molecules generated from the methylated and unmethylated libraries were used to train a statistical model (the remaining 50% were used for validation). In this case, it is a convolutional neural network (CNN) model. As an example, a CNN model can have one or more convolutional layers (e.g., 1D or 2D layers). A convolutional layer can use one or more different filters, each using a kernel that operates on matrix values local to (e.g., nearby or surrounding) a particular matrix element, thereby providing a new value for that particular matrix element. In one implementation, two 1D convolutional layers were used (each with 100 filters, each with a kernel size of 4). Filters can be applied individually and then combined (e.g., by weighted averaging). The resulting matrix can be smaller than the input matrix.
[0104] The convolutional layer is followed by a ReLU (rectified linear unit) layer, followed by a dropout layer with a dropout rate of 0.5. ReLU is an example of an activation function that manipulates individual values to obtain a new matrix (image) from the convolutional layer(s). Other activation functions (e.g., sigmoid, softmax, etc.) can also be used. One or more of such layers can be used. Dropout layers can be used in ReLU layers or max pooling layers and act as regularizers to prevent overfitting. Dropout layers can be used during the training process to ignore different (e.g., random) values during various iterations of the optimization process performed as part of training (e.g., to reduce cost / loss functions).
[0105] A max-pooling layer (e.g., pool size 2) can be used after the ReLU layer. A max-pooling layer functions similarly to a convolutional layer, except that instead of obtaining a dot product between the input and the kernel, it obtains the maximum value of the region from the input that overlaps with the kernel. Additional convolutional layers can be used. For example, the data from the pooling layer can be input to two other 1D convolutional layers (e.g., each with 128 filters with a kernel size of 2 followed by a ReLU layer), and a dropout layer with a dropout rate of 0.5 can be used. A max-pooling layer with a pool size of 2 was used. Finally, a fully connected layer (e.g., with 10 neurons followed by a ReLU layer) can be used. An output layer with one neuron can be followed by a sigmoid layer, which yields the methylation probability. Various settings for layers, filters, and kernel sizes can be adjusted. This training dataset used 468,596 and 432,761 CpG sites from methylated and unmethylated libraries.
[0106] A. Results on training and test datasets Figure 17A shows the methylation probability for each CpG site of each single DNA molecule in the training dataset. The methylation probability was much higher in the methylated library than in the unmethylated library. When the methylation probability cutoff was 0.5, 94.7% of the unmethylated CpG sites were correctly predicted to be unmethylated, and 84.7% of the methylated CpGs were correctly predicted to be methylated.
[0107] Figure 17B shows the performance of the test dataset. Using the model trained with the training dataset, we predicted the methylation status of 469,729 and 432,024 CpG sites in independent test datasets from methylated and unmethylated libraries. When the cutoff for the probability of being methylated was 0.5, 94.0% of the unmethylated CpG sites were correctly predicted to be unmethylated, and 84.1% of the methylated CpG sites were correctly predicted to be methylated. These results suggested that the use of new transformations of kinetics in combination with sequence context may enable the determination of the methylation status of DNA (e.g., from human subjects).
[0108] The ability of each feature (sequence context, IPD, and PW) in predicting CpG methylation status was assessed by including a subset of features in the model. In the training dataset, models with (i) sequence context only, (ii) IPD only, and (iii) PW only yielded area under the curve (AUC) values of 0.5, 0.74, and 0.86, respectively. Combining IPD and sequence context improved performance, with an AUC of 0.86. A combined analysis of sequence context ("Seq"), IPD, and PW significantly improved performance, with an AUC of 0.94 (Figure 18A). Performance on an independent test dataset was comparable to the training dataset (Figure 18B).
[0109] The subread depth of a CpG site was defined as the average number of subreads covering that site and its surrounding 10 bp. As shown in Figures 19A and 19B, the higher the subread depth of a CpG site, the higher the accuracy of methylation detection achieved. For example, as shown in the test dataset (Figure 19B), when the depth of each CpG site is at least 10, the AUC for predicting methylation status is 0.93. However, when the subread depth of each CpG site is at least 300, the AUC for predicting methylation status is 0.98. Meanwhile, even when the depth is 1, an AUC of 0.9 is achieved. This indicates that our approach can achieve methylation prediction even at low sequencing depths.
[0110] To test the effect of strand information on the performance of methylation analysis, sequence contexts derived from the Watson and Crick strands, IPD and PW, were used for training according to embodiments present in this disclosure. Figures 20A and 20B show that using a single strand, i.e., either the Watson or Crick strand, for training and testing is feasible, as AUCs of up to 0.91 and 0.87 can be achieved in the training and testing datasets. Using both strands (e.g., as illustrated in Figures 6-8), including the Watson and Crick strands, yielded the best performance (AUCs of 0.94 and 0.90 in the training and testing datasets, respectively), suggesting that strand information is important for achieving optimal performance.
[0111] According to the embodiments developed in the present disclosure, different numbers of nucleotides upstream and downstream of the CpG site were further tested to study how this parameter affects performance. Figures 21A and 21B show that the number of nucleotides upstream and downstream of a cytosine in a CpG context affects the accuracy of methylation prediction. For example, for illustrative purposes, considering, but not limited to, 2 nucleotides (nt), 3 nt, 4 nt, 6 nt, 8 nt, 10 nt, 15 nt, and 20 nt upstream and downstream of the investigated cytosine, the AUC of the method using 2 nt upstream and downstream of the investigated cytosine is only 0.50 in both the training and test datasets, while the AUC of the method using 15 nt upstream and downstream of the investigated cytosine increases to 0.95 and 0.92. These results suggest that optimal performance can be found by varying the length of the upstream and downstream regions adjacent to the analyzed cytosine. In one embodiment, as shown in Figure 21B, 3 nt upstream and downstream of the cytosine can be used to determine the methylation status, achieving an AUC of 0.89.
[0112] In one embodiment, to perform the analysis according to the present embodiment of the present disclosure, asymmetric sequences adjacent to the investigated cytosine can be used. For example, 2 nt upstream can be used in combination with 1 nt, 3 nt, 4 nt, 5 nt, 6 nt, 7 nt, 8 nt, 9 nt, 10 nt, 11 nt, 12 nt, 13 nt, 14 nt, 15 nt, 16 nt, 17 nt, 18 nt, 19 nt, 20 nt, 25 nt, 30 nt, 35 nt, and 40 nt downstream of cytosine, and 1 nt, 2 nt, 4 nt, 5 nt, 6 nt, 7 nt, 8 nt, 9 nt, 10 nt, 11 nt, 12 nt, 13 nt, 14 nt, 15 nt, 16 nt, 17 nt, 18 nt, 19 nt, 20 nt, 25 nt, 30 nt, 35 nt, and 40 nt downstream of cytosine. 3 nt upstream can be used in combination with 6 nt, 17 nt, 18 nt, 19 nt, 20 nt, 25 nt, 30 nt, 35 nt, and 40 nt downstream, and 4 nt upstream can be used in combination with 1 nt, 2 nt, 3 nt, 5 nt, 6 nt, 7 nt, 8 nt, 9 nt, 10 nt, 11 nt, 12 nt, 13 nt, 14 nt, 15 nt, 16 nt, 17 nt, 18 nt, 19 nt, 20 nt, 25 nt, 30 nt, 35 nt, and 40 nt downstream of cytosine. As another example, 2 nt downstream of cytosine can be used in combination with 1 nt, 3 nt, 4 nt, 5 nt, 6 nt, 7 nt, 8 nt, 9 nt, 10 nt, 11 nt, 12 nt, 13 nt, 14 nt, 15 nt, 16 nt, 17 nt, 18 nt, 19 nt, 20 nt, 25 nt, 30 nt, 35 nt, and 40 nt upstream of cytosine, Three nt downstream can be used in combination with 16 nt, 17 nt, 18 nt, 19 nt, 20 nt, 25 nt, 30 nt, 35 nt, and 40 nt upstream, and four nt downstream can be used in combination with 1 nt, 2 nt, 3 nt, 5 nt, 6 nt, 7 nt, 8 nt, 9 nt, 10 nt, 11 nt, 12 nt, 13 nt, 14 nt, 15 nt, 16 nt, 17 nt, 18 nt, 19 nt, 20 nt, 25 nt, 30 nt, 35 nt, and 40 nt upstream of cytosine.Utilizing the IPD, PW, strand information, and sequence context associated with the n-nt upstream and m-nt downstream of a cytosine can provide improved accuracy in determining methylation status in certain embodiments. Such different measurement windows can be applied to the analysis of other types of base modifications, such as 5hmC, 6mA, 4mC, and oxoG, or any of the modifications disclosed herein. Such different measurement windows can include the analysis of DNA secondary structures, such as G-quadruplexes and stem-loop structures. Examples of such structures are described above. Such secondary structure information can also be added as separate columns in a matrix.
[0113] 22A and 22B show that it is feasible to determine methylation status using kinetic patterns associated with only bases at least three downstream. According to embodiments present in the present disclosure, using features associated with cytosine and its downstream 3, 4, 6, 8, and 10 bases, the AUCs for determining methylation status in the training dataset were 0.91, 0.92, 0.94, 0.94, and 0.94, respectively, and in the test dataset, the AUCs were 0.87, 0.88, 0.90, 0.90, and 0.90, respectively.
[0114] However, Figures 23A and 23B show that when only features related to upstream bases are used, the ability to discriminate methylation status appears to be reduced: in the training and test datasets, the AUCs were all 0.50 for 2 to 10 upstream bases.
[0115] Figures 24 and 25 show that different combinations of upstream and downstream bases can achieve optimal classification when determining methylation status. For example, features associated with 8 bases upstream and 8 bases downstream of cytosine achieved the best performance in this dataset, with AUCs of 0.94 and 0.91 for the training and test datasets, respectively.
[0116] Figure 26 shows the relative importance of features for classifying methylation status at CpG sites. The "W" and "C" in parentheses indicate strand information, with "W" indicating Watson strand and "C" indicating Crick strand. The importance of each feature, including sequence context, IPD, and PW, was determined using random forest. Random forest tree analysis showed that the feature importance of IPD and PW peaked downstream of the cytosine under investigation, revealing that the main contribution to classification power was the IPD and PW downstream of the cytosine under investigation.
[0117] The random forest was composed of multiple decision trees. During the construction of the decision trees, Gini impurity was used to determine which decision logic to use for each decision node. Important features that have a greater impact on the final classification result are likely to be located in nodes closer to the root of the decision tree, while unimportant features that have less impact on the final classification result are likely to be located in nodes further away from the root. Therefore, the importance of a feature can be estimated by calculating the average distance to the root of all decision trees in the random forest.
[0118] In some embodiments, consensus of methylation calls at CpG sites between the Watson and Crick strands can be further used to improve specificity. For example, both strands showing methylation should be called the methylated state, and both strands showing unmethylation should be called the unmethylated state. Since methylation at CpG sites is known to be typically symmetric, confirmation from each strand can improve specificity.
[0119] In various embodiments, global kinetic features from the entire molecule can be used to determine the methylation status. For example, global methylation of the molecule affects the dynamics of the entire molecule during single-molecule real-time sequencing. Modeling the sequencing dynamics of the entire template DNA molecule, including IPD, PW, fragment size, strand information, and sequence context, can improve the accuracy of classification as to whether the molecule is methylated. As an example, the measurement window can be the entire template molecule. Statistics (e.g., mean, median, mode, percentile, etc.) of IPD, PW, or other dynamic features can be used to determine the methylation of the entire molecule.
[0120] B. Limitations of Other Analytical Techniques Detection of methylation based on the IPD of a specific C in a specific sequence motif is very poor; for example, a sensitivity of only 1.9% was reported (Clark et al., 2013). We also attempted to replicate such an analysis by combining different sequence motifs with IPDs, without using the PW metric and using only the IPD cutoff, rather than the data structure described herein. For example, we extracted the 3-nt upstream and downstream neighbors of the CpG under investigation. The IPD of that CpG was stratified into different groups (4096 groups for 6 positions) depending on the context of the 6-nt neighboring sequences (i.e., the 3-nt upstream and downstream, respectively) centered on that CpG. The IPD between methylated and unmethylated CpGs within the same sequence motif was studied using ROC. For example, the unmethylated "AATCGGAC" motif and the methylated "AAT m When comparing the IPD of CpGs in the "CGGAC" motif, the AUC was 0.48. Therefore, using a cutoff for a specific sequence group did not perform as well as the embodiment using a different one.
[0121] Figure 27 shows the performance of the motif-based IPD analysis described above for detecting methylation without using a pulse width signal (Beckmann et al., BMC Bioinformatics, 2014). The vertical bar graphs represent the average AUC (i.e., the number of bases surrounding the CpG site being investigated) across different k-mer motifs flanking the CpG site being investigated. Figure 27 shows that the average AUC for IPD-based discrimination between methylated and unmethylated cytosines across different k-mer motifs (e.g., 2-mer, 3-mer, 4-mer, 6-mer, 8-mer, 10-mer, 15-mer, and 20-mer motifs surrounding the CpG site in question) was found to be less than 60%. These results suggested that considering the IPD of a candidate nucleotide in a given motif context without considering the IPD of neighboring nucleotides (Flusberg et al., 2010) is inferior to the method disclosed herein for determining CpG methylation.
[0122] We also tested the method described in Flusberg et al.'s study (Flusberg et al., 2010). A total of 5,948,348 DNA segments, 2 nt upstream and 6 nt downstream of the cytosine subjected to methylation analysis, were analyzed. There were 2,828,848 methylated segments and 3,119,500 unmethylated segments. As shown in Figure 28, the signals estimated from principal component analysis using IPD and PW were found to largely overlap between fragments containing methylated cytosine (mC) and unmethylated cytosine (C), suggesting that the method described by Flusberg et al. lacks practically meaningful accuracy. These results suggest that principal component analysis, which linearly combines PW and IPD values at bases and neighboring bases, as used in Flusberg et al.'s study (Flusberg et al., 2010), cannot reliably or meaningfully distinguish between 5-methylcytosine and unmethylated cytosine.
[0123] Figure 29 shows that the AUC of the principal component analysis-based method in which two principal components were used in Flusberg et al.'s study (Flusberg et al., 2010), including IPD and PW, was 0.55, which is much less accurate than the convolutional neural network-based approach including IPD and PW, and the sequence context-based approach shown in our disclosure (0.94).
[0124] C. Other mathematical / statistical models In another embodiment, other mathematical / statistical models, including but not limited to random forests and logistic regression, can be trained by adapting the above-developed features. For the CNN model, training and test datasets were constructed from DNA using M.SssI treatment (methylated) and PCR amplification (unmethylated), which were used to train the random forest (Breiman, 2001). In this random forest analysis, each nucleotide was described using a four-component binary vector encoding six features: IPD, PW, and base identity. In such a binary vector, A, C, G, and T are coded as [1,0,0,0], [0,1,0,0], [0,0,1,0], and [0,0,0,1], respectively. For each CpG site analyzed, we incorporated information from its 10 nt upstream and downstream on both strands to form a 252-dimensional (252D) vector, with each feature representing one dimension. The above-described training dataset with 252D vectors was used to train a random forest model and a logistic regression model. The trained model was used to predict the methylation status of an independent test dataset. The random forest consisted of 100 decision trees. Bootstrapping was used during tree construction. When splitting each decision tree node, Gini impurity was used to determine the optimal split, and up to 15 features were considered at each split. Each leaf of the decision tree was required to contain at least 60 samples.
[0125] Figures 30A and 30B show the performance of methods using random forest and logistic regression for methylation prediction. Figure 30A shows the AUC values of the training dataset for CNN, random forest, and logistic regression. Figure 30B shows the AUC values of the test dataset for CNN, random forest, and logistic regression. The method using random forest achieved AUC of 0.93 and 0.86 for the training dataset and the test dataset, respectively.
[0126] A logistic regression model was trained using the training dataset described with the same 252D vector. The trained model was used to predict the methylation status of an independent test dataset. A logistic regression model with L2 regularization (Ng and Y., 2004) was fitted to the training dataset. As shown in Figures 30A and 30B, the method using logistic regression achieved AUCs of 0.87 and 0.83 in the training and test datasets, respectively.
[0127] Therefore, these results suggested that certain models other than CNN (e.g., but not limited to, random forest and logistic regression) can be used for methylation analysis using the features and analysis protocols developed in the present disclosure. These results also suggested that the CNN implemented in accordance with embodiments of the present disclosure had an AUC of 0.90 on the test dataset (Figure 30B), outperforming both random forest (AUC: 0.86) and logistic regression (AUC: 0.83).
[0128] D. Determination of 6mA Modifications of Nucleic Acids In addition to methylated CpG, the methods described herein can also detect other DNA base modifications, such as methylated adenine, including forms of 6mA.
[0129] 1. Detection of 6mA using kinetic features and sequence context To evaluate the performance and usefulness of the disclosed embodiments for determining nucleic acid base modifications, the inventors further analyzed N6-adenine methylation (6mA). In one embodiment, approximately 1 ng of human DNA (e.g., extracted from placental tissue) was amplified to obtain 100 ng of DNA product through whole genome amplification using unmethylated adenine (uA), unmethylated cytosine (C), unmethylated guanine (G), and unmethylated thymine (T).
[0130] Figure 31A shows an example of one approach for generating molecules with unmethylated adenines by whole genome amplification. In this diagram, "uA" indicates unmethylated adenine and "mA" indicates methylated adenine. Whole genome amplification is performed using exonuclease-resistant phosphorothioate-modified random hexamers as primers. The primers bind randomly across the genome, allowing a polymerase (e.g., Phi29 DNA polymerase) to amplify the DNA (e.g., isothermal linear amplification). In stage 3102, double-stranded DNA is denatured. In stage 3106, the amplification reaction begins when several random hexamers (e.g., 3110) anneal to the denatured template DNA (i.e., single-stranded DNA). As shown in 3114, hexamer-mediated DNA synthesis of strand 3118 proceeds in the 5' to 3' direction. Upon reaching the next hexamer-mediated DNA synthesis site, the polymerase displaces the newly synthesized DNA strand (3122) and continues strand elongation. The displaced strand becomes a single-stranded DNA template to which random hexamers can rebind and initiate new DNA synthesis. Repeated hexamer annealing and strand displacement in an isothermal process produces amplified DNA products in high yields. The amplification described here may be referred to as multiple displacement amplification (MDA).
[0131] The amplified DNA product can be further fragmented into fragments having sizes ranging from, but not limited to, 100 bp, 200 bp, 300 bp, 400 bp, 500 bp, 600 bp, 700 bp, 800 bp, 900 bp, 1 kb, 5 kb, 10 kb, 20 kb, 30 kb, 40 kb, 50 kb, 60 kb, 70 kb, 80 kb, 90 kb, 100 kb, or other desired size ranges. The fragmentation process may include enzymatic digestion, nebulization, hydrodynamic shearing, sonication, etc. As a result, original base modifications such as 6mA can be nearly eliminated by whole genome amplification with unmethylated A (uA). Figure 31A shows possible fragments (3126, 3130, and 3134) of the DNA product, with unmethylated A on both strands. Such mA-free whole genome amplified DNA products were subjected to single molecule real-time sequencing to generate uA datasets.
[0132] Figure 31B shows an example of one approach for generating molecules with methylated adenines by whole genome amplification. In this figure, "uA" indicates unmethylated adenine, and "mA" indicates methylated adenine. Approximately 1 ng of human DNA was amplified to obtain 10 ng of DNA product through whole genome amplification using 6 mA and unmethylated C, G, and T. Methylated adenines can be generated through a series of chemical reactions (JD Engel et al. J Biol Chem. 1978;253:927-34). As shown in Figure 31B, whole genome amplification is performed using exonuclease-resistant phosphorothioate-modified random hexamers as primers, which, similar to Figure 31A, bind randomly on the genome and allow a polymerase (e.g., Phi29 DNA polymerase) to amplify the DNA (e.g., by isothermal linear amplification). Exonuclease-resistant phosphorothioate-modified random hexamers are resistant to the 3' to 5' exonuclease activity of proofreading DNA polymerases, thus protecting the random hexamers from degradation during amplification.
[0133] The amplification reaction was initiated when several random hexamers annealed to denatured template DNA (i.e., single-stranded DNA). Hexamer-mediated DNA synthesis proceeded in the 5' to 3' direction until the next hexamer-mediated DNA synthesis site was reached, at which point the polymerase displaced the newly synthesized DNA strand and continued strand elongation. The displaced strand became a single-stranded DNA template to which the random hexamers rebinded and initiated new DNA synthesis. Repeated hexamer annealing and strand displacement in an isothermal process produced amplified DNA products in high yields.
[0134] The amplified DNA products may be further fragmented into lengths, such as, but not limited to, 100 bp, 200 bp, 300 bp, 400 bp, 500 bp, 600 bp, 700 bp, 800 bp, 900 bp, 1 kb, 5 kb, 10 kb, 20 kb, 30 kb, 40 kb, 50 kb, 60 kb, 70 kb, 80 kb, 90 kb, 100 kb, or other combinations. As shown in Figure 31B, the amplified DNA products may contain different methylation patterns across the adenine sites of each strand. For example, both strands of a double-stranded molecule may be methylated at adenine (molecule I), generated when the two strands originate from DNA synthesis during whole genome amplification.
[0135] As another example, one strand of a double-stranded molecule may contain an interlaced methylation pattern across adenine sites (molecule II). An interlaced methylation pattern is defined as containing a mixture of methylated and unmethylated bases present in the DNA strand. In the following example, we use an interlaced adenine methylation pattern, which contains a mixture of methylated and unmethylated adenines present in the DNA strand. This type of double-stranded molecule (molecule II) may be generated because an unmethylated hexamer containing an unmethylated adenine binds to a DNA strand and initiates DNA elongation. Such an amplified DNA product containing a hexamer with an unmethylated adenine would be sequenced. Alternatively, this type of double-stranded molecule (molecule II) may be initiated by fragmented DNA from the original template DNA containing an unmethylated adenine, because such fragmented DNA may bind to the DNA strand as a primer. Such an amplified DNA product containing a portion of the original DNA with an unmethylated adenine in the strand would be sequenced. Since the unmethylated hexamer primer is a very small fraction of the resulting DNA strand, the majority of the fragments still contain 6mA.
[0136] As another example, one strand of a double-stranded DNA molecule may be methylated across the adenine sites, while the other strand may be unmethylated (molecule III). This type of double-stranded molecule can be generated when an original DNA strand without methylated adenines serves as a template DNA molecule to generate a new strand with methylated adenines.
[0137] Both strands may be unmethylated (molecule IV). This type of double-stranded molecule may result from the reannealing of two original DNA strands that do not contain methylated adenines.
[0138] Fragmentation processes can include enzymatic digestion, nebulization, hydrodynamic shearing, and sonication. Such whole-genome amplified DNA products can be methylated primarily at A sites. This DNA with mA was subjected to single-molecule real-time sequencing to generate an mA dataset.
[0139] For the uA dataset, 262,608 molecules with a median length of 964 bp were sequenced using single-molecule real-time sequencing. The median subread depth was 103x. 48% of the subreads could be aligned to the human reference genome using the BWA aligner (Li H et al. Bioinformatics. 2009;25:1754-60). As an example, single-molecule real-time sequencing can be performed using the Sequel II system (Pacific Biosciences). Fragmented DNA molecules were subjected to the construction of single-molecule real-time (SMRT) sequencing templates using the SMRTbell Express Template Prep Kit 2.0 (Pacific Biosciences). Sequencing primer annealing and polymerase binding conditions were calculated using SMRT Link v8.0 software (Pacific Biosciences). Briefly, sequencing primer v2 was annealed to the sequencing template, and then polymerase was allowed to bind to the template using the Sequel II Binding and Internal Control Kit 2.0 (Pacific Biosciences). Sequencing was performed on a Sequel II SMRT Cell 8M. Sequencing videos were collected for 30 hours on the Sequel II system using the Sequel II Sequencing Kit 2.0 (Pacific Biosciences).
[0140] For the mA dataset, 804,469 molecules with a median length of 826 bp were sequenced using single-molecule real-time sequencing. The median subread depth was 34x. 27% of the subreads could be aligned to the human reference genome using the BWA aligner (Li H et al. Bioinformatics. 2009;25:1754-60).
[0141] In one embodiment, kinetic characteristics, including but not limited to IPD and PW, were analyzed in a strand-specific manner. For sequence results derived from Watson strands, 644,318 A sites without methylation randomly selected from the uA dataset and 718,586 A sites with methylation randomly selected from the mA dataset were used to construct a training dataset. This training dataset was used to establish a classification model and / or threshold for distinguishing between methylated and unmethylated adenines. The test dataset consisted of 639,702 A sites without methylation and 723,320 A sites with methylation. This test dataset was used to verify the performance of the model / threshold estimated from the training dataset.
[0142] The sequence results derived from the Watson strand were analyzed. Figure 32A shows the interpulse distance (IPD) values across the training datasets for the uA and mA datasets. For the training dataset, the IPD values across the sequenced A sites were observed to be higher in the mA dataset (median: 1.09, range: 0-9.52) than in the uA dataset (median: 0.20, range: 0-9.52) (P value < 0.0001, Mann-Whitney U test).
[0143] Figure 32B shows the IPD of the test datasets of the uA and mA datasets. When examining the IPD values across the sequenced A sites in the test dataset, it was observed that the IPD values of the mA dataset were higher than those of the uA dataset (median 1.10 vs. 0.19, P value < 0.0001, Mann-Whitney U test).
[0144] Figure 32C shows the area under the receiver operating characteristic (ROC) curve using the IPD cutoff. The true positive rate is on the y-axis, and the false positive rate is on the x-axis. The area under the receiver operating characteristic curve (AUC) for distinguishing between methylated and unmethylated sequence A bases in template DNA molecules using the corresponding IPD values was 0.86 for both the training and test datasets.
[0145] In addition to the results from the Watson strand, we also analyzed the sequence results from the Crick strand. Figure 33A shows the IPD values across the training datasets for the uA and mA datasets. For the training dataset, we observed that the IPD values across the sequenced A sites were higher in the mA dataset (median: 1.10, range: 0-9.52) than in the uA dataset (median: 0.19, range: 0-9.52) (P value < 0.0001, Mann-Whitney U test).
[0146] Figure 34B shows the IPD values of the test datasets for the uA and mA datasets. Compared to the uA dataset, higher IPD values were observed across the sequenced A sites in the mA dataset from the test dataset (median 1.10 vs. 0.19, P value < 0.0001, Mann-Whitney U test).
[0147] Figure 33C shows the area under the ROC curve. The true positive rate is on the y-axis, and the false positive rate is on the x-axis. The area under the ROC curve (AUC) values for distinguishing between sequenced A bases in methylated and unmethylated template DNA molecules using the corresponding IPD values were 0.86 and 0.87 for the training and test datasets, respectively.
[0148] Figure 34 shows an illustration of 6mA determination of Watson strands using a measurement window according to an embodiment of the present invention. Such a measurement window can include kinetic features such as IPD and PW and nearby sequence context. 6mA determination can be performed similarly to determining methylated CpG.
[0149] 35 shows an illustration of 6 mA determination of a Crick strand using a measurement window, according to an embodiment of the present invention. Such a measurement window can include kinetic features such as IPD and PW, as well as nearby sequence context.
[0150] As an example, a measurement window was constructed using 10 bases from each side of the sequenced A base of the template DNA being investigated. Using feature values including IPD, PW, and sequence context, a convolutional neural network (CNN) was used to train a model according to the method disclosed herein. In other embodiments, statistical models may include, but are not limited to, linear regression, logistic regression, deep recurrent neural networks (e.g., long short-term memory, LSTM), Bayesian classifiers, hidden Markov models (HMMs), linear discriminant analysis (LDA), k-means clustering, density-based spatial clustering for noisy applications (DBSCAN), random forest algorithms, support vector machines (SVMs), etc.
[0151] Figures 36A and 36B show the determined methylation probabilities for sequenced A bases in the Watson strand between the uA and mA datasets using a measurement window-based CNN model. Figure 36A shows that the CNN model was trained from the training dataset. As an example, the CNN model utilized two 1D convolutional layers (each with 64 filters with a kernel size of 4 followed by a ReLU (rectified linear unit) layer) followed by a dropout layer with a dropout rate of 0.5. A max-pooling layer with a pool size of 2 was used. This then flowed into two 1D convolutional layers (each with 128 filters with a kernel size of 2 followed by a ReLU layer) followed by a dropout layer with a dropout rate of 0.5. A max-pooling layer with a pool size of 2 was used. Finally, a fully connected layer containing 10 neurons, followed by a ReLU layer with an output layer containing one neuron, followed by a sigmoid layer, yielded the methylation probability. Other settings for layers, filters, and kernel sizes can be adapted as described herein, for example, for other methylations (e.g., CpGs). For this training dataset of Watson strand sequencing results, 644,318 and 718,586 A bases from the unmethylated and methylated libraries were used.
[0152] Based on the CNN model, for the Watson strand-associated data, sequenced A bases in template DNA molecules from the mA database had a much higher probability of methylation compared to those A bases present in the uA database in both the training and test datasets (P value < 0.0001, Mann-Whitney U test). For the training dataset, the median probability of methylation at A sites in the uA dataset was 0.13 (interquartile range, IQR: 0.09-0.15), while the value for the mA dataset was 1.000 (IQR: 0.998-1.000).
[0153] Figure 36A shows the methylation probabilities determined for the test dataset. For the test dataset, the median probability of methylation at the A site for the uA dataset was 0.13 (IQR: 0.10-0.15), while the value for the mA dataset was 1.000 (IQR: 0.997-1.000). Figures 36A and 36B show that a measurement window-based CNN model can be trained to detect methylation in the test dataset.
[0154] Figure 37 shows the ROC curve for detecting 6mA using a measurement window-based CNN model for sequenced A bases in the Watson strand. The true positive rate is on the y-axis, and the false positive rate is on the x-axis. This figure shows that the AUC values for distinguishing between methylated and unmethylated sequenced A sites using the CNN model are 0.94 and 0.93, respectively, for the training and test datasets composed of Watson strand sequencing results. This suggests that it is feasible to use the present disclosure to determine the methylation status of A sites using Watson strand data. Using a determined methylation probability of 0.5 as the cutoff, a specificity of 99.3% and a sensitivity of 82.6% can be achieved for detecting 6mA. Figure 37 shows that the measurement window-based CNN model can be used to detect 6mA with high specificity and sensitivity. The accuracy of the model can be compared to techniques using only the IPD metric.
[0155] Figure 38 shows a performance comparison of IPD metric-based 6mA detection and measurement window-based 6mA detection. Sensitivity is plotted on the y-axis, and specificity is plotted on the x-axis. Figure 38 shows that performance using the measurement window-based 6mA classification disclosed herein (AUC: 0.94) was superior to the conventional method using only the IPD metric (AUC: 0.87) (P-value < 0.0001, DeLong's test). The measurement window-based CNN model outperformed the IPD metric-based detection.
[0156] Figures 39A and 39B show the determined probabilities of methylation for the sequenced A bases of the Crick strand between the uA and mA datasets using a measurement window-based CNN model. Figure 39A shows the training dataset, and Figure 39B shows the test dataset. Both figures plot the probability of methylation on the y-axis. Figures 39A and 39B show that for the Crick strand-related data, based on the CNN model, the sequenced A bases of the template DNA molecule from the mA database have a much higher probability of methylation in both the training and test datasets compared to those A bases present in the uA database (P value < 0.0001, Mann-Whitney U test).
[0157] Figure 40 shows the performance of 6mA detection using a measurement window-based CNN model for sequenced A bases in Crick strands. The true positive rate is on the y-axis. The false positive rate is on the x-axis. Figure 40 shows that the AUC values for distinguishing between methylated and unmethylated sequenced A sites using the CNN model are 0.95 and 0.94, respectively, for the training and test datasets consisting of Crick strand sequencing results. Performance using the CNN approach disclosed herein (AUC: 0.94) was also shown to be superior to performance using the IPD metric alone (0.87) (P-value < 0.0001). This result suggested that it is feasible to determine the methylation status of A sites using Crick strand data using the methods disclosed herein. Using a determined methylation probability of 0.5 as a cutoff, a specificity of 99.3% and a sensitivity of 83.0% can be achieved for the detection of 6mA. Figure 40 demonstrates that 6mA can be detected with high specificity and sensitivity using a measurement window-based CNN model.
[0158] Figure 41 shows examples of methylation states across A bases in molecules containing Watson and Crick strands. White dots represent unmethylated adenines. Black dots represent methylated adenines. The horizontal dotted lines represent the strands of a double-stranded DNA molecule. Molecule 1 shows that both the Watson and Crick strands are determined to be methylated across A bases. Molecule 2 shows that the Watson strand is nearly entirely unmethylated, while the Crick strand is nearly entirely methylated. Molecule 3 shows that both the Watson and Crick strands are determined to be nearly entirely methylated across A bases.
[0159] 2. Reinforced training using selective datasets As shown in Figures 36A, 36B, 39A, and 39B, there was a bimodal distribution of methylation probabilities across the sequenced A bases of the template DNA molecules in the mA dataset. In other words, there were some molecules with uA signals in the mA dataset. This was further evidenced by the presence of fully unmethylated and hemimethylated molecules in the mA dataset (Figure 41). One possible reason is that molecules containing 6mA reduce the amplification efficiency of DNA during the whole genome amplification step, so molecules containing uA in the DNA template still account for a significant portion of the mA dataset after whole genome amplification. This explanation was supported by the fact that 1 ng of genomic DNA amplified with 6mA only produced 10 ng of DNA product, whereas 1 ng of genomic DNA amplified with unmethylated A produced 100 ng of DNA product under the same amplification conditions. Thus, for the mA dataset, the original template DNA molecule, in which adenines are typically unmethylated (e.g., 0.051%) (Xiao CL et al. Mol 2018;71:306-318), would account for approximately 10% of the total adenines.
[0160] In one embodiment, when training a CNN model to distinguish between mA and uA, selectively using those A bases with relatively high IPD values in the mA dataset reduces the impact of uA data on training the model for mA detection. Only A bases with IPD values above a certain cutoff value can be used. The cutoff value may correspond to a percentile. In one embodiment, those A bases in the mA dataset with IPD values greater than the value at the 10th percentile will be used. In some embodiments, those A bases with IPD values greater than the value at the 1st, 5th, 15th, 20th, 30th, 40th, 50th, 60th, 70th, 80th, 90th, or 95th percentile will be used. The percentile may be based on data from all nucleic acid molecules in a reference sample or multiple reference samples.
[0161] Figure 42 shows the performance of enhanced training by selectively using A bases in the mA dataset with IPD values greater than the 10th percentile. Figure 42 shows the true positive rate on the y-axis and the false positive rate on the x-axis. The figure shows that using A bases in the mA dataset with IPD values greater than the 10th percentile to train the CNN model increased the AUC for distinguishing between mA and uA bases to 0.98, which was superior to the model trained on data without selection by IPD value before training (AUC: 0.94). This suggests that using IPD values to select mA sites in creating the training dataset helps improve discrimination.
[0162] To further confirm the presence of molecules with uA bases in the mA dataset, we hypothesized that the 6mA present in the molecule slows down polymerase extension during new strand generation compared to molecules without 6mA, and therefore the percentage of uA in the mA dataset increases in wells with more subreads.
[0163] Figure 43 shows a graph of the number of subreads in each well versus the percentage of unmethylated adenines in the mA dataset. The y-axis shows the percentage of uA in the mA dataset. The x-axis shows the number of subreads in each well. After removing A sites with IPD values below the 10th percentile, the test dataset was reanalyzed using the enriched model trained by using mA sites. As the number of subreads per well increased (including from 1 to 10 subreads per sequencing well, from 10 to 20 subreads per well, from 40 to 50 subreads per well, from 60 to 70 subreads per well, and above 70 subreads per well), a gradual increase in uA was observed (i.e., from 14.6% to 55.05%). Therefore, wells with a high number of subreads tend to have lower mA. Methylation of A can slow the progression of the sequencing reaction. Thus, sequencing wells with a large depth of subreads are more likely to be unmethylated with respect to A. This behavior can be exploited to detect unmethylated molecules using a cutoff value for the number of subreads associated with a molecule; for example, more than 70 subreads can be identified as mostly unmethylated.
[0164] Figure 44 shows the methyladenine patterns between the Watson and Crick strands of double-stranded DNA molecules in the test dataset. Because the methylation of A is asymmetric, the two strands behave differently. Most molecules are methylated by mA incorporation, with some remaining unmethylated A. The y-axis shows the methyladenine levels of the Crick strand. The x-axis shows the methyladenine levels of the Watson strand. Each point represents a double-stranded molecule. Using a reinforcement model trained with selected mA sites, the double-stranded molecules can be classified into different groups according to the methylation level of each strand as follows: (a) For double-stranded DNA molecules, the methyladenine levels of both the Watson and Crick strands were greater than 0.8. Such double-stranded molecules were defined as fully methylated with respect to adenine sites (Figure 44, region A). The methyladenine level of a strand was defined as the percentage of A sites determined to be methylated among all A sites in that strand. (b) For double-stranded DNA molecules, the methyladenine level in one strand was greater than 0.8, while the other strand was less than 0.2. Such molecules were defined as hemimethylated at the adenine site (Figure 44, regions B1 and B2). (c) For double-stranded DNA molecules, the methyladenine levels in both the Watson and Crick strands were less than 0.2. Such double-stranded molecules were defined as fully unmethylated at the adenine site (Figure 44, region C). (d) For double-stranded DNA molecules, the methyladenine levels in the Watson and Crick strands did not belong to groups a, b, or c. Such double-stranded molecules were defined as molecules with an interlaced methylation pattern with respect to the adenine sites (Figure 44, region D). An interlaced methylation pattern was defined as a mixture of methylated and unmethylated adenines present in the DNA strands.
[0165] In some other embodiments, the cutoff level of methyladenine for defining an unmethylated strand can be, but is not limited to, less than 0.01, 0.05, 0.1, 0.2, 0.3, 0.4, and 0.5, and the cutoff level of methyladenine for defining a methylated strand can be, but is not limited to, greater than 0.5, 0.6, 0.7, 0.8, 0.9, 0.95, and 0.99.
[0166] Figure 45 is a table showing the percentages of molecules with fully unmethylated, hemimethylated, fully methylated, and interlaced methyladenine patterns in the training and test datasets. The molecules in the test dataset can be classified into fully unmethylated (7.0%), hemimethylated (9.8%), fully methylated (79.4%), and molecules with interlaced methyladenine patterns (3.7%) at adenine sites. These results are comparable to those shown in the training dataset, which contained fully unmethylated (7.0%), hemimethylated (10.0%), fully methylated (79.4%), and molecules with interlaced methyladenine patterns (3.6%) at adenine sites.
[0167] Figure 46 shows representative examples of molecules that are fully unmethylated, hemimethylated, fully methylated, and have an interlaced methyladenine pattern at adenine sites. White dots represent unmethylated adenines. Black dots represent methylated adenines. The dotted horizontal lines represent the strands of a double-stranded DNA molecule.
[0168] In embodiments, increasing the purity of the 6mA bases used to train the CNN model can improve its performance in distinguishing between methylated and unmethylated adenines. To this end, increasing the duration of the DNA amplification reaction to increase the amount of newly generated DNA product can reduce the effect of unmethylated adenines contributed from the original DNA template. In other embodiments, biotinylated bases can be incorporated during DNA amplification using 6mA. The newly generated DNA products containing 6mA can be pulled down and concentrated using streptavidin-coated magnetic beads.
[0169] Use of the 3.6mA methylation profile The 6mA modification of DNA is present in the genomes of bacteria, archaea, protists, and fungi (Didier W et al. Nat Rev Micorbiol. 2009;4:183-192). It has also been reported that 6mA is present in the human genome, accounting for 0.051% of all adenines (Xiao CL et al. Mol Cell. 2018;71:306-318). Considering the low content of 6mA in the human genome, in one embodiment, a training dataset can be created by adjusting the ratio of 6mA in the dNTP mix (N represents unmodified A, C, G, and T) during the whole genome amplification step. For example, the ratio of 6mA to dNTP can be 1:10, 1:100, 1:1000, 1:10000, 1:100000, or 1:1000000. In another embodiment, the adenine DNA methyltransferase M.EcoGII can be used to generate a 6mA training dataset.
[0170] The amount of 6mA is lower in gastric cancer and liver cancer tissues, and this downregulation of 6mA correlates with increased tumor formation (Xiao CL et al. Mol Cell. 2018;71:306-318). On the other hand, high levels of 6mA have been reported in glioblastoma (Xie et al. Cell. 2018;175:1228-1243). Therefore, the 6mA approach disclosed herein may be useful for studying cancer genomics (Xiao CL et al. Mol Cell. 2018;71:306-318; Xie et al. Cell. 2018;175:1228-1243). Furthermore, 6mA has been found to be more prevalent and abundant in mammalian mitochondrial DNA and has been shown to be associated with hypoxia (Hao Z et al. Mol Cell. 2020; doi:10.1016 / j.molcel.2020.02.018). Therefore, the approach for detecting 6mA in this disclosure will be useful for studying mitochondrial stress responses under different clinical conditions, such as pregnancy, cancer, and autoimmune diseases.
[0171] IV. Results and Applications A. Methylation Detection The detection of methylation at CpG sites using the above method has been performed on various biological samples and genomic regions. As an example, methylation determination using bisulfite sequencing was validated against cell-free DNA in the plasma of pregnant women using single-molecule real-time sequencing. Methylation results can be used for various purposes, including copy number determination and disorder diagnosis. The method described below is not limited to CpG sites and can be applied to any of the modifications described herein.
[0172] 1. Detection of methylation of long DNA molecules in placental tissues Single-molecule real-time sequencing can sequence DNA molecules up to kilobases in length (Nattestad et al., 2018). By synergistically utilizing the long-read information from single-molecule real-time sequencing, the methylation status of CpG sites can be deciphered using the present invention described herein, enabling the inference of haplotype information from the methylation status. To demonstrate the feasibility of inferring the methylation status and haplotype information from long reads, placental tissue DNA was sequenced using 478,739 molecules covered by 28,913,838 subreads. Seven molecules were over 5 kb in size. Each was covered by an average of three subreads.
[0173] Figure 47 shows the methylation status along a 6,265-bp long DNA molecule (i.e., haplotype block). It was sequenced in ZMW with ZMW hole number m54276_180626_162240 / 40763503 and mapped to the genomic location chr1:113246546-113252811 in the human genome. "-" represents a non-CpG nucleotide. "U" represents the unmethylated state of the CpG site. "M" represents the methylated state of the CpG site. The yellow-highlighted region 4710 indicates a CpG island region, which is generally known to be unmethylated (Figure 47). The majority of the CpG sites in the CpG island were predicted to be unmethylated (96%). In contrast, 75% of the CpG sites outside the CpG island were predicted to be unmethylated. These results suggest that methylation levels outside CpG islands (e.g., at the shore / shelf of CpG islands) are higher than those within CpG islands. The mixture of methylated and unmethylated states in haplotype configurations outside CpG islands indicates variable methylation patterns. This observation is generally consistent with current understanding (Zhang et al., 2015; Feinberg and Irizarry, 2010). Therefore, this disclosure enables the identification of different methylation states along long stretches of DNA, including methylated and unmethylated states, implying that haplotype information on methylation states may be graded. Haplotype information refers to the association of the methylation state of CpG sites with a contiguous stretch of DNA.
[0174] In one embodiment, this approach can be used herein to analyze methylation status along haplotypes to detect and analyze imprinted regions. Imprinted regions are subject to epigenetic regulation that results in a parent-of-origin methylation state. For example, one important imprinted region is located on human chromosome 11p15.5, which contains the imprinted genes IGF2, H19, and CDKN1C (P57), which are potent regulators of fetal growth. kip2) (Brioude et al., Nat Rev Endocrinol. 2018;14:229-249). Genetic and epigenetic abnormalities in imprinted regions may be associated with disease. Beckwith-Wiedemann syndrome (BWS) is an overgrowth syndrome in which patients often present with macroglossia, abdominal wall defects, hemihypertrophy, enlarged abdominal organs, and an increased risk of embryonal tumors during childhood. BWS is thought to result from genetic or epigenetic defects within the 11p15.5 region (Brioude et al., Nat Rev Endocrinol. 2018;14:229-249). A region called ICR1 (imprinted control region 1), located between H19 and IGF2, is variably methylated on the paternal allele. ICR1 induces parent-of-origin-specific expression of IGF2. Therefore, genetic and epigenetic abnormalities in ICR1 lead to abnormal IGF2 expression, which may be one of the reasons for BWS. Therefore, detection of the methylation status along imprinted regions is of clinical importance.
[0175] The inventors downloaded data for 92 imprinted genes from a public database (http: / / www.geneimprint.org / ) that curates currently reported imprinted genes. 5 kb upstream and downstream regions of these imprinted genes were used for further analysis. Within these regions, 160 CpG islands are associated with these imprinted genes. The inventors obtained 324,248 circular consensus sequences from placenta samples. After removing low-quality circular consensus sequences and short regions overlapping with CpG islands (e.g., less than 50% of the length of the associated CpG island), nine circular consensus sequences overlapping with nine CpG islands corresponding to eight imprinted genes were obtained.
[0176] Figure 48 is a table showing that nine DNA molecules were sequenced by single-molecule real-time sequencing and overlapped with imprinted regions, including H19, WT1-AS, WT1, DLK1, MEG3, ATP10A, LRRTM1, and MAGI2. The sixth column contained DNA stretches overlapping with CpG islands containing imprinted regions. "U" represents unmethylated cytosine in a CpG context. "M" represents methylated cytosine in a CpG context. "*" represents a CpG site not covered by the sequencing results. "-" represents a nucleotide from a non-CpG site. If a molecule overlaps with a single nucleotide polymorphism (SNP), the genotype is indicated in parentheses. The seventh column indicates the methylation status of the entire molecule. According to embodiments present in the present disclosure, a molecule can be called methylated if the majority (e.g., more than 50%) of its CpG sites are shown to be methylated. Otherwise, it is called unmethylated.
[0177] Of the nine DNA molecules, five (55.6%) were called methylated, which did not deviate significantly from the expectation that 50% of the DNA molecules would be methylated. As shown in the sixth column of the table in Figure 48, the majority of CpG sites were shown to be methylated or unmethylated in concert (i.e., as a methylation haplotype). In one embodiment, according to embodiments present in the present disclosure, if the majority (e.g., more than 50%) of the CpG sites are shown to be methylated, the molecule may be called methylated; otherwise, it is called unmethylated. Other cutoffs for determining whether a molecule is methylated can be used, including, but not limited to, at least 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and 100% of the CpG sites in the analyzed molecule are considered to be methylated.
[0178] In another embodiment, molecules that simultaneously include analysis of at least one SNP and at least one CpG site can be used to determine whether a region is associated with an imprinted region or whether a known imprinted gene is abnormal (e.g., loss of imprint). For illustrative purposes, Figure 49 shows a first molecule from an imprinted region with an "A" allele and a second molecule from the imprinted region with an "G" allele. Assuming the imprinted region is paternally imprinted, the first molecule from the maternal haplotype was fully unmethylated, and the second molecule from the paternal haplotype was fully methylated. In one embodiment, such an assumption provides a ground truth of the methylation status, allowing for testing the performance of base modification detection according to embodiments present in the present disclosure.
[0179] Figure 49 shows an example of determining methylation patterns in imprinted regions. DNA from a biological sample was extracted and ligated with hairpin adapters to form circular DNA molecules. The sequence information and base modifications (e.g., methylation status of CpG sites) for these circular DNA molecules were unknown. These circular DNA molecules were subjected to single-molecule real-time sequencing. After subreads were mapped to a reference genome, the IPD, PW, and sequence context for each base in each subread derived from the circular DNA molecules were determined. Furthermore, the genotypes of these molecules were determined. The IPD, PW, and sequence context of the measurement window associated with the CG site were compared with a reference kinetic pattern according to embodiments present in the present disclosure to determine the methylation status of each CpG. If two molecules with different alleles exhibit different methylation patterns, such as one being completely unmethylated and the other being fully methylated, the genomic region associated with these two molecules is likely an imprinted region. In one embodiment, for example, as shown in Figure 49, if such a genomic region happened to be a known imprinted region, the methylation patterns of these two molecules matched the methylation pattern expected under normal circumstances (i.e., ground truth), which may indicate the accuracy of a method for classifying methylation status according to embodiments present in the present disclosure. In one embodiment, a derivation between the measured methylation pattern and the expected methylation pattern according to embodiments present in the present disclosure would indicate an imprinting abnormality, e.g., loss of imprinting.
[0180] Figure 50 shows an example of determining methylation patterns in an imprinted region. In one embodiment, the imprinting pattern can be further determined through analyzing the methylation pattern of that region across a particular family tree. For example, analysis of the methylation patterns and allelic information across the paternal and maternal genomes and offspring can be performed. Such a family tree may further include the genomes of the paternal or maternal grandfather, paternal or maternal grandmother, or other related genomes. In another embodiment, such analysis can be extended to a family trio (mother, father, and child) dataset in a particular population, obtaining methylation and genotype information for each individual, for example, according to the embodiments present herein.
[0181] As shown after classification, both the genotype (allele in the box) and methylation status can be determined. For each molecule, the methylation pattern at each site can be provided to identify which parent the molecule was inherited from (e.g., all methylated or all unmethylated). Alternatively, the methylation density can be determined, and one or more cutoffs can be used to classify the molecule as hypermethylated (e.g., >80% or other percentage from one parent) or hypomethylated (e.g., <20% or other percentage from the other parent).
[0182] 2. Detection of methylation of cfDNA molecules As another example, cell-free DNA (cfDNA) methylation is also increasingly recognized as an important molecular signal for noninvasive prenatal testing. For example, we demonstrated that tissue-specific methylation of cfDNA molecules can be used to determine the proportional contributions from different tissues, such as neutrophils, T cells, B cells, liver, and placenta, in maternal plasma (Sun et al., 2015). The feasibility of using plasma DNA methylation to detect trisomy 21 has also been demonstrated (Lun et al., 2013). cfDNA molecules in maternal plasma were fragmented to a median size of 166 bp, which is much shorter than artificially fragmented E. coli DNA, which is approximately 500 bp in size. It has been reported that cfDNA is not randomly fragmented. For example, terminal motifs in plasma DNA are associated with tissue origin, such as placental origin. These distinctive properties of cfDNA provide a very different sequence context from artificially fragmented E. coli DNA. Therefore, it remains unclear whether the kinetics of such polymerases typically allow for quantitative estimation of the methylation level of cell-free DNA molecules. The disclosure in this patent application is applicable, for example, but not limited to, to methylation analysis of cell-free DNA in the plasma of pregnant women by using a methylation prediction model trained from the above tissue DNA molecules.
[0183] Using single-molecule real-time sequencing, six plasma DNA samples from pregnant women with male fetuses were sequenced, yielding a median of 30,738,399 subreads (range: 1,431,215-105,835,846) corresponding to a median of 111,834 CCS (range: 61,010-503,582). Each plasma DNA was sequenced a median of 262 times (range: 173-320). The dataset was generated from DNA prepared with the Sequel I Sequencing Kit 3.0.
[0184] To evaluate the detection of methylation in cfDNA molecules, we analyzed the methylation of the six plasma DNA samples from pregnant women using bisulfite sequencing (Jiang et al., 2014). A median of 66 million paired-end reads was obtained (ranging from 58 to 82 million paired-end reads). The median overall methylation was found to be 69.6% (range, 67.1% to 72.0%).
[0185] Figure 51 shows a comparison of methylation levels estimated by the new approach and traditional bisulfite sequencing. The y-axis is the methylation level predicted according to the embodiment present in this patent application. The x-axis is the methylation level estimated by bisulfite sequencing. A median of 314,675 CpG sites (range: 144,546-1,382,568) were analyzed for plasma DNA results generated from single-molecule real-time sequencing. The median percentage of CpG sites predicted to be methylated was 64.7% (range: 60.8-68.5%), which appeared comparable to the results estimated from bisulfite sequencing. As shown in Figure 51, there was a good correlation (r: 0.96, p-value = 0.0023) between the global methylation levels estimated by single-molecule real-time sequencing with this methylation prediction approach and bisulfite sequencing.
[0186] Due to the shallow depth of bisulfite sequencing, it may not be robust enough to estimate the methylation level of each CpG in the human genome (i.e., the proportion of sequenced CpGs that are methylated). Instead, we calculated the methylation level of some regions with multiple CpG sites by aggregating read signals covering CpG sites in genomic regions where any two consecutive CpG sites are within 50 nt and the number of CpG sites is at least 10. The percentage of sequenced cytosines out of the total number of sequenced cytosines and thymines across all CpG sites in a region indicated the methylation level of that region. Regions were divided into different groups according to their methylation level. The probability of methylation predicted by the model trained from a previous training dataset (i.e., tissue DNA) increased as the methylation level increased (Figure 52A). These results further suggested the feasibility and validity of using single-molecule real-time sequencing to predict the methylation status of cfDNA molecules in pregnant women. Figure 52B showed that methylation levels in a 10 Mb genomic window estimated using single molecule real-time sequencing according to embodiments present in the present disclosure were well corrected with those from bisulfite sequencing (r=0.74, p-value<0.0001).
[0187] Figure 53 shows that the genomic representation (GR) of the Y chromosome in maternal plasma of pregnant women measured by single-molecule real-time sequencing correlated well with that measured by BS-seq (r = 0.97, P value = 0.007). These results suggest that single-molecule real-time sequencing also enables accurate quantification of DNA molecules derived from non-hematopoietic tissues, such as the placenta, which generally contribute a small amount of DNA. In other words, this disclosure demonstrates the feasibility of simultaneously analyzing copy number abnormalities and methylation status of natural molecules without base conversion and amplification prior to sequencing.
[0188] 3. CpG block-based methods Some embodiments may perform methylation analysis on several genomic regions having multiple CpG sites, including, for example, but not limited to, 2, 3, 4, 5, 10, 20, 30, 40, 50, 100 CpG sites. The size of such genomic regions may be, for example, but not limited to, 50, 100, 200, 300, and 500 nt. The distance between CpG sites in this region may be, for example, but not limited to, 10, 20, 30, 40, 50, 100, 200, 300 nt. In one embodiment, any two consecutive CpG sites within 50 nt may overlap to form a CpG block, such that the number of CpG sites in this block is 11 or more. Such block-based methods combine multiple regions into a single window represented as a single matrix, allowing the regions to be effectively processed together.
[0189] As an example, as shown in Figure 54, the dynamics of all subreads associated with a CpG block were used for methylation analysis. The predicted IPD profiles of the upstream and downstream 10-nt neighbors of each CpG within that block were artificially aligned to the CpG site to calculate an average IPD profile (Figure 54). The term "projected" refers to the alignment of the subread dynamics signals to each corresponding CpG site in question. The average IPD profile of a CpG block was used to train a model (e.g., an artificial neural network, or ANN) to identify the methylation status of each block. The ANN analysis included an input layer, two hidden layers, and an output layer. Each CpG block was characterized by a feature vector of 21 IPD values, which were input to the ANN. The first hidden layer included 10 neurons with ReLu as the activation function. The second hidden layer included 5 neurons with ReLu as the activation function. Finally, the output layer included one neuron with a sigmoid activation function, which output the methylation probability. CpG sites with a methylation probability greater than 0.5 were considered methylated; otherwise, they were considered unmethylated. The average IPD profile can be used to analyze the methylation status of the entire molecule. If a certain number of sites (e.g., 0, 1, 2, 3, etc.) exceed a threshold value and are methylated, or if the molecule has a certain methylation density, the entire molecule can be considered methylated.
[0190] The unmethylated and methylated libraries contained 9,678 and 9,020 CpG blocks, respectively, each containing at least 10 CpG sites. These CpG blocks covered 176,048 and 162,943 CpG sites in the unmethylated and methylated libraries, respectively. As shown in Figures 55A and 55B, we were able to achieve an overall accuracy of over 90% in predicting methylation status in both the training and test datasets. However, such an embodiment relying on CpG blocks would significantly reduce the number of CpGs that can be evaluated. By definition, the requirement of a minimum number of CpG sites limits methylation analysis to specific genomic regions (e.g., analysis of CpG islands).
[0191] B. Determination of Origin or Disorder Methylation profiles can be used to detect tissue origin or determine disorder classification. Methylation profile analysis can be used in combination with other clinical data, including imaging, conventional blood panels, and other medical diagnostic information. Methylation profiles can be determined using any of the methods described herein.
[0192] 1. Determining copy number abnormalities In this section, we show that SMRT is accurate for determining copy number and therefore methylation and copy number profiles can be analyzed simultaneously.
[0193] Copy number abnormalities have been shown to be revealed by sequencing tumor tissue (Chan (2013)). Here, we demonstrate that cancer-associated copy number abnormalities can be identified by sequencing tumor tissue using single-molecule real-time sequencing. For example, for case TBR3033, we obtained 589,435 and 1,495,225 consensus sequences (minimum requirement of 5 subreads used to construct each consensus sequence) for tumor DNA and its paired adjacent non-tumor liver tissue DNA, respectively. The dataset was generated from DNA prepared using the Sequel II Sequencing Kit 1.0. In one embodiment, the genome was divided in silico into 2-Mb windows. The percentage of consensus sequences mapping to each window was calculated to obtain a genome representation (GR) at 2-Mb resolution. GR was determined by the number of reads at a position and normalized by the total sequence reads across the genome.
[0194] Figure 56A shows the GR ratio between tumor DNA and its paired adjacent non-tumor tissue DNA using single-molecule real-time sequencing. The copy number ratio of tumor DNA to its paired adjacent normal tissue DNA is shown on the y-axis, and the genomic bin index for each 2 Mb window encompassing chromosomes 1-22 is shown on the x-axis. In this figure, regions with GR ratios below the 5th percentile of all 2 Mb windows were classified as having copy number losses, whereas regions with GR ratios above the 95th percentile of all 2 Mb windows were classified as having copy number gains. Chromosome 13 showed a copy number loss, while chromosome 20 showed a copy number gain. These gains and losses are correct results.
[0195] Figure 56B shows the ratio of GR between a tumor and its paired adjacent non-tumor tissue using bisulfite sequencing. The copy number ratio of tumor DNA to its paired adjacent normal tissue DNA is shown on the y-axis, and the genomic bin index for each 2 Mb window containing chromosomes 1-22 is shown on the x-axis. The copy number changes identified by single-molecule real-time sequencing in Figure 56A were verified with the concordant bisulfite sequencing results in Figure 56B.
[0196] For case TBR3032, 413,982 and 2,396,054 consensus sequences were obtained for tumor DNA and its paired adjacent non-tumor tissue DNA, respectively (minimum requirement of 5 subreads used to construct each consensus sequence). In one embodiment, the genome was divided in silico into 2 Mb windows. The percentage of consensus sequences mapping to each window, i.e., 2 Mb genome representation (GR), was calculated.
[0197] Figure 57A shows the GR ratio between tumor DNA and its paired adjacent non-tumor tissue DNA using single-molecule real-time sequencing. The copy number ratio of tumor DNA to its paired adjacent normal tissue DNA is shown on the y-axis, and the genomic bin index for each 2 Mb window encompassing chromosomes 1-22 is shown on the x-axis. In this figure, regions with GR ratios below the 5th percentile of all 2 Mb windows were classified as having copy number losses, whereas regions with GR ratios above the 95th percentile of all 2 Mb windows were classified as having copy number gains. Chromosomes 4, 6, 11, 13, 16, and 17 exhibited copy number losses, while chromosomes 5 and 7 exhibited copy number gains.
[0198] Figure 57B shows the ratio of GR between a tumor and its paired adjacent non-tumor tissue using bisulfite sequencing. The copy number ratio of tumor DNA to its paired adjacent normal tissue DNA is shown on the y-axis, and the genomic bin index for each 2 Mb window containing chromosomes 1-22 is shown on the x-axis. The copy number changes identified by single-molecule real-time sequencing in Figure 57A were verified with the concordant bisulfite sequencing results in Figure 57B.
[0199] Therefore, methylation and copy number profiles can be analyzed simultaneously. In this example, because tumor tissues generally have a low tumor purity, amplified regions contribute relatively more to tumor DNA, whereas deleted regions contribute relatively less to tumor DNA. Because tumor genomes are characterized by global hypomethylation, amplified regions have a lower methylation level compared to deleted regions. For example, in case TBR3033, the methylation level of chromosome 22 (increased copy number) measured using the present invention was 48.2%, lower than the methylation level of chromosome 3 (decreased copy number) (methylation level: 54.0%). In case TBR3032, the methylation level of chromosome 5p arm (increased copy number) measured using the present invention was 46.5%, lower than the methylation level of chromosome 5q arm (decreased copy number) (methylation level: 54.9%).
[0200] 2. Plasma DNA tissue mapping of pregnant women As shown in Figure 58, the accuracy of the methylation analysis allows us to compare the plasma DNA methylation profile of a pregnant woman with the methylation profiles of different reference tissues (e.g., liver, neutrophils, lymphocytes, placenta, T cells, B cells, heart, brain, etc.). Therefore, the contribution of DNA from different cell types in a pregnant woman's plasma DNA pool can be estimated using the following procedure: The CpG methylation levels of a DNA mixture (e.g., plasma DNA) determined according to embodiments present in the present disclosure are recorded in a vector (X), and the reference methylation levels retrieved across different tissues are recorded in a matrix (M) that can be quantified (including, but not limited to, bisulfite sequencing). The proportional contribution (p) from different tissues to the DNA mixture can be solved by, but not limited to, quadratic programming. Here, we use a mathematical equation to describe the estimation of the proportional contribution of different organs to the DNA mixture. The mathematical relationship between the methylation density of different sites in a DNA mixture and the methylation density of corresponding sites in different tissues can be expressed as follows:
number
number
[0201] Additional criteria can be included in the algorithm to improve accuracy. For example, the aggregate contribution of all cell types can be constrained to be 100%. Σ k p k =100% Furthermore, the contributions of all organs must be non-negative. P k ≧0,∀k
[0202] Due to biological variation, the observed global methylation pattern may not be completely identical to the methylation pattern estimated from tissue methylation. In such situations, mathematical analysis is required to determine the most likely proportional contribution of each tissue. In this regard, the difference between the observed methylation pattern in DNA and the methylation pattern estimated from tissue is denoted W.
number
[0203] each p kThe most likely value of can be determined by minimizing W, the difference between the observed and estimated methylation patterns. This equation can be solved using mathematical algorithms, including but not limited to, quadratic programming, linear / nonlinear regression, expectation-maximization (EM) algorithm, maximum likelihood estimation algorithm, maximum a posteriori probability estimation, and least squares.
[0204] As shown in Figure 59, using the method of plasma DNA tissue mapping shown in Figure 58, we observed that the placental DNA contribution to maternal plasma of pregnant women with male fetuses correlated well with the fetal DNA fraction estimated by Y chromosome reads. This result suggested the feasibility of using kinetics to track the tissue of origin of plasma DNA in pregnant women.
[0205] 3. Quantification of methylation levels in regions This section describes techniques for determining representative levels of methylation in selected genomic regions. This can be done using relatively low-level sequencing. Methylation levels can be determined per strand, per molecule, or per region using the number of methylated sites and the total number of methylated sites. Methylation levels in various tissues are also analyzed.
[0206] Eleven human tissue DNA samples were sequenced to a median of 30.7 million subreads per sample (range: 9.1-88.6 million) that could be aligned to the human reference genome (hg19). Subreads for each sample were generated from a median of 3.8 million Pacific Biosciences Single Molecular Real-Time (SMRT) sequencing wells (range: 1.1-11.5 million), with each well containing at least one subread that could be aligned to the human reference genome. On average, each molecule in an SMRT well was sequenced 9.9 times (range: 6.5-13.4 times). The human tissue DNA samples included one maternal buffy coat sample from a pregnant subject, one placenta sample, two hepatocellular carcinoma (HCC) tumor tissues, two HCC tissues and two adjacent non-tumor tissues paired with the two aforementioned HCC tissues, four buffy coat samples from healthy control subjects (M1 and M2 from male subjects, F1 and F2 from female subjects), and one HCC cell line (HepG2). A detailed summary of the sequencing data is shown in Figure 60.
[0207] Figure 60 shows different tissue groups in the first column and sample names in the second column. "Total subreads" indicates the total number of sequences generated from the SMRT well, including those from Watson and Crick strands. "Mapped subreads" lists the number of subreads that could be aligned to the human reference genome. "Subread mappability" refers to the percentage of subreads that could be aligned to the human reference genome. "Average subread depth per SMRT well" indicates the average number of subreads generated from each SMRT well. "Number of SMRT wells" refers to the number of SMRT wells that generated detectable subreads. "Mappable wells" indicates the number of wells containing at least one alignable subread. "Percentage of mappable wells" is the percentage of wells containing at least one alignable subread.
[0208] a) Methylation level and pattern analysis techniques In one embodiment, the methylation density of a single nucleic acid strand (e.g., DNA or RNA) can be measured, defined as the number of methylated bases in the strand divided by the total number of methylable bases in that strand. This measurement is also referred to as the "single-strand methylation level." This single-strand measurement is particularly feasible in the context of the present disclosure, since single-molecule real-time sequencing platforms can obtain sequencing information from each of the two strands of a double-stranded DNA molecule. This is facilitated by the use of hairpin adapters in preparing sequencing libraries, which allow the Watson and Crick strands of a double-stranded DNA molecule to be linked in a circular form and sequenced together. Indeed, this structure allows the partner Watson and Crick strands of the same double-stranded DNA molecule to be sequenced in the same reaction, allowing the methylation states of corresponding complementary sites in the Watson and Crick strands of any double-stranded DNA molecule to be determined individually and directly compared (e.g., Figures 20A and 20B).
[0209] These strand-based methylation analyses could not be easily achieved with other techniques. Without the direct methylation analysis methods disclosed in this application, other means must be applied to distinguish methylated from unmethylated bases, for example, by bisulfite conversion. Bisulfite conversion requires treatment of DNA with sodium bisulfite so that methylated and unmethylated cytosines can be distinguished as cytosine and thymine, respectively. Under the denaturing conditions of many bisulfite conversion protocols, the two strands of a double-stranded DNA molecule dissociate from each other. In many sequencing applications, for example, using the Illumina platform, the bisulfite-converted DNA is then amplified by polymerase chain reaction (PCR), which involves dissociation of the double-stranded DNA into single strands.
[0210] In Illumina sequencing, methylated adapters can be used prior to bisulfite conversion to prepare sequencing libraries without PCR. Even with this strategy, each DNA strand of a double-stranded DNA molecule is randomly selected for bridge amplification in the flow cell. Due to the random nature of sequencing, it is highly unlikely that each strand from the same DNA molecule will be sequenced in the same reaction. Even if two or more sequences from the same locus are analyzed in the same run, there is no simple way to determine whether the two reads are from the partner Watson and Crick strands of a single double-stranded DNA molecule, or from two different double-stranded DNA molecules. This consideration is important in certain embodiments of the present invention, as the two strands of a double-stranded DNA molecule may exhibit different methylation patterns. When single-strand methylation densities of multiple nucleic acid strands (e.g., DNA or RNA) are measured, a "multi-strand methylation level" can also be determined based on the concepts and formulas for "Methylation Level of Genomic Region of Interest" in Figure 61.
[0211] Figure 61 illustrates various methods for analyzing methylation patterns. A double-stranded DNA molecule (X) with unknown sequence and methylation information is ligated to an adapter, forming, in one example, a hairpin loop structure. As a result, the two single strands of the DNA molecule, including both the Watson strand X (a) and the Crick strand X (b), are physically linked in a circular fashion. The methylation status of sites in both the Watson and Crick strands can be obtained using methods described herein (e.g., using kinetic, electronic, electromagnetic, optical, or other types of physical signals from a sequencer). The Watson and Crick strands of the circularized DNA molecule can be examined in the same reaction. After sequencing, the adapter sequence is removed.
[0212] From the analysis, different methylation levels can be determined. In Figure 61 (I), the methylation pattern of only a single-stranded molecule, such as either X(a) or X(b), can be analyzed. This analysis can be referred to as single-strand methylation pattern analysis. The analysis can include, but is not limited to, determining the methylation status or methylation pattern of a site. In Figure 61, the single-stranded molecule X(a) exhibits a methylation pattern of 5'-UMMUU-3', where "U" indicates an unmethylated site and "M" indicates a methylated site, while its complementary single-stranded molecule X(b) exhibits a methylation pattern of 3'-UMUUU-5'. Therefore, X(b) has a different methylation pattern from X(a). The corresponding single-strand methylation levels of X(a) and X(b) are 40% and 20%, respectively.
[0213] In contrast, as shown in (II), methylation patterns can be analyzed at the single double-stranded DNA molecule level (i.e., considering the methylation patterns of both the Watson and Crick strands). This analysis can be referred to as single-molecule double-stranded DNA methylation pattern analysis. The methylation level of single-molecule double-stranded DNA for this exemplary molecule X is 30%. In one variant of this analysis, kinetic signals from both the Watson and Crick strands are combined to analyze modifications. In particular, because methylation at CpG sites is generally symmetric, kinetic signals from the Watson and Crick strands can be combined for a site before determining the methylation status of the site. In some situations, the performance of determining base modifications using the combined kinetic signals from the Watson and Crick strands of a molecule is superior to the performance of using the kinetic signals of a single strand independently. For example, as shown in Figure 20B, using the combined kinetic signals from both strands, including the Watson and Crick strands, yields a larger AUC (0.90) in the test dataset compared to using the single strands independently (AUC: 0.85).
[0214] In Figure 61 (III), the methylation level of a genomic region of interest is determined, and different DNA molecules with different molecular sizes and different numbers of methylable sites (e.g., CpG sites) may contribute to the genomic region of interest. This analysis is sometimes called multi-stranded methylation level analysis. The term "multi-stranded" can refer to multiple single-stranded DNA molecules, or multiple double-stranded DNA molecules, or any combination thereof. In this example, there are three double-stranded DNA molecules covering the genomic region of interest: molecule "X", molecule "Y", and molecule "Z", each having an "a" strand and a "b" strand. The corresponding methylation level of this region is 9 / 28, or 32%. The size of the genomic region to be analyzed can be 1 nt, 10 nt, 20 nt, 30 nt, 40 nt, 50 nt, 100 nt, 1 knt (kilonucleotide, i.e., 1000 nucleotides), 2 knt, 3 knt, 4 knt, 5 knt, 10 knt, 20 knt, 30 knt, 40 knt, 50 knt, 100 knt, 200 knt, 300 knt, 400 knt, 500 knt, 1 Mnt (meganucleotide, i.e., 1 million nucleotides), 2 Mnt, 3 Mnt, 4 Mnt, 5 Mnt, 10 Mnt, 20 Mnt, 30 Mnt, 40 Mnt, 50 Mnt, 100 Mnt, or 200 Mnt. The genomic region can be a chromosome arm or the entire genome.
[0215] Methylation patterns can also be determined after determining the methylation status of sites within a molecule. For example, in a scenario where there are three consecutive CpG sites on a single double-stranded DNA molecule, the methylation patterns of each of the Watson and Crick strands can be revealed as methylated (M), unmethylated (N), and methylated (M) for the three sites. This pattern, e.g., MNM for the Watson strand, can be referred to as the "methylation haplotype" of the Watson strand in this region. Due to the presence of DNA methylation maintenance activity, the methylation patterns of the Watson and Crick strands of a double-stranded DNA molecule can be complementary to each other. For example, if a CpG site on the Watson strand is methylated, the complementary CpG site on the Crick strand may also be methylated. Similarly, an unmethylated CpG site on the Watson strand may be complementary to an unmethylated CpG site on the Crick strand.
[0216] In one embodiment, the methylation level of a single DNA molecule can be measured, which is defined as the number of methylated bases or nucleotides in the molecule divided by the total number of methylable bases or nucleotides in that molecule. This measurement is also referred to as the "single-molecule methylation level." This single-molecule measurement may be particularly useful in the context of the present disclosure due to the long read lengths possible with single-molecule real-time sequencing platforms. When measuring the single-molecule methylation levels of multiple DNA molecules, a "multiple-molecule methylation level" can also be determined based on the concept and formula in Figure 61. For example, the "multiple-molecule methylation level" can be the average or median of the single-molecule methylation levels.
[0217] In some embodiments, one or more genetic polymorphisms (e.g., single nucleotide polymorphisms (SNPs)) can be analyzed on a DNA molecule along with the methylation status of sites on the molecule, thus revealing both the genetic and epigenetic information of the molecule. Such analysis reveals the "methylation haplotype" of the analyzed DNA molecule. Methylation haplotype analysis is useful, for example, for studying genomic imprinting in maternal plasma and cell-free nucleic acids (containing a mixture of cell-free DNA molecules with maternal and fetal genetic and epigenetic characteristics).
[0218] b) Comparison of methylation results Genome-wide methylation densities for the tissues in the table in Figure 60 are determined using bisulfite sequencing and single-molecule real-time sequencing, as described in this disclosure. Figure 62A shows methylation densities quantified by bisulfite sequencing on the y-axis, with tissue type on the x-axis. Figure 62B shows methylation densities quantified by single-molecule real-time sequencing, as described in this disclosure, on the y-axis, with tissue type on the x-axis.
[0219] Figure 62A shows the methylation density across different tissues using bisulfite sequencing (i.e., samples were bisulfite converted and then subjected to Illumina sequencing) (Lister et al. Nature. 2009;462:315-322), including HepG2, HCC tumor tissue, matched HCC tumor-adjacent normal liver tissue (i.e., adjacent normal tissue), placenta tissue, and buffy coat sample. HepG2 showed the lowest methylation level at 40.4%. Buffy coat sample showed the highest methylation level at 76.5%. The mean methylation density of HCC tumor tissue (51.2%) was found to be lower than that of matched adjacent normal tissue (71.0%). This is consistent with the expectation that HCC tumors are hypomethylated at the genome-wide level compared to adjacent normal tissue (Ross et al. Epigenomics. 2010;2:245-69). The dataset was generated from DNA prepared by the Sequel II Sequencing Kit 1.0.
[0220] A portion of the same tissue was subjected to methylation analysis using single-molecule real-time sequencing and the method disclosed herein. The results are shown in Figure 62B. Methylation analysis using the single-molecule real-time sequencing method disclosed herein showed that the HepG2 cell line was the most hypomethylated, followed by the analyzed HCC tumor tissue, and then the placenta tissue. The adjacent non-tumorous liver tissue sample was more methylated than other tissues, including HCC and placenta tissue, with the buffy coat being the most hypermethylated.
[0221] Figures 63A, 63B, and 63C show the correlation of global methylation levels quantified by bisulfite sequencing and single-molecule real-time sequencing according to the methods described herein. Figure 63A shows the methylation levels quantified by bisulfite sequencing on the x-axis and the methylation levels quantified by single-molecule real-time sequencing using the methods described herein on the y-axis. The solid black line is the fitted regression line. The dashed line indicates where the two measurements are equal.
[0222] There was a very high correlation between methylation levels between bisulfite sequencing and single-molecule real-time sequencing according to the present disclosure (r = 0.99, P value < 0.0001). These data demonstrated that methylation analysis using the single-molecule real-time sequencing method disclosed herein is an effective means for determining methylation levels between tissues and allows for comparison of the methylation status and methylation profiles between these tissues. We noted that the slope of the regression line in Figure 63A deviated from 1 for the two measurements of methylation levels. These results suggested that there may be a deviation between the two measurements (in some contexts, this deviation may be referred to as bias) in the determination of methylation levels using single-molecule real-time sequencing according to the present disclosure compared to conventional massively parallel bisulfite sequencing.
[0223] In one embodiment, linear or LOESS (locally weighted smoothing) regression can be used to quantify bias. As an example, if massively parallel bisulfite sequencing (Illumina) is considered as the reference, the results determined by single-molecule real-time sequencing according to the present disclosure can be transformed using regression coefficients to adjust the readout between different platforms. In Figure 63A, the linear regression equation is Y = aX + b, where "Y" represents the methylation level determined by single-molecule real-time sequencing according to the present disclosure, "X" represents the methylation level determined by bisulfite sequencing, "a" represents the slope of the regression line (e.g., a = 0.62), and "b" represents the intercept of the y-axis (e.g., b = 17.72). In this case, the adjusted methylation value determined by single-molecule real-time sequencing is calculated by (Yb) / a. In another embodiment, the deviation (ΔM) of two measurements and the corresponding mean of the two measurements are used.
number
number
[0224] Figure 63B shows the relationship between ΔM and
number
number
[0225] Figure 63C shows the average of two measurements.
number
number
[0226] Conventional whole-genome bisulfite sequencing (Illumina) has been reported to have considerable variability in quantification of methylation levels between methods in certain genomic regions, resulting in significantly biased sequence output and overestimated overall methylation (Olova et al. Genome Biol. 2018;19:33). The methods disclosed herein can be performed without bisulfite conversion, which dramatically degrades DNA, and without PCR amplification, which can complicate the process or introduce additional errors into the determination of methylation levels.
[0227] Figures 64A and 64B show methylation patterns at 1 Mb resolution. Figure 64A shows the methylation pattern of an HCC cell line (HepG2). Figure 64B shows the methylation pattern of a buffy coat sample from a healthy control subject. Chromosome ideograms (the outermost rings in each figure) are organized clockwise from the p-terminus to the q-terminus. The second outermost ring (also called the middle ring) shows methylation levels determined by bisulfite sequencing. The innermost ring shows methylation levels determined by single-molecule real-time sequencing according to the present disclosure. Methylation levels are classified into five grades: 0-20% (light green), 20-40% (green), 40-60% (blue), 60-80% (light red), and 80-100% (red). As shown in Figures 64A and 64B, the methylation profiles at 1 Mb resolution were consistent between bisulfite sequencing (middle track) and single molecule real-time sequencing (innermost track) according to the present disclosure. The methylation level of the maternal buffy coat sample was shown to be higher than that of the HCC cell line (HepG2).
[0228] Figures 65A and 65B show scatter plots of methylation levels measured at 1 Mb resolution. Figure 65A shows the methylation levels of an HCC cell line (HepG2). Figure 65B shows the methylation levels of a buffy coat sample from a healthy control subject. For both Figures 65A and 65B, methylation levels quantified by bisulfite sequencing are on the x-axis, and methylation levels measured by single-molecule real-time sequencing according to the present disclosure are on the y-axis. The solid line is the fitted regression line. The dashed line indicates where the two measurement techniques are equal. For HCC cell lines, methylation levels determined by single-molecule real-time sequencing at 1 Mb resolution correlated well with those measured by bisulfite sequencing (r = 0.99, P < 0.0001) (Figure 65A). Correlation was also observed for data from buffy coat samples (r = 0.87, P < 0.0001) (Figure 65B).
[0229] Figures 66A and 66B show scatter plots of methylation levels measured at 100 kb resolution. Figure 66A shows the methylation levels of an HCC cell line (HepG2). Figure 66B shows the methylation levels of a buffy coat sample from a healthy control subject. For both Figures 66A and 66B, methylation levels quantified by bisulfite sequencing are on the x-axis, and methylation levels measured by single-molecule real-time sequencing according to the present disclosure are on the y-axis. The solid line is the fitted regression line. The dashed line indicates where the two measurement techniques are equal. When the analytical resolution was increased by 100 kb (or 100 knt) windows, a high degree of correlation was also observed between methylation quantification measurements between the two methods at 1 Mb (or 1 Mnt) resolution. All these data indicate that the disclosed single-molecule real-time approach is an effective tool for quantifying methylation levels or densities within genomic regions varying at different degrees of resolution, e.g., 1 Mb (or 1 Mnt) or 100 kb (or 100 knt). The data also indicate that the present invention is an effective tool for assessing methylation profiles or patterns across regions or samples.
[0230] Figures 67A and 67B show methylation patterns at 1 Mb resolution. Figure 67A shows the methylation pattern of HCC tumor tissue (TBR3033T). Figure 67B shows the methylation pattern of adjacent normal tissue (TBR3033N). Chromosome ideograms (the outermost rings in each figure) are organized clockwise from the p-terminus to the q-terminus. The second outermost ring (also called the middle ring) shows methylation levels determined by bisulfite sequencing. The innermost ring shows methylation levels determined by single-molecule real-time sequencing according to the present disclosure. Methylation levels are classified into five grades: 0-20% (light green), 20-40% (green), 40-60% (blue), 60-80% (light red), and 80-100% (red). As shown in Figure 67A, hypomethylation in HCC tumor tissue DNA (TBR3033T) can be detected and distinguished from adjacent normal liver tissue DNA (TBR3033N) in Figure 67B. The methylation levels and patterns determined by bisulfite sequencing (middle track) and single molecule real-time sequencing (innermost track) according to the present disclosure were consistent. The methylation levels of adjacent normal tissue DNA were shown to be higher than those of HCC tumor tissue DNA.
[0231] Figures 68A and 68B show scatter plots of methylation levels measured at 1 Mb resolution. Figure 68A shows the methylation levels of HCC tumor tissue (TBR3033T). Figure 68B shows the methylation levels of adjacent normal tissue. For both Figures 68A and 68B, methylation levels quantified by bisulfite sequencing are on the x-axis, and methylation levels measured by single-molecule real-time sequencing according to the present disclosure are on the y-axis. The solid line is the fitted regression line. The dashed line indicates where the two measurement techniques are equal. For HCC tumor tissue DNA, methylation levels measured by single-molecule real-time sequencing at 1 Mb resolution correlated well with those determined by bisulfite sequencing (r = 0.96, P value < 0.0001) (Figure 68A). Data from adjacent normal liver tissue samples also correlated (r = 0.83, P value < 0.0001) (Figure 68B).
[0232] Figures 69A and 69B show scatter plots of methylation levels measured at 100 kb resolution. Figure 69A shows the methylation levels of HCC tumor tissue (TBR3033T). Figure 69B shows the methylation levels of adjacent normal tissue (TBR3033N). For both Figures 69A and 69B, methylation levels quantified by bisulfite sequencing are on the x-axis, and methylation levels measured by single-molecule real-time sequencing according to the present disclosure are on the y-axis. The solid line is the fitted regression line. The dashed line indicates where the two measurement techniques are equal. Such a high degree of correlation of methylation quantification data between the two methods at 1 Mb resolution was observed even when methylation level measurements were performed at higher resolution, e.g., in 100 kb windows.
[0233] Figures 70A and 70B show the methylation patterns at 1 Mb resolution for other tumor and normal tissues. Figure 70A shows the methylation pattern for HCC tumor tissue (TBR3032T). Figure 70B shows the methylation pattern for adjacent normal tissue (TBR3032N). Chromosome ideograms (the outermost rings in each figure) are organized clockwise from the p-terminus to the q-terminus. The second outermost ring (also called the middle ring) shows the methylation levels determined by bisulfite sequencing. The innermost ring shows the methylation levels determined by single-molecule real-time sequencing according to the present disclosure. Methylation levels are classified into five grades: 0-20% (light green), 20-40% (green), 40-60% (blue), 60-80% (light red), and 80-100% (red). As shown in Figure 70A, we were able to detect hypomethylation in HCC tumor tissue DNA (TBR3032T) and distinguish it from adjacent normal liver tissue DNA (TBR3032N) in Figure 70B. The methylation levels and patterns determined by bisulfite sequencing (middle track) and single-molecule real-time sequencing using the present invention (innermost track) were consistent. The methylation levels of adjacent normal tissue DNA were shown to be higher than those of HCC tumor tissue DNA.
[0234] Figures 71A and 71B show scatter plots of methylation levels measured at 1 Mb resolution. Figure 71A shows the methylation levels of HCC tumor tissue (TBR3032T). Figure 71B shows the methylation levels of adjacent normal tissue. For both Figures 71A and 71B, methylation levels quantified by bisulfite sequencing are on the x-axis, and methylation levels measured by single-molecule real-time sequencing according to the present disclosure are on the y-axis. The solid line is the fitted regression line. The dashed line indicates where the two measurement techniques are equal. For HCC tumor tissue DNA, methylation levels measured by single-molecule real-time sequencing at 1 Mb resolution correlated well with those determined by bisulfite sequencing (r = 0.98, P < 0.0001) (Figure 71A). Data from adjacent normal liver tissue samples also correlated (r = 0.87, P < 0.0001) (Figure 71B).
[0235] Figures 72A and 72B show scatter plots of methylation levels measured at 100 kb resolution. Figure 72A shows the methylation levels of HCC tumor tissue (TBR3032T). Figure 72B shows the methylation levels of adjacent normal tissue (TBR3032N). For both Figures 72A and 72B, methylation levels quantified by bisulfite sequencing are on the x-axis, and methylation levels measured by single-molecule real-time sequencing according to the present disclosure are on the y-axis. The solid line is the fitted regression line. The dashed line indicates where the two measurement techniques are equal. Such a high degree of correlation of methylation quantification data between the two methods at 1 Mb resolution was observed even when methylation level measurements were performed at higher resolution, e.g., in 100 kb windows.
[0236] 4. Variable methylation regions between tumor and adjacent normal tissue Methylomic abnormalities are common in regions of the cancer genome. One example of such an abnormality is hypomethylation and hypermethylation of selected genomic regions (Cadieux et al. Cancer Res. 2006;66:8469-76, Graff et al. Cancer Res. 1995;55:5195-9, Costello et al. Nat Genet. 2000;24:132-8). Another example is abnormal patterns of methylated and unmethylated bases in selected genomic regions. This section demonstrates that techniques for determining methylation can be used to perform quantitative analysis and diagnosis when analyzing tumors.
[0237] Figure 73 shows an example of an abnormal methylation pattern near the tumor suppressor gene CDKN2A. The coordinates are highlighted in blue, and the underlines indicate CpG islands. Black dots indicate methylated sites. Open dots indicate unmethylated sites. The numbers in parentheses to the right of each dotted horizontal line indicate the fragment size, single-molecule methylation density, and number of CpG sites. For example, (3.3 kb, MD: 17.9%, CG: 39) means that the fragment size is 3.3 kb, the fragment methylation level is 17.9%, and the number of CpG sites is 39. MD represents methylation density.
[0238] As shown in Figure 73, the CDKN2A (cyclin-dependent kinase inhibitor 2A) gene encodes two proteins, INK4A (p16) and ARF (p14), which act as tumor suppressors. Two molecules (molecule 7301 and molecule 7302) were found to cover the region overlapping with the CDKN2A gene in non-tumor tissue adjacent to the tumor tissue. The methylation levels of the single double-stranded DNA molecules of molecules 7301 and 7302 were shown to be 17.9% and 7.6%, respectively. In contrast, the methylation level of the single double-stranded DNA molecule of molecule 7303 present in the tumor tissue was found to be 93.9%, which was much higher than the methylation level of the molecule present in the paired adjacent non-tumor tissue. Meanwhile, the multistranded methylation level could also be calculated using molecules 7301 and 7302 present in the non-tumor tissue adjacent to the tumor tissue. The resulting multistranded methylation level was 9.7%, which was lower than the methylation level in the tumor tissue (93.9%). The differential methylation levels suggest that single-stranded molecule methylation levels and / or multi-stranded methylation levels can be used to detect or monitor diseases such as cancer.
[0239] Figures 74A and 74B show variably methylated regions detected by single-molecule real-time sequencing according to an embodiment of the present invention. Figure 74A shows hypomethylation in a cancer genome. Figure 74B shows hypermethylation in a cancer genome. The x-axis shows the coordinates of CpG sites. The coordinates are highlighted in blue, and underlines indicate CpG islands. Black dots indicate methylated sites. Open dots indicate unmethylated sites. The numbers in parentheses to the right of each dotted horizontal line indicate the fragment size, fragment-level methylation density, and number of CpG sites. For example, (3.1 kb, MD: 88.9%, CG: 180) means that the fragment size is 3.1 kb, the fragment methylation density is 88.9%, and the number of CpG sites is 180.
[0240] Figure 74A shows a region near the GNAS gene that shows more hypomethylated fragments in HCC tumor tissues compared to adjacent normal liver tissues. Figure 74B shows a region near the ESR1 gene that displays hypermethylated fragments in HCC tissues, while DNA fragments from bare adjacent non-tumor tissues that align to the corresponding region instead show hypomethylation. As shown in Figure 74B, the methylation profiles or methylation haplotypes of individual DNA molecules were sufficient to reveal the abnormal methylation status of these genomic regions, i.e., GNAS and ESR1, when cancer samples were compared with non-cancer samples.
[0241] These data demonstrate that the single-molecule real-time sequencing methylation analysis disclosed herein can determine the methylation status at each CpG site (methylated or unmethylated) on individual DNA fragments. The read lengths of single-molecule real-time sequencing are typically much longer (on the order of kilobases) than those of Illumina sequencing, which can range from 100 to 300 nt per read (De Maio et al. Micob Genom. 2019;5(9)). Combining the long read length characteristics of single-molecule real-time sequencing with the methylation analysis methods disclosed herein can easily determine the methylation haplotypes of multiple CpG sites along any single DNA molecule. A methylation profile refers to the methylation status of CpG sites from one genomic coordinate to another within a contiguous stretch of DNA (e.g., on the same chromosome, within a bacterial plasmid, or within a single DNA stretch within a virus).
[0242] Because single-molecule real-time sequencing analyzes each DNA molecule individually without requiring prior amplification, the methylation profile determined for an individual DNA molecule is actually a methylation haplotype, which refers to the methylation status of CpG sites from one end of the same DNA molecule to another. When more than one molecule is sequenced from the same genomic region, the % methylation (i.e., methylation level or methylation density) of each CpG site across all sequenced CpG sites within the genomic region can be aggregated from data from multiple DNA fragments using the same formula, as shown in Figure 61. The % methylation of each CpG site is reported for all sequenced CpG sites, providing a methylation profile for the sequenced genomic region. Alternatively, data from all reads and all sites within the sequenced genomic region can be aggregated to provide a 1% methylation value for the region in the same manner as the methylation level of a 1 Mb or 1 kb region is calculated, as shown in Figures 64-72.
[0243] 5. Viral DNA Methylation Analysis This section demonstrates that the methylation techniques of the present disclosure can be used to accurately determine the methylation levels of viral DNA.
[0244] Figure 75 shows the methylation patterns of hepatitis B virus DNA between two pairs of HCC tissue samples and adjacent non-tumor tissue samples using single-molecule real-time sequencing. Each arrow represents a gene annotation of the HBV genome. Arrows with "P," "S," "X," and "C" indicate gene annotations for the HBV genome, encoding polymerase, surface antigen, X protein, and core protein, respectively. We identified one fragment (molecule I, spanning 2,278–3,141 of the HBV genome highlighted by a dashed rectangle) measuring 1,183 bp in size from the adjacent non-tumor tissue, which showed a 12% methylation level. We also identified three fragments (molecule II, molecule III, and molecule IV) measuring 3,215 bp, 2,961 bp, and 3,105 bp from the tumor tissue. Among them, two fragments (molecule III and molecule IV) from the HCC tumor overlapped with the HBV genome region spanned by molecule I from the non-tumor tissue. In contrast to the low methylation level (12%) of the HBV region highlighted by the dashed rectangle (HBV genome locations: 2,278–3,141), the methylation level was higher (i.e., 24% and 30%) in those fragments (molecule III and molecule IV) of HCC tissues. These results suggested that an approach using single-molecule real-time sequencing is feasible to determine the methylation pattern of the viral genome and identify variably methylated regions (DMRs) of HBV between HCC and non-HCC tissues. Therefore, the determination of the methylation status of the entire viral genome using single-molecule real-time sequencing according to this disclosure will provide a new tool for studying clinical relevance using tissue biopsies.
[0245] This DMR region happened to overlap with genes P, C, and S. This region was also reported to be hypermethylated in HCC tissues compared with non-cancer liver tissues in the presence of HBV infection (Jain et al. Sci Rep. 2015;5:10478, Fernandez et al. Genome Res. 2009;19:438-51).
[0246] We pooled the bisulfite sequencing results of liver tissue from four patients with cirrhosis but without HCC and obtained 1,156 HBV fragments for methylation analysis. Figure 76A shows the methylation levels of hepatitis B virus DNA in liver tissue from patients with cirrhosis but without HCC. Furthermore, we pooled the bisulfite sequencing results of HCC tumor tissue from 15 patients and obtained 736 HBV fragments for methylation analysis. Figure 76B shows the methylation levels of hepatitis B virus DNA in HCC tumor tissue. As shown in Figures 76A and 76B, massively parallel bisulfite sequencing also revealed that the HBV DMR region (HBV genome positions 1,982–2,435) had higher methylation levels in HCC tissue than in cirrhotic liver tissue. These results suggested that our approach for determining the methylation status of the viral genome is effective.
[0247] 6. Variant-associated Methylation Analysis Different alleles can be associated with different methylation profiles. For example, an imprinted gene may have one allele that has a higher level of methylation than the other alleles. This section demonstrates that methylation profiles can be used to identify alleles in a particular genomic region.
[0248] A single-molecule real-time sequencing well containing a single DNA template generates several subreads. The subreads include kinetic characteristics (e.g., interpulse interval (IPD) and pulse width (PW)) and nucleotide composition. In one embodiment, the subreads from one single-molecule real-time sequencing well can be used to generate a consensus sequence (also called a circular consensus sequence, CCS), which can dramatically reduce sequencing errors (e.g., mismatches, insertions, or deletions). Details of CCS are described herein. In one embodiment, a consensus sequence can be constructed using the subreads aligned to the human reference genome. In another embodiment, a consensus sequence can be constructed by mapping the subreads to the longest subread within the same single-molecule real-time sequencing well.
[0249] Figure 77 shows the principle of stepwise methylation haplotype analysis. Filled lollipops represent CpG sites classified as methylated. Unfilled lollipops represent CpG sites classified as unmethylated.
[0250] As shown in one embodiment of Figure 77, the subreads were aligned to the human reference genome. The aligned subreads from one single-molecule real-time sequencing well were combined to form a consensus sequence. The consensus sequence can generally be determined using the most frequent nucleotide present in the subreads across each aligned position. Thus, nucleotide variants, including but not limited to single nucleotide polymorphisms, insertions, and deletions, can be identified from the consensus sequence. According to the present disclosure, methylation patterns can be determined using the averaged IPD and PW within the same molecule tagged by the nucleotide variants. Thus, variant-associated methylation patterns can be further determined. The methylation status of the same molecule can be considered a methylation haplotype. Methylation haplotypes may not be easily and directly constructed from two or more short DNA molecules due to the lack of molecular markers that can distinguish whether two or more fragmented short DNA molecules are derived from a single original molecule or whether two or more different original molecules contribute. Synthetic long-read technologies (such as linked-read sequencing developed by 10X Genomics) offer the possibility to distribute a single long DNA molecule into partitions (such as droplets) and tag short DNA molecules derived from that long DNA molecule with the same molecular barcode sequence. However, this barcoding step involves PCR amplification, which does not preserve the original methylation state.
[0251] Furthermore, when attempting to process long DNA molecules using bisulfite, because bisulfite only acts on single-stranded DNA molecules under certain chemical conditions, the first step prior to bisulfite treatment involves DNA denaturation under destructive conditions, converting double-stranded DNA to single-stranded DNA. This DNA denaturation step breaks long DNA molecules into short fragments, resulting in the loss of the original methylation haplotype information. A second drawback of bisulfite-based methylation analysis is that the bisulfite conversion step denatures double-stranded DNA into single-stranded DNA, i.e., Watson and Crick strands. For a single molecule, there is a 50% chance of sequencing the Watson strand and a 50% chance of sequencing the Crick strand. With millions of Watson and Crick strands, the chance of simultaneously sequencing both the Watson and Crick strands of a molecule is extremely low. Even assuming that both the Watson and Crick strands of a molecule are sequenced, it is still impossible to reliably determine whether such Watson and Crick strands originate from a single original fragment or whether two or more distinct original fragments contribute. Liu et al. recently introduced a bisulfite-free sequencing method for detecting methylated and hydroxymethylated cytosines under mild conditions that reduce DNA degradation using a ten-eleven translocation (TET) enzyme-based conversion (Liu et al. Nat Biotechnol. 2019;37:424-429). However, the enzymatic reaction involves two sequential steps. A low conversion rate in either step of the enzymatic reaction dramatically affects the overall conversion rate. Furthermore, even with this bisulfite-free sequencing method for detecting methylated cytosines, it is still difficult to distinguish the Watson and Crick strands of a molecule from the sequencing results.
[0252] In contrast, in embodiments of the present invention, the Watson and Crick strands of a molecule are covalently linked via bell-shaped adapters to form a circular DNA molecule, so that both the Watson and Crick strands of the molecule can be sequenced in the same reaction well to determine the methylation state of each strand.
[0253] One advantage of embodiments of the present invention is the ability to ascertain methylation and genetic (i.e., sequence) information for long, contiguous DNA molecules (equibases or kilonucleotides in length). Generating such information using short-read sequencing technologies is more challenging. With short-read sequencing technologies, combining sequencing information for multiple short reads using genetic or epigenetic anchors is necessary to be able to infer long stretches of methylation and genetic information. However, this will prove challenging in many scenarios due to the distance between such genetic or epigenetic anchors. For example, while there is an average of one SNP per kb, current short-read sequencing technologies can typically sequence a maximum of 300 nt per read, or 600 nt even in paired-end format.
[0254] In one embodiment, variant-associated methylation haplotype analysis can be used to study the methylation patterns of imprinted genes. Imprinted regions are subject to epigenetic regulation (e.g., CpG methylation) in a parent-of-origin manner. For example, in the table in Figure 60, one buffy coat DNA sample (M2) was sequenced to obtain approximately 152 million subreads. In this sample, 53% of single-molecule real-time sequencing wells generated at least one subread that could be aligned to the human reference genome. The average subread depth for each SMRT well was 7.7x. In total, approximately 3 million consensus sequences were obtained. Approximately 91% of the reference genome was covered by the consensus sequence at least once. For the covered regions, the sequencing depth was 7.9x. The dataset was generated from DNA prepared by Sequel II Sequencing Kit 1.0.
[0255] Figure 78 shows the size distribution of sequenced molecules determined from the consensus sequence, with a median size of 6,289 bp (range: 66-198,109 bp). Fragment size (bp) is shown on the x-axis, and the frequency (%) associated with fragment size is shown on the y-axis.
[0256] Figures 79A, 79B, 79C, and 79D show examples of allele methylation patterns in imprinted regions. The x-axis indicates the coordinates of the CpG sites. The coordinates are highlighted in blue, and underlines indicate CpG islands. Black dots indicate methylated CpG sites. Unfilled dots indicate unmethylated CpG sites. The embedded letters between the filled and unfilled dots (i.e., CpG sites) in each horizontal series indicate the alleles at the SNP site. The numbers in parentheses to the right of each horizontal series indicate the fragment size, fragment-level methylation density, and number of CpG sites. For example, (10.0 kb, MD: 79.1%, CG: 139) indicates that the corresponding fragment size is 10.0 kb, the fragment methylation density is 79.1%, and the number of CpG sites is 139. The dashed rectangle outlines the most variably methylated region within each gene.
[0257] Figure 79A shows 11 sequenced fragments from the SNURF gene, with a median size of 11.2 kb (range: 1.3-25 kb). The SNURF gene is maternally imprinted, meaning that the copy of the gene an individual inherits from their mother is methylated and transcriptionally silent. As shown in Figure 79A, in the dashed rectangle, fragments associated with the C allele were highly methylated, while fragments associated with the T allele were highly unmethylated. High methylation indicates methylation of 70%, 80%, 90%, 95%, or 99% or more of the site. Allele-specific methylation patterns could be observed in other imprinted genes, including PLAGL1 (Figure 79B), NAP1L5 (Figure 79C), and ZIM2 (Figure 79D). Figure 79B shows that in the case of PLAGL1, fragments associated with the T allele were highly unmethylated, whereas fragments associated with the C allele were highly methylated. Figure 79C shows that for NAP1L5, the fragment associated with the C allele was highly unmethylated, whereas the fragment associated with the T allele was highly methylated. Figure 79D shows that for ZIM2, the fragment associated with the C allele was highly unmethylated, whereas the fragment associated with the T allele was highly methylated.
[0258] Figures 80A, 80B, 80C, and 80D show examples of allele methylation patterns in non-imprinted regions. The x-axis shows the coordinates of CpG sites. The coordinates are highlighted in blue, and underlines indicate CpG islands. Black dots indicate methylated CpG sites. Open dots indicate unmethylated CpG sites. The embedded letters between each horizontal series of filled and unfilled dots (i.e., CpG sites) indicate the alleles at the SNP site. The numbers in parentheses to the right of each horizontal series of dots indicate the fragment size, fragment-level methylation density, and number of CpG sites. The dashed rectangles indicate randomly selected regions for calculating the methylation densities reported in parentheses. In contrast to the results in Figures 79A-79D, no observable allelic methylation patterns were observed in non-imprinted genes. Figure 80A shows no differences in allelic methylation patterns in the chr7 region. Figure 80B shows no differences in allelic methylation patterns in the chr12 region. Figure 80C shows no differences in allelic methylation patterns in the chr1 region. Figure 80D shows no differences in allelic methylation patterns in another chr1 region.
[0259] Figure 81 shows a table of allele-specific fragment methylation levels. The first column lists the categories of "imprinted genes" and "randomly selected regions." The second column lists the specific genes. The third column lists the first alleles of the gene's SNPs. The fourth column lists the second alleles of the gene's SNPs. The fifth column shows the methylation levels of fragments linked to the first alleles. The sixth column shows the methylation levels of fragments linked to the second alleles. The methylation levels of fragments linked to allele 2 (mean: 88.6%, range: 84.6-91.1%) are much higher than those of fragments linked to allele 1 of those imprinted genes (mean: 12.2%, range: 7.6-15.7%) (P value = 0.03), indicating the presence of allele-specific methylation. In contrast, there was no significant change in methylation levels between those randomly selected regions (P value = 1), suggesting the absence of allele-specific methylation.
[0260] 7. Cell-free DNA analysis during pregnancy In this example, we demonstrate that the method disclosed herein can be applied to the analysis of cell-free nucleic acids in plasma or serum obtained from a pregnant woman with at least one fetus.During pregnancy, cell-free DNA molecules and cell-free RNA molecules from placental cells are found in maternal circulation.Such placenta-derived cell-free nucleic acid molecules are also referred to as cell-free fetal nucleic acids in maternal plasma or circulating cell-free fetal nucleic acids.Cell-free fetal nucleic acids exist in maternal plasma in the context of maternal cell-free nucleic acids.For example, circulating cell-free fetal DNA molecules exist as rare species in the context of cell-free maternal DNA in maternal plasma and serum.
[0261] It is known that genetic or epigenetic means, or a combination thereof, can be used to distinguish cell-free fetal DNA from cell-free maternal DNA in maternal plasma or serum. Genetically, the fetal genome may differ from the maternal genome due to paternally inherited fetal-specific SNP alleles, paternally inherited mutations, or de novo mutations. Epigenetically, the placental methylome is generally hypomethylated compared to the methylome of maternal blood cells (Lun et al. Clin Chem. 2013;59:1583-94). Because the placenta is the major contributor of cell-free fetal DNA, while maternal blood cells are the major contributor of cell-free maternal DNA in maternal circulation (plasma or serum), cell-free fetal DNA molecules are generally hypomethylated compared to cell-free maternal DNA in plasma or serum. There are certain genomic loci that are hypermethylated in the placenta compared to maternal blood cells. For example, the promoter and exon 1 regions of RASSF1A are more methylated in the placenta than in maternal blood cells (Chiu et al. Am J Pathol. 2007;170:941-950). Thus, circulating cell-free fetal DNA derived from this RASSF1A locus is hypermethylated compared with circulating cell-free maternal DNA derived from the same locus.
[0262] In embodiments, cell-free fetal DNA can be distinguished from cell-free maternal DNA molecules based on the different methylation states between the two pools of circulating nucleic acids.For example, if the CpG sites along the cell-free DNA molecule are found to be mostly unmethylated, this molecule may be derived from the fetus.If the CpG sites along the cell-free DNA molecule are found to be mostly methylated, this molecule is likely to be derived from the mother.There are several methods known to those skilled in the art to confirm whether such molecules are actually derived from the fetus or the mother.One approach is to compare the methylation pattern of the sequenced molecule with the known methylation profile of the corresponding locus in placenta or maternal blood cells.
[0263] Figure 82 shows an example of using methylation profiles to determine the placental origin of plasma DNA during pregnancy. Coordinates are highlighted in blue, and underlines indicate CpG islands. Black dots indicate methylated sites. Open dots indicate unmethylated sites. Numbers in parentheses near each dotted horizontal line indicate fragment size, single-molecule methylation density, and number of CpG sites.
[0264] As shown in Figure 82, if a maternal plasma cell-free DNA molecule aligns to the promoter region of RASSF1A (a region known to be specifically methylated in placental tissue) and the sequencing data generated using the methods of the present invention are hypermethylated, the molecule is likely derived from the fetus or placenta. In contrast, molecules showing hypomethylation are likely derived from maternal background DNA (primarily of hematopoietic origin).
[0265] Figure 83 shows an approach for fetal-specific methylation analysis. This approach involves utilizing sequenced molecules containing fetal-specific SNP alleles or fetal-specific mutations (e.g., paternally inherited or essentially de novo). When such fetal-specific genetic features are identified, the methylation status of bases present in the same cell-free DNA molecule reflects the methylation profile of cell-free fetal DNA or placental methylome. When plasma cell-free DNA sequencing reveals alleles or mutations not present in the maternal genome (e.g., by analysis of maternal genomic DNA), or when known to be transmitted familialally (e.g., by analysis of DNA from a proband), fetal-specific genetic features can be revealed.
[0266] Methylation of fetal-specific DNA molecules can be determined by analyzing those DNA fragments that have alleles different from the homozygous alleles in the maternal genome. Methylation of fetal DNA molecules can be expected to be lower than that of maternal DNA molecules.
[0267] As an example, we sequenced buffy coat DNA and corresponding placental DNA from one pregnant woman to obtain 59x and 58x haploid genome coverage, respectively. We identified a total of 822,409 informative SNPs, homozygous in the mother and heterozygous in the fetus. Through single-molecule real-time sequencing, we found 2,652 fetal-specific fragments and 24,837 shared fragments (i.e., fragments with shared alleles, primarily maternal-derived) in maternal plasma (M13160). The fetal DNA fraction was 19.3%. According to the present disclosure, the methylation profiles of these fetal-specific and shared fragments were estimated. The results showed that the methylation level of fetal-specific fragments was 57.4%, while the methylation level of shared fragments was 69.9%. This finding was consistent with current knowledge that fetal DNA methylation levels are lower than maternal DNA in the plasma of pregnant women (Lun et al., Clin Chem. 2013;59:1583-94).
[0268] Methylation patterns can be used for diagnostic or monitoring purposes. For example, methylation profiles of maternal plasma samples have been used to determine gestational age (https: / / www.ncbi.nlm.nih.gov / pubmed / 27979959). One application is as a quality control step. Another potential application is monitoring the "biological" and "chronological" age of a pregnancy. This application can be used to detect or assess the risk of preterm birth. Other embodiments can be used to analyze fetal cells in maternal blood. In still other embodiments, such fetal cells can be identified (e.g., on the cell surface or within the cytoplasm) by antibody-based approaches or by selective staining using cell markers, or enriched by flow cytometry, micromanipulation, microdissection, or physical methods (e.g., differential flow through a chamber, surface, or container).
[0269] C. Methylation detection using different reagents In this section, we demonstrate that the methylation technique is not limited to any particular reagent system.
[0270] Methylation analysis was performed using different reagent systems to confirm the applicability of the technology. As an example, SMRT-seq was performed using the Sequel II system (Pacific Biosciences) to perform single-molecule real-time sequencing. Sheared DNA molecules were subjected to the construction of single-molecule real-time (SMRT) sequencing templates using the SMRTbell Express Template Prep Kit 2.0 (Pacific Biosciences). Sequencing primer annealing and polymerase binding conditions were calculated using SMRT Link v8.0 software (Pacific Biosciences). Briefly, sequencing primer v2 was annealed to the sequencing template, and then polymerase was bound to the template using the Sequel II Binding and Internal Control Kit 2.0 (Pacific Biosciences). Sequencing was performed using a Sequel II SMRT Cell 8M. Sequencing videos were collected for 30 hours on the Sequel II system using the Sequel II Sequencing Kit 2.0 (Pacific Biosciences). In other embodiments, other chemical reagents and reaction buffers will be used for SMRT-seq. In one embodiment, polymerases will have different kinetic characteristics for incorporating nucleotides along the DNA template strand depending on their methylation status (Huber et al. Nucleic Acids Res. 2016;44:9881-9890). In this disclosure, results are generated using sequencing primer v1 unless otherwise specified.
[0271] To demonstrate the use of the present invention in the present disclosure using different reagents, we analyzed SMRT-seq data generated based on different sequencing kits, including, but not limited to, Sequel I Sequencing Kit 3.0, RS II, Sequel II Sequencing Kit 1.0, and Sequel II Sequencing Kit 2.0. RS II contains 150,000 ZMW per SMRT cell. Sequel uses 1,000,000 ZMW per SMRT cell. Sequel II uses 8 million ZMW per SMRT cell using two sequencing kits (1.0 and 2.0). This analysis included two datasets. The first dataset was prepared based on DNA after whole genome amplification and represents the unmethylated state. The second type of dataset was prepared based on DNA after M.SsssI methyltransferase treatment and represents the methylated state. These data were generated using Sequel Sequencing Kit 3.0 for the Sequel sequencer, and Sequel II Sequencing Kit 1.0 and Sequel II Sequencing Kit 2.0 for the Sequel II sequencer. Thus, the inventors obtained three datasets with kinetic profiles generated with different reagents (e.g., polymerases). Each dataset was divided into a training dataset and a test dataset to evaluate performance using the CNN model according to the present disclosure.
[0272] 1. Measurement window Figures 84A, 84B, and 84C show the performance of different measurement window sizes across different reagent kits for SMRT-seq on a training dataset containing whole-genome amplification data (unmethylated CpG sites) and M.SsssI-treated data (methylated CpG sites). The true positive rate is plotted on the y-axis, and the false positive rate is plotted on the x-axis. Figure 84A shows SMRT-seq data generated using Sequel Sequencing Kit 3.0. Figure 84B shows SMRT-seq data generated using Sequel II Sequencing Kit 1.0. Figure 84C shows SMRT-seq data generated using Sequel II Sequencing Kit 2.0. In the figures, upstream signals of the analyzed CpG cytosine sites are indicated by "-". Downstream signals of the analyzed CpG cytosine sites are indicated by "+". For example, "-6nt" represents a signal 6 nt upstream of the analyzed CpG cytosine site. "+6nt" indicates a 6-nt downstream signal from the analyzed CpG cytosine site. "±6nt" indicates both a 6-nt upstream signal and a 6-nt downstream signal from the analyzed CpG cytosine site (i.e., a total of 12 nt of sequence adjacent to the CpG cytosine site).
[0273] As shown in Figure 84A, for a training dataset based on Sequel Sequencing Kit 3.0, using a measurement window including the signal of the analyzed CpG cytosine and the signal 6 nt upstream of that cytosine (denoted by -6 nt) (e.g., IPD, PW, relative position, sequence composition), an AUC value of 0.50 suggested no discriminatory power in distinguishing methylated CpG cytosines from unmethylated cytosines. However, for training datasets based on Sequel II Sequencing Kits 1.0 and 2.0, the corresponding AUC values were 0.62 (Figure 84B) and 0.75 (Figure 84C). These data indicate that there are distinct kinetic profiles inherent to the different reagents used in SMRT-seq. These data demonstrate that the methods disclosed herein are easily adapted to the use of different reagents. Furthermore, the accuracy of detecting base modifications can potentially be improved by further development of the reagents, for example, the use of different polymerases and other chemistries.
[0274] As another example, as shown in Figure 84A, for a training dataset based on Sequel Sequencing Kit 3.0, using a measurement window containing a signal 10 bp upstream of the CpG cytosine site (labeled -10nt), an AUC value of 0.50 suggested no discriminatory power for distinguishing methylated CpG cytosines from unmethylated cytosines. However, for training datasets based on Sequel II Sequencing Kits 1.0 and 2.0, the corresponding AUC values were 0.66 (Figure 84B) and 0.79 (Figure 84C), indicating an improvement compared to a measurement window containing a 6-nt upstream signal. These data confirmed that there are distinct kinetic profiles inherent to the different reagents used for SMRT-seq. These data indicate that the methods disclosed herein are easily adapted to the use of different reagents.
[0275] In contrast to measurement windows with upstream signals, measurement windows with downstream signals can lead to significant improvements in classification performance. For example, as shown in Figure 84A, for a training dataset based on Sequel Sequencing Kit 3.0 using a measurement window including 6 nt downstream signals (+6 nt) of a CpG cytosine site, the AUC value was 0.94, much larger than when using 6 nt upstream signals (AUC: 0.5). For training datasets based on Sequel II Sequencing Kits 1.0 and 2.0, the corresponding AUC values were 0.95 (Figure 84B) and 0.92 (Figure 84C), respectively, demonstrating improvement compared to a measurement window including 6 nt upstream. These data suggest that kinetic features linked to sequence context can improve classification power using, but are not limited to, CNN models. These data also suggest that the disclosure herein is applicable to datasets generated by different reagents and different sequencing conditions (e.g., different polymerases, other chemical reagents, their concentrations, and sequencing reaction parameters (e.g., duration)) through adjusting the measurement window. Similar conclusions are drawn from analyses using measurement windows encompassing the signal 10 nt downstream of the CpG cytosine site (Figures 84A, 84B, and 84C).
[0276] In another embodiment, a measurement window can be used that includes the signal on the cytosine being analyzed and both the upstream and downstream signals of that cytosine. For example, as shown in Figures 84A, 84B, and 84C, using a measurement window that includes the 6-nt upstream signal and the 6-nt downstream signal (denoted ±6 nt), AUC values were found to be 0.94, 0.95, and 0.92 for training datasets based on Sequel Sequencing Kit 3.0, Sequel II Sequencing Kit 1.0, and 2.0, respectively. Using a measurement window that includes the 10-nt upstream signal and the 10-nt downstream signal (denoted ±10 nt), AUC values were found to be 0.94, 0.95, and 0.94 for training datasets based on Sequel Sequencing Kit 3.0, Sequel II Sequencing Kit 1.0, and 2.0, respectively. These data suggest that the disclosure herein is broadly applicable to datasets generated with different reagents and different sequencing reaction parameters.
[0277] When applying the CNN model trained on the training dataset, the results obtained from the test datasets using different measurement windows across different sequencing kits are shown in Figures 85A, 85B, and 85C. The true positive rate is plotted on the y-axis, and the false positive rate is plotted on the x-axis. The legend labeling is equivalent to the labeling used in Figures 84A, 84B, and 84C. Figure 85A shows SMRT-seq data generated based on Sequel Sequencing Kit 3.0. Figure 85B shows SMRT-seq data generated based on Sequel II sequencing Kit 1.0. Figure 85C shows SMRT-seq data generated based on Sequel II Sequencing Kit 2.0. All conclusions drawn from the training datasets can be verified with these independent test datasets that were not involved in the training process. Furthermore, among three independent test datasets, analysis of two datasets (2 / 3) involving Sequel II Sequencing Kit 1.0 and 2.0 showed that using a measurement window including 10 nt upstream and 10 nt downstream signals (denoted ±10 nt) was superior to the other datasets.
[0278] 2. Comparison with bisulfite sequencing Figures 86A, 86B, and 86C show the correlation between global methylation levels quantified by bisulfite sequencing and SMRT-seq (Sequel II Sequencing Kit 2.0). In Figure 86A, the methylation level as a percentage quantified by SMRT-seq is shown on the y-axis. In Figure 86B, the methylation level as a percentage quantified by bisulfite sequencing is shown on the x-axis. The black line is the fitted regression line. The dashed line is the diagonal line where the two scales are equal. Figure 86B shows a Bland-Altman plot. The x-axis shows the average methylation levels quantified by SMRT-seq and bisulfite sequencing according to the present disclosure. The y-axis shows the difference in methylation levels between SMRT-seq and bisulfite sequencing (i.e., Pacific Biosciences methylation-bisulfite-based methylation) according to the present disclosure. The dashed line corresponds to a horizontal line crossing zero where there is no difference between the two scales. Data points outside the dashed line indicate deviations between the scales. Figure 86C shows the percentage change relative to the value quantified by bisulfite sequencing. The x-axis shows the average methylation levels quantified by SMRT-seq and bisulfite sequencing according to the present disclosure. The y-axis shows the percentage difference in methylation levels between the two scales relative to the average methylation level. The dashed line corresponds to a horizontal line crossing zero where there is no difference between the two scales. Data points outside the dashed line indicate deviations between the scales.
[0279] With reference to Figure 86A, the linear regression equation is Y = aX + b, where "Y" represents the methylation level determined by SMRT-seq according to the present disclosure, "X" represents the methylation level determined by bisulfite sequencing, "a" represents the slope of the regression line (e.g., a = 1.45), and "b" represents the y-axis intercept (e.g., b = -20.98). In this case, the methylation value determined by SMRT-seq is calculated by (Yb) / a. This graph shows that methylation levels determined by SMRT-seq can be converted to methylation levels determined by bisulfite sequencing, and vice versa, for Sequel II Sequencing Kit 2.0 as well as Sequel II Sequencing Kit 1.0.
[0280] Figure 86B is a Bland-Altman plot showing the bias in methylation quantification between SMRT-seq and bisulfite sequencing according to the present disclosure. The x-axis shows the average methylation levels quantified by SMRT-seq and bisulfite sequencing according to the present disclosure, and the y-axis shows the difference in methylation levels quantified by SMRT-seq and bisulfite sequencing according to the present disclosure. The median difference between the two measurements was -6.85% (range: -10.1 to 1.7%). The median percentage change in methylation levels quantified by the present disclosure relative to the values by bisulfite sequencing was -9.96% (range: -14.76 to 3.21%). This difference varies depending on the mean. The higher the mean of the two measures, the greater the bias.
[0281] Figure 86C shows the same data as Figure 86B, but the difference in methylation levels is divided by the average of the two methylation levels. Figure 86C also shows that the greater the average of the two measurements, the greater the bias.
[0282] The error may be due to bisulfite sequencing and is unrelated to the SMRT-seq method. Conventional whole-genome bisulfite sequencing (Illumina) has been reported to have significant variability in quantification of methylation levels between methods in certain genomic regions, resulting in significantly biased sequence output and overestimated overall methylation (Olova et al. Genome Biol. 2018;19:33). The embodiments disclosed herein have several exemplary advantages, including the ability to perform without bisulfite conversion, which dramatically degrades DNA, and the ability to perform without PCR amplification.
[0283] 3.Tissue origin According to an embodiment of the present disclosure, single molecule real-time sequencing (SMRT-seq, Pacific Biosciences) was used to perform methylation analysis across various cancer types. The cancer types used for SMRT-seq included, but were not limited to, colorectal cancer (n=3), esophageal cancer (n=2), breast cancer (n=2), renal cell carcinoma (n=2), lung cancer (n=2), ovarian cancer (n=2), prostate cancer (n=2), gastric cancer (n=2), and pancreatic cancer (n=1). Matched adjacent non-tumor tissues were also included in the SMRT-seq. The dataset was generated from DNA prepared using the Sequel II Sequencing Kit 2.0.
[0284] Figures 87A and 87B show a comparison of global methylation levels between various tumor tissues and paired adjacent non-tumor tissues. Methylation levels are on the y-axis as a percentage. In Figure 87A, methylation levels are quantified by SMRT-seq. In Figure 87B, methylation levels are quantified by bisulfite sequencing. Tissue type (i.e., tumor tissue or adjacent non-tumor tissue) is on the x-axis. Different symbols represent different tissues of origin.
[0285] Figure 87A shows that the global methylation levels of tumor tissues, including breast, colon, esophageal, liver, lung, ovarian, pancreatic, renal cell carcinoma, and gastric cancer, were significantly lower than those of corresponding non-tumor tissues (including breast, colon, esophagus, liver, lung, ovary, pancreas, prostate, kidney, and stomach, respectively) (P value = 0.006, paired-sample Wilcoxon signed-rank test). The median difference in methylation levels between tumor and paired non-tumor tissues was -2.7% (IQR: -6.4 to -0.8%).
[0286] Figure 84B confirms lower methylation levels in tumor tissue. These results therefore suggest that methylation patterns across various cancer types and tissues can be accurately determined by SMRT-seq according to the present disclosure, implicating broad applications of the present disclosure for early detection, prognosis, diagnosis, and treatment of cancer based on tissue biopsy. The differential extent of reduced methylation levels across various tumor types may suggest that methylation patterns are related to cancer type, allowing for the determination of the tissue of origin of the cancer.
[0287] D. Enhanced Detection and Other Techniques In some embodiments, analysis of base modifications (e.g., methylation) can be performed using one or more of the following parameters: sequence context, IPD, and PW. IPD and PW can be determined from sequencing reactions without aligning to a reference genome. Aspects of single-molecule real-time sequencing approaches can further enhance the accuracy of determining sequence context, IPD, and PW. One aspect is the ability to generate circular consensus sequences that can measure specific positions on a sequence template multiple times, allowing for determination of sequence context, IPD, and PW based on the average or distribution of values from these multiple reads. In certain embodiments, analysis of base modifications without an alignment process can increase computational efficiency, reduce turnaround time, and reduce the cost of analysis. Embodiments can be performed without an alignment process. In still other embodiments, an alignment process can be used, and may be preferred, for example, when using an alignment process to confirm the clinical or biological significance of detected base modifications (e.g., when a tumor suppressor is hypermethylated) or when using an alignment process to select a subset of sequencing data corresponding to a particular genomic region of interest for further analysis. In embodiments where data from selected genomic regions is desired, these embodiments may involve targeting such regions using one or more enzymes or enzyme-based methodologies that can cut regions of interest within the genome, for example, restriction enzymes or CRISPR-Cas9 systems. PCR amplification typically does not preserve information about DNA base modifications, so CRISPR-Cas9 systems may be preferable to PCR-based methods. The methylation level of such selected regions (bioinformatically, for example, through alignment or through methods such as CRISPR-Cas9) can be analyzed to provide information about tissue origin, fetal disorders, pregnancy disorders, and cancer.
[0288] 1. Methylation analysis using subreads without aligning to a reference genome In embodiments, methylation analysis can be performed using a measurement window containing kinetic features and sequence context from subreads without aligning to a reference genome. As shown in Figure 88, a consensus sequence 8802 (also known as a circular consensus sequence (CCS)) was constructed using subreads derived from zero-mode waveguiding (ZMW). Average kinetic values at each position of the CCS, including but not limited to PW and IPD values, were calculated. The sequence context surrounding a CpG site was determined from the CCS based on the upstream and downstream sequences of the CpG site. Thus, the measurement window defined in the present disclosure is constructed for training, and the measurement window includes PW, IPD value, and sequence context according to the subreads with kinetic features associated with the CCS. This procedure eliminates the need to align the subreads to a reference genome.
[0289] To test the principle shown in Figure 88, a training dataset was created using 601,942 unmethylated CpG sites from whole genome amplified DNA and 163,527 methylated CpG sites from CpG methyltransferase (e.g., M.SssI)-treated DNA. A test dataset was created using 546,393 unmethylated CpG sites from whole genome amplified DNA and 193,641 methylated CpG sites from CpG methyltransferase (e.g., M.SssI)-treated DNA. The dataset was generated from DNA prepared by Sequel II Sequencing Kit 2.0.
[0290] As shown in Figure 89, in one embodiment, kinetic features and sequence context associated with subreads and CCSs are used to train a convolutional neural network (CNN) model for determining methylation, which can achieve AUC values of 0.94 and 0.95 for distinguishing methylated from unmethylated CpG sites in the test and training datasets, respectively. In other embodiments, other neural network models, deep learning algorithms, artificial intelligence, and / or machine learning algorithms can be used.
[0291] Setting the methylation probability cutoff at 0.2, we obtained a sensitivity of 82.4% and a specificity of 91.7% for detecting methylated CpG sites. These results demonstrate that subreads with kinetic features can be used to distinguish between methylated and unmethylated CpG sites without prior alignment to a reference genome.
[0292] In another embodiment, kinetic features can be used directly from subreads along with sequence context to determine the methylation status of entire CpG sites without CCS information and without prior alignment to a reference genome. To train a CNN model for determining methylation status, kinetic features including PW and IPD values at positions spanning 20 nt upstream and 20 nt downstream of the CpG present in the subread were used. As shown in Figure 90, the AUCs of the ROC curves using subread-associated kinetic features according to embodiments of the present disclosure were 0.70 and 0.69, respectively, for detecting methylated CpG sites in the training and test datasets. These data suggested that it is feasible to use embodiments of the present disclosure to infer the methylation pattern of DNA molecules using subread-associated kinetic features but without prior alignment and construction of a consensus sequence. However, the performance of this embodiment for determining methylation was inferior to embodiments that utilize alignment information or consensus sequences in combination, as described herein. It is believed that the increased accuracy in generating subreads and kinetic values will improve the performance of determining base modifications using subreads and their associated kinetic features.
[0293] 2. Methylation analysis of the deleted region using targeted single-molecule real-time sequencing The methods described herein can also be applied to analyze one or more selected genomic regions. In one embodiment, the region(s) of interest can first be enriched by a hybridization method, which allows DNA molecules from the region(s) of interest to hybridize with synthetic oligonucleotides having complementary sequences. When analyzing base modifications using the methods described herein, the target DNA molecules cannot be amplified by PCR before being subjected to sequencing, because the base modification information of the original DNA molecules is not transferred to the PCR product. Several methods have been developed to enrich these target regions without PCR amplification.
[0294] In another embodiment, the target region(s) can be enriched through the use of the CRISPR-Cas9 system (Stevens et al. PLOS One 2019;14(4):e0215441, Watson et al. Lab Invest 2020;100:135-146). In one embodiment, the ends of DNA molecules in a DNA sample are first dephosphorylated to prevent them from being directly ligated to sequencing adapters. The region(s) of interest are then guided by the Cas9 protein with a guide RNA (crRNA) to create a double-stranded break. The region(s) of interest on both sides of the double-stranded break are then ligated to sequencing adapters specified by the sequencing platform of choice. In another embodiment, the DNA can be treated with an exonuclease to degrade DNA molecules not bound to the Cas9 protein (Stevens et al. PLOS One 2019;14(4):e0215441). These methods do not involve PCR amplification, and therefore can determine the sequence of the original DNA molecule containing the base modification, and determine the base modification. In one embodiment, this method can be used to target multiple regions that share homologous sequences, such as long interspersed nucleotide sequences (LINEs). In one example, such analysis can be used to analyze circulating cell-free DNA in maternal plasma to detect fetal aneuploidy (Kinde et al. PLOS One 2012;7(7):e41162).
[0295] As shown in Figure 91, the CRISPR (clustered regularly interspaced short palindromic repeats) / Cas9 (CRISPR-associated protein 9) system can be used to implement targeted single-molecule real-time sequencing. A DNA fragment (e.g., molecule 9102) bearing a 5' phosphoryl group (i.e., 5'-P) and a 3' hydroxyl group (i.e., 3'-OH) was subjected to an end-blocking process by removing the 5'-P and ligating the 3'-OH with a dideoxynucleotide (i.e., ddNTP). The resulting end-modified molecule (e.g., molecule 9104) could not be ligated with an adapter for subsequent DNA library preparation. However, the end-blocked molecule was subjected to targeted cleavage mediated by the CRISPR / Cas9 system, introducing 5'-P and 3'-OH ends into the molecule of interest. Such a newly cleaved DNA molecule (e.g., molecule 9106) bearing a 5'-P and 3'-OH end can now be ligated with a hairpin adapter to form a circular molecule 9108. The unligated adapter, linear DNA, and molecules with only one cleavage were subjected to digestion with exonucleases III and VII. As a result, molecules ligated with two hairpin adapters were enriched and subjected to single-molecule real-time sequencing. These target molecules were suitable for base modification analysis (i.e., targeted single-molecule real-time sequencing) according to embodiments of the present disclosure.
[0296] As shown in Figure 92, the Cas9 protein of the CRISPR / Cas9 system interacted with a guide RNA (i.e., gRNA) comprising a CRISPR RNA (crRNA, involved in DNA targeting) and a transactivating crRNA (tracrRNA, involved in forming a complex with Cas9) (Pickar-Oliver et al. Nat Rev Mol Cell biol. 2019;20:490-507). The curved shape represents the Cas9 protein, an enzyme that recognizes and cleaves specific strands of DNA complementary to portions of the CRISPR sequence using the CRISPR sequence as a guide. The crRNA was annealed to the tracrRNA. In one embodiment, a synthetic single RNA sequence contained both the crRNA and tracrRNA sequences, referred to as the single guide RNA (sgRNA). A segment of the crRNA, referred to as the spacer sequence, guides the Cas9 protein to recognize and cleave the specific strand of double-stranded DNA (dsDNA) through complementary base pairing to the target region. In one embodiment, there is no mismatch involved in the complementarity between the spacer sequence and the target dsDNA.In another embodiment, the complementary base pairing between the spacer sequence and the target dsDNA will allow mismatch.For example, the number of mismatches is not limited to, but is 1, 2, 3, 4, 5, 6, 7, 8, etc.In one embodiment, the CRISPR sequence is programmable according to the cutting efficiency, specificity, sensitivity, and the ability to multiplex different CRISPR / Cas complex designs.
[0297] As shown in Figure 93, we designed a pair of CRISPR / Cas9 complexes to target two cleavages spanning an Alu element in the human genome. "XXX" indicates the three nucleotides adjacent to the Cas9 nuclease cleavage site. "YYY" indicates the three corresponding nucleotides complementary to "XXX." 5'-NGG represents the protospacer adjacent motif (PAM) sequence. In other CRISPR / Cas systems, the PAM sequence may differ, resulting in different sequences adjacent to the Cas nuclease cleavage site. In this figure, the size of the Alu region was 223 bp. There were 1,175,329 Alu regions, each containing a homolog of such an Alu element in the human genome. A median of five CpG sites was located in this Alu element (range: 0-34). As an example, this design included a 36-nt crRNA containing a 20-nt spacer sequence. Detailed gRNA sequence information is as follows:
[0298] First CRISPR / Cas9 complex to introduce the initial cut: (all sequences from 5' to 3') crRNA:GCCUGUAAUCCCAGCACUUUGUUUUAGAGCUAUGCU tracrRNA:AGCAUAGCAAGUUAAAAUAAGGCUAGUCCGUUAUCAACUUGAAAAAGUGGCACCGAGUCGGUGCUUU
[0299] A second CRISPR / Cas9 complex to introduce the second cut: crRNA:AGGGUCUCGCUCUGUCGCCCGUUUUAGAGCUAUGCU tracrRNA:AGCAUAGCAAGUUAAAAUAAGGCUAGUCCGUUAUCAACUUGAAAAAGUGGCACCGAGUCGGUGCUUU
[0300] The crRNA molecule was annealed to the tracrRNA (e.g., 67 nt) to form the backbone of the gRNA. The Cas9 nuclease containing the designed gRNA can cleave both strands of the end-blocked molecule containing the target cleavage site with a certain level of specificity. There were 116,184 Alu regions of interest in the human genome that were predicted to be cleaved by the designed CRISPR / Cas9 complex. Therefore, after targeted cleavage by the Cas9 complex, these Alu regions can be ligated to hairpin adapters. These molecules ligated to hairpin adapters can be sequenced by single-molecule real-time sequencing. The methylation patterns of these Alu regions can be determined in a targeted manner. In one embodiment, the spacer sequences from the two Cas9 complexes can base-pair with the same strand (e.g., Watson strand or Crick strand) of a double-stranded DNA substrate. In one embodiment, the spacer sequences of the gRNAs from the two Cas9 complexes can base-pair with different strands of a double-stranded DNA substrate. For example, the spacer sequence of one of the Cas9 complexes was complementary to the Watson strand of the double-stranded DNA substrate, and the spacer sequence of the other of the Cas9 complexes was complementary to the Crick strand of the double-stranded DNA substrate, and vice versa.
[0301] In one embodiment, the DNA molecules ligated to hairpin adapters are in a circular form that is resistant to exonuclease digestion. Therefore, the DNA products ligated to adapters can be treated with exonucleases (e.g., exonucleases III and VII) to remove linear DNA (e.g., off-target DNA molecules). This step using exonucleases can further enrich the target molecules. The size of the target molecule to be sequenced depended on the span size between the two cleavage sites introduced by one or more Cas9 nucleases (e.g., including, but not limited to, 10 bp, 20 bp, 30 bp, 40 bp, 50 bp, 100 bp, 200 bp, 300 bp, 400 bp, 500 bp, 1000 bp, 2000 bp, 3000 bp, 4000 bp, 5000 bp, 10 kb, 20 kb, 30 kb, 40 kb, 50 kb, 100 kb, 200 kb, 300 kb, 500 kb, and 1 Mb).
[0302] As an example, using Cas9 containing a gRNA targeting the Alu region, we sequenced 187,010 molecules from a human hepatocellular carcinoma (HCC) tumor tissue sample using single-molecule real-time sequencing. Among them, 113,491 molecules had on-target cleavage (i.e., the on-target cleavage rate was approximately 60.7% of the molecules). The dataset was generated from DNA prepared using the Sequel II Sequencing Kit 2.0. In other words, in this example, the specificity of the cleavage site introduced into the molecule of interest by the Cas9 complex was 60.7%. In other embodiments, the specificity of the cleavage site introduced into the molecule of interest by Cas9 or other Cas complexes may vary, including, but not limited to, 1%, 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and 100%. To determine the methylation status at CpG sites in Alu sequences, we used IPD values, PW values, and sequence context derived from CCS and subreads without aligning to the reference genome.
[0303] As shown in Figure 94, a similar distribution of methylation was observed between the methylation levels determined by bisulfite sequencing and single-molecule real-time sequencing according to the present disclosure. Figure 94 shows a histogram of methylation density (percentage) from bisulfite sequencing and single-molecule real-time sequencing (Pacific Biosciences). The y-axis indicates the percentage of molecules in a sample with a specific methylation density indicated on the x-axis. This result suggested that it is feasible to determine methylation patterns using Cas9-mediated targeted single-molecule real-time sequencing. This result also suggested that methylation can be determined using subread-related kinetic features, including PW and IPD values, without aligning to a reference genome. As shown in Figure 94, a significant amount of Alu regions exhibiting hypomethylation was observed, which was consistent with previous findings that cancer genomes are demethylated at Alu repeat regions (Rodriguez et al. Nucleic Acids Res. 2008;36:770-784).
[0304] Figure 95 shows the distribution of methylation levels determined by single-molecule real-time sequencing according to the present disclosure on the y-axis, and the methylation density determined by bisulfite sequencing on the x-axis. As shown in Figure 95, the methylation levels of Alu regions were classified into five categories according to the results of bisulfite sequencing: 0-20%, 20-40%, 40-60%, 60-80%, and 80-100%. The methylation levels of the same set of Alu regions were further determined by a model using a measurement window containing the kinetic features and sequence context (y-axis) of each category of Alu regions. The distribution of methylation levels determined by our model gradually increased according to the ascending order of methylation levels across the binned categories. Again, these results suggest that it is possible to determine methylation patterns using Cas9-mediated targeted single-molecule real-time sequencing. Methylation can be determined using subread-related kinetic features, including PW and IPD values, without aligning to a reference genome.
[0305] In yet another embodiment, other types of CRISPR / Cas systems can be used to perform targeted single molecule real-time sequencing, including but not limited to, Cas12a, Cas3, and other orthologues (e.g., Staphylococcus aureus Cas9) or engineered Cas proteins (enhanced Acidaminococcus spp. Cas12a).
[0306] In one embodiment, non-activated Cas9 (dCas9) without nuclease activity can be used to concentrate target molecules without cleavage. For example, target DNA molecules are bound to a complex containing biotinylated dCas9 and target sequence-specific gRNA. Because dCas9 lacks nuclease activity, such target DNA molecules may not be cleaved by dCas9. Target DNA molecules can be concentrated using streptavidin-coated magnetic beads.
[0307] In one embodiment, after incubation with Cas protein, exonuclease can be used to digest the DNA mixture. Exonuclease can degrade DNA molecules not bound to Cas protein, while exonuclease can not degrade or degrade DNA molecules bound to Cas protein with significantly reduced efficiency. Therefore, information about target molecules bound to Cas protein can be further enriched in the final sequencing results.
[0308] Figure 96 shows a table of tissues and the methylation levels of Alu regions within the tissues. Most tissues exhibit methylation levels ranging from 85% to 92%, including a range of 88% to 92%. HCC tumor tissues and placenta tissues exhibited methylation levels below 80%. As seen in Figure 96, HCC tumors were frequently hypomethylated in the Alu regions targeted by our design. Therefore, the Alu methylation determinations presented in this disclosure can be used for cancer detection, staging, and monitoring during tumor progression or tumor treatment using DNA extracted from tumor biopsies or other tissues or cells.
[0309] Hypomethylation of placental tissue across the Alu region can be used to perform noninvasive prenatal testing using plasma DNA from pregnant women. For example, a high degree of hypomethylation may indicate a high fetal DNA fraction in pregnant women. In another example, if a woman is carrying a chromosomal aneuploid fetus, the number of Alu fragments derived from the affected chromosome detected by this approach may be quantitatively different (i.e., either increased or decreased) from women carrying euploid fetuses. Therefore, if a fetus has trisomy 21, the number of Alu fragments derived from chromosome 21 detected by this approach may be increased compared to women carrying euploid fetuses. On the other hand, if a fetus has monosomy, the number of Alu fragments derived from that chromosome detected by this approach may be decreased compared to women carrying euploid fetuses. The presence of extra hypomethylation of affected chromosomes (13, 18, or 21) in plasma compared to unaffected chromosomes can be used as a molecular indicator to distinguish between women carrying normal and abnormal fetuses.
[0310] 3. Methylation analysis of Alu regions targeted by Cas9 complexes in different cancer types Although the targeted Alu repeats were highly methylated in different tissues, the inventors hypothesized that different types of cancers would have different demethylation patterns across their Alu repeats. In one embodiment, Cas9-based targeted single-molecule real-time sequencing can be used to analyze methylation patterns and determine different cancer types according to the disclosures herein.
[0311] Figure 97 shows a cluster analysis of methylation signals associated with Alu repeats in different cancer types. Cancer subjects from the TCGA database (www.cancer.gov / about-nci / organization / ccg / research / structural-genomics / tcga) had their methylation status analyzed at CpG sites using microarray technology (Infinium HumanMethylation450 BeadChip, Illumina Inc). The methylation status was analyzed across 3,024 CpG sites present on the microarray chip that overlap with the Alu regions targeted by the CRISPR / Cas9 complex. Several CpGs originate from the Alu regions of interest in each patient. The methylation level of each CpG was quantified by microarray (also known as a methylation index or beta value). Hierarchical clustering was performed based on the number of methylation levels at those CpG sites across patients. Therefore, patients with similar methylation level patterns at those CpG sites were clustered together to form clades. The similarity of methylation patterns across different patients is indicated by the height value of the clustering dendrogram. In this example, the height was calculated according to Euclidean distance. In other embodiments, other distance metrics are used, including but not limited to Minkowski, Chebyshev, Mahalanobis, Manhattan, cosine, correlation, Spearman, Hamming, Jaccard distance, etc. As used herein, height represents the value of the distance metric between clusters and reflects the relatedness between clusters. For example, if two clusters are observed to overlap at a height x, it is suggested that the distance between those clusters is x (e.g., the average distance between all inter-cluster patients).
[0312] Using the methylation status of CpG sites, cluster analysis showed that patients were clustered into distinct groups according to cancer type. Cancer types included bladder urothelial carcinoma (BLCA), invasive breast carcinosarcoma (BRCA), ovarian serous cystadenocarcinoma (OV), pancreatic adenocarcinoma (PAAD), HCC, lung adenocarcinoma (LUAD), gastric adenocarcinoma (STAD), cutaneous melanoma (SKCM), and uterine carcinosarcoma (UCS). The numbers following the cancer type in the figure indicate the patient. Therefore, clustering suggests that the methylation signals of the Alu repeats we selected were useful for classifying cancer types, including those not shown in Figure 97. In one embodiment, primary and secondary tumors can be distinguished based on methylation patterns in tissue biopsies.
[0313] 4. Sublead depth and size cutoff This section demonstrates that subread depth and / or size cutoffs can be used to improve the accuracy and / or efficiency of methylation detection. Library preparation may be modified to test specific subread depths or sizes.
[0314] Based on the Sequel II Sequencing Kit 2.0, we analyzed the effect of read depth on the quantification of global methylation levels in test datasets generated from samples after whole genome amplification or M.SsssI treatment. Genomic sites covered by subreads with at least a specific cutoff were investigated, for example, but not limited to, 1x, 10x, 20x, 30x, 40x, 50x, 60x, 70x, 80x, 90x, 100x, etc.
[0315] Figure 98A shows the effect of read depth on quantitating global methylation levels in a test dataset involving whole genome amplification. Figure 98B shows the effect of read depth on quantitating global methylation levels in a test dataset involving M.SsssI treatment. The y-axis shows global methylation levels as a percentage. The fold axis shows sub-read depth. The dashed line shows the expected global methylation level.
[0316] As shown in Figure 98A, for the dataset containing whole genome amplification, global methylation ranged from 5.7% to 5.2% and decreased at the first few cutoffs, such as 1x, 10x, 20x, 40x, and 50x. Methylation levels gradually stabilized at approximately 5% at cutoffs of 50x and above.
[0317] On the other hand, in Figure 98B, for the dataset generated from the sample after M.SsssI treatment, the overall methylation increased in the range of 70% to 83% at the first few cutoffs, such as 1x, 10x, 20x, 40x, and 50x. The methylation level gradually stabilized at approximately 83% at cutoffs above 50x.
[0318] In one embodiment, the subread depth cutoff can be adjusted to make the performance of base modification analysis acceptable across different applications. In other embodiments, using a slightly more relaxed subread depth cutoff can obtain a larger ZMW (i.e., number of molecules) suitable for downstream analysis. In yet another embodiment, the methylation level readout determined by SMRT-seq according to the present disclosure can be calibrated with a second measurement (e.g., without limitation, BS-seq, digital droplet PCR (on bisulfite-converted samples), methylation-specific PCR, or methylated cytosine-binding antibodies or other proteins). In another embodiment, the second measurement is obtained by subjecting whole-genome-amplified DNA molecules maintained at 5mC to BS-seq, digital droplet PCR (on bisulfite-converted samples), methylation-specific PCR, or methyl-CpG-binding domain (MBD) protein-enriched genome sequencing (MBD-seq). As an example, 5mC-retaining whole genome amplification can be mediated by DNA primase TthPrimPol, polymerase phi29, and DNMT1 (DNA methyltransferase 1).
[0319] Methylation levels were analyzed across various types of cancer and non-tumor tissues at different subread depths. Methylation levels determined by SMRT-seq according to the present disclosure were also compared with BS-seq sequencing results. Using the Sequel II Sequencing Kit 2.0, a median of 43 million subreads (interquartile range (IQR): 30-52 million) were obtained, which allowed for the generation of a median of 4.6 million circular consensus sequences (CCSs) that were aligned to the human reference genome (IQR: 2.8-5.8 million). Of these samples, 22 were also subjected to massively parallel bisulfite sequencing (BS-seq), a well-established method for determining methylation patterns, providing a second measurement for comparing methylation levels.
[0320] Figure 99 shows a comparison between global methylation levels determined by SMRT-seq (Sequel II Sequencing Kit 2.0) according to the present disclosure and BS-seq using different subread depth cutoffs. Methylation levels as a percentage determined by SMRT-seq are shown on the y-axis. Methylation levels as a percentage determined by bisulfite sequencing are on the x-axis. Symbols indicate different subread depths of 1x, 10x, and 30x. Three diagonal lines indicate lines approximating the different subread depths.
[0321] Figure 99 shows that when analyzing genomic sites covered at least once by subreads (i.e., a subread depth cutoff of 1x or greater), the methylation levels of CpG sites determined by SMRT-seq according to the present disclosure correlate well with those determined by BS-seq (r=0.8, P-value<0.0001). These results suggest that the present embodiments can be used to measure methylation levels in different tissue types, including, but not limited to, colon cancer, colorectal tissue, esophageal cancer, breast cancer, non-cancerous breast tissue, renal cell carcinoma, kidney tissue, lung cancer, and lung tissue. We also observed that increasing the subread depth cutoff to 10x and 30x, respectively, improved the correlation between these two measurements to 0.87 (P-value<0.0001) and 0.95 (P-value<0.0001). In some embodiments, increasing the subread depth or selecting genomic regions that cover more subreads will improve the performance of SMRT-seq-based methylation determination according to the present disclosure.
[0322] Figure 100 is a table showing the effect of subread depth on the correlation of methylation levels between two measurements by SMRT-seq (Sequel II Sequencing Kit 2.0) and BS-seq. The first column shows the subread depth cutoff. The second column shows the correlation coefficient, Pearson's r. The third column shows the number of CpG sites associated with the cutoff, with the range of site numbers in parentheses.
[0323] As shown in Figure 100, the correlation of methylation levels between the two measurements by SMRT-seq and BS-seq varied with different subread depth cutoffs. In one embodiment, the relationship between the subread depth cutoff and the correlation coefficient (e.g., Pearson's correlation coefficient) between the two measurements can be utilized to determine the optimal subread depth cutoff for distinguishing methylated from unmethylated cytosines. Figure 100 shows that at a subread depth cutoff of 30x (i.e., 30x or greater), the methylation levels measured by SMRT-seq according to the present disclosure showed the highest correlation with the results generated by BS-seq (Pearson's r=0.952). In other embodiments, sub-read depth cutoffs such as, but not limited to, 1x, 10x, 30x, 40x, 50x, 60x, 70x, 80x, 900x, 100x, 200x, 300x, 400x, 500x, 600x, 700x, 800x, etc., can be used.
[0324] The number of CpG sites used for methylation analysis decreases with increasing subread depth cutoff, as shown in Figure 100. At a subread depth cutoff of 100x, a lower correlation (Pearson's r = 0.875) was observed between the two measures of methylation levels compared to a subread depth cutoff of 30x (Pearson's r = 0.952). The lower correlation at higher subread cutoffs may be due to the smaller number of CpG sites meeting the more stringent subread depth cutoff. In one embodiment, a trade-off can be considered between the subread depth requirement and the number of molecules that can be used for methylation analysis. For example, if the goal is to scan the entire genome for methylation patterns, more molecules may be desirable. If targeted SMRT-seq is used to focus on a specific region, a higher subread depth may be desirable to obtain the methylation pattern of that region.
[0325] Figure 101 shows the subread depth distribution with respect to fragment size for data generated by the Sequel II Sequencing Kit 2.0. The y-axis shows the subread depth, and the x-axis shows the DNA molecule length, which was estimated from the size of the circulating consensus sequence (CCS).
[0326] Subread depth can affect the performance of methylation determination using SMRT-seq data. Because subread depth is a function of the strand length of the DNA molecules being sequenced, the size of the DNA molecules can be important for obtaining an optimal subread depth for analyzing the methylation pattern of a sample. As shown in Figure 101, the longer the DNA, the shallower the subread depth. For example, for a population of molecules 1 kb in size, the median subread depth was 50x. For a population of molecules 10 kb in size, the median subread depth was 15x.
[0327] In one embodiment, as shown in Figure 100, the optimal cutoff for subread depth may be at least 30x, resulting in the highest correlation coefficient. To further improve the throughput of molecules that meet the optimal subread depth cutoff of 30x, the relationship between subread depth and the strand length of the DNA template molecule can be utilized. For example, in Figure 101, 30x is the median subread depth for molecules with a strand length of approximately 4 kb. Therefore, before preparing the SMRT-seq library, 4 kb DNA molecules can be fractionated and sequencing can be limited to 4 kb DNA molecules. In other embodiments, other size cutoffs can be used for fractionation of DNA molecules, including but not limited to 100bp, 200bp, 300bp, 400bp, 500bp, 600bp, 700bp, 800bp, 900bp, 1kb, 2kb, 3kb, 4kb, 5kb, 6kb, 7kb, 9kb, 10kb, 20kb, 30kb, 40kb, 50kb, 60kb, 70kb, 80kb, 90kb, 100kb, 500kb, 1Mb, or combinations of different size cutoffs.
[0328] 5. Restriction Enzyme-based Targeted Single-molecule Real-time Sequencing This section describes the use of restriction enzymes to improve the practicality, throughput, and / or cost-effectiveness of modification detection. The DNA fragments generated by restriction enzymes can be used to identify the origin of a sample.
[0329] a) Restriction enzymes are used to digest DNA molecules In embodiments, DNA molecules can be digested using one or more restriction enzymes prior to single-molecule real-time sequencing (e.g., using a Pacific Biosciences system). Because the distribution of restriction enzyme recognition sites is uneven across the human genome, DNA digested by restriction enzymes can produce a skewed size distribution. Genomic regions with more restriction enzyme recognition sites are digested into smaller fragments, while genomic regions with fewer restriction enzyme recognition sites may be digested into longer fragments. In embodiments, size ranges can be used to selectively obtain DNA molecules from one or more regions with similar cleavage patterns of one or more restriction enzymes. The size range required for size selection can be determined by in silico cleavage analysis of one or more restriction enzymes. A computer program can be used to determine the number of recognition sites for a restriction enzyme of interest in a reference genome (e.g., a human reference genome). Such a reference genome was sheared in silico into fragments according to their recognition sites, providing size information for the genomic regions of interest.
[0330] Figure 126 shows a method for MspI-based targeted single-molecule real-time sequencing using DNA end repair and A-tailing. In an embodiment, as shown in Figure 126, MspI, which recognizes 5'C^CGG3' sites, can be used to digest an organism's DNA sample, for example, but not limited to, a human DNA sample. The digested DNA fragments with 5'CG overhangs were subjected to size selection to enrich for DNA molecules derived from CpG islands. Genomic regions enriched in G and C residues (also known as GC content) may generate shorter fragments. Therefore, the fragment size range for selection can be determined based on the GC content of the region of interest. Various DNA fragment size selection tools are available to those skilled in the art, including, but not limited to, gel electrophoresis, size exclusion electrophoresis, capillary electrophoresis, chromatography, mass spectrometry, filtration approaches, precipitation-based approaches, microfluidics, and nanofluidics. The size-fractionated DNA molecules were subjected to DNA end repair and A-tailing, and the desired DNA products were ligated with hairpin adapters with 5′ T overhangs to form circular DNA templates.
[0331] For example, without limitation, after using exonucleases (exonucleases III and VII) to remove unligated adapters, linear DNA, and incomplete circular DNA, DNA molecules ligated to hairpin adapters can be used in single-molecule real-time sequencing to determine the IPD, PW, and sequence context in determining the methylation profiles disclosed herein. By analyzing CpG-enriched genomic regions, DNA obtained from different tissues or tissues or biological samples with different diseases and / or physiological conditions can be distinguished and classified by their methylation profiles determined by the sequencing data analysis methods disclosed herein.
[0332] In an embodiment, for the step involving size selection in Figure 126, the desired size range can be determined by in silico cleavage analysis of MspI. A total of 2,286,541 MspI cleavage sites were determined in the human reference genome. The human reference genome was sheared into fragments in silico according to the MspI cleavage sites. A total of 2,286,565 fragments were obtained. The size of each individual fragment was determined by the total number of nucleotides in the fragment.
[0333] Figures 127A and 127B show the size distribution of fragments digested with MspI. The y-axis in these figures is the frequency (percentage) of fragments of a particular size. Figure 127A has a logarithmic scale for the x-axis ranging from 50 to 500,000 bp. Figure 127B has a linear scale for the x-axis ranging from 50 to 1,000 bp.
[0334] As shown in Figures 127A and 127B, the MspI-digested DNA molecules have a skewed size distribution. The median size of the MspI-digested fragments was 404 bp (IQR: 98-1,411 bp). Approximately 53% of these MspI-digested fragments were less than 1 kb. The size profile contained a series of spike peaks that could be attributed to repetitive elements. Specific repeat elements may share similar patterns of MspI cleavage sites, leading to sets of molecules derived from MspI digestion with similar fragment sizes. For example, the most frequent spike peak (i.e., a total of 49,079) corresponded to a size of 64 bp. Among them, 45,894 (94%) overlapped with Alu repeats. DNA molecules 64 bp in size can be selected to enrich for DNA molecules derived from Alu repeats. The data suggest that size selection can be used to enrich for desired DNA molecules for downstream methylation analysis according to the present disclosure.
[0335] Figure 128 shows a table of the number of DNA molecules in a particular selected size range. The first column shows the size range in base pairs. The second column shows the percentage of molecules within the size range relative to all fragments. The third column shows the number of molecules within the size range that overlap with a CpG island. The fourth column shows the percentage of molecules within the size range that overlap with a CpG island. The fifth column shows the number of CpG sites that are sequenced. The sixth column shows the number of CpG sites that are within a CpG island. The seventh column shows the percentage of CpG sites that are subject to size selection and are within a CpG island. As shown in Figure 128, the amount of DNA molecules generated from a human genome subjected to MspI digestion varied according to the different size ranges in question. The number of DNA molecules that overlap with a CpG island varies depending on the size range.
[0336] Because the CCGG motif occurs preferentially in CpG islands, molecules below a certain size cutoff can be selected to enrich for DNA molecules derived from CpG islands. For example, in the 50-200 bp size range, the number of molecules was 526,543, accounting for 23.03% of all DNA fragments derived from the human genome subjected to MspI digestion. Of the 526,543 DNA molecules, 104,079 (19.76%) overlapped with CpG islands. In the 600-800 bp size range, the number of molecules was 133,927, accounting for 5.86% of all DNA fragments derived from the human genome subjected to MspI digestion. Of the 133,927 molecules, 3,673 (2.74%) overlapped with CpG islands. As an example, a size range of 50-200 bp can be selected to enrich for DNA fragments derived from CpG islands.
[0337] To calculate the enrichment of CpG sites overlapping CpG islands via MspI-based targeted single-molecule real-time sequencing, we performed simulations of sheared DNA by sonication. We generated 526,543 fragments from ZMWs with a mean size of 200 bp and a standard deviation of 20 bp based on a normal distribution. Only 0.88% of the DNA molecules overlapped with CpG islands. A total of 71,495 CpG sites overlapped with CpG islands. As shown in Figure 128, when MspI-digested fragments in the 50-200 bp range were selected, 19.8% of the fragments overlapped with CpG islands. Therefore, these data suggest that DNA prepared by MspI digestion may be 22.5-fold enriched in DNA fragments derived from CpG islands compared to DNA prepared by sonication. Furthermore, we analyzed the CpG sites enriched in CpG islands through MspI digestion. Selection of MspI-digested fragments in the 50-200 bp range could yield 885,041 CpG sites overlapping CpG islands, accounting for 37.5% of the total CpG sites from sequenced fragments within that size range. Compared to DNA prepared by sonication, CpG sites overlapping CpG islands were enriched 12.3-fold (i.e., 885,041 / 71,495). Based on the information shown in Figure 128, a suitable size range can be selected to contain the desired number of CpG sites and the desired enrichment factor for CpG sites within CpG islands.
[0338] Figure 129 is a graph of the percent coverage of CpG sites within CpG islands versus the size of DNA fragments after restriction enzyme digestion. The y-axis indicates the percentage of CpG sites within CpG islands covered by fragments of a given size. The x-axis indicates the upper limit of the size range of DNA fragments after restriction enzyme digestion. Figure 129 shows the percentage of CpG sites within CpG islands covered by expanding the size selection range. In Figure 129, the size range is from 50 bp to the size shown on the x-axis. In other embodiments, the lower limit of the size range can be customized, for example, but not limited to, 60 bp, 70 bp, 80 bp, 90 bp, 100 bp, 200 bp, 300 bp, 400 bp, and 500 bp. It can be seen that as the size range is expanded by increasing the upper limit, the percent coverage of CpG sites within CpG islands gradually increases, leveling off at 65%. Some CpG sites are not covered because they are within DNA fragments less than 50 bp or within fragments within very long molecules (e.g., >100,000 bp).
[0339] In some embodiments, DNA samples can be analyzed using two or more different restriction enzymes (with different restriction sites), thereby increasing the coverage of CpG sites within CpG islands. Digestion of DNA samples with different enzymes can be performed in separate reactions, with only one type of restriction enzyme present in each reaction. For example, AccII, which recognizes CG^CG sites, can be used to preferentially cleave CpG islands. In other embodiments, other restriction enzymes that contain a CG dinucleotide as part of their recognition site can be used. There are 678,669 AccII cleavage sites in the human genome. Using AccII restriction, we performed in silico cleavage of the human reference genome, obtaining a total of 678,693 fragments. We then performed in silico size selection of these fragments and calculated the percent coverage of CpG sites within CpG islands, following the method described above for MspI digestion. It can be seen that the percent coverage of CpG sites gradually increases with increasing size selection. The percent coverage plateaus at approximately 50%. CpG site coverage is further increased by combining data from two enzyme digestion experiments (i.e., MspI digestion and AccII digestion). 80% of CpG sites within CpG islands are covered by selecting DNA fragments between 50 bp and 400 bp in size. This percentage is higher than the respective values for digestion experiments using either of these enzymes alone. Coverage can be further increased by analyzing DNA samples using other restriction enzymes. If the DNA sample is divided into two aliquots, one aliquot is digested with MspI and the other with AccII. The two digested DNA samples are mixed at equimolar concentrations and sequenced using single-molecule real-time sequencing with 5 million ZMW. Based on in silico analysis, 83% of CpG sites within CpG islands (i.e.,...
Claims
1. 1. A method for detecting methylation of a nucleotide in a nucleic acid molecule, comprising: receiving input data obtained by measuring pulses of an optical signal corresponding to a window of nucleotides in a sample nucleic acid molecule, wherein the window includes at least two nucleotides, and wherein the input data has the following characteristics: For each nucleotide in the window Identification of said nucleotide; the position of the nucleotide relative to a target position within each of the windows; the width of the pulse corresponding to the nucleotide; and a pulse interval representing the time between the pulse corresponding to the nucleotide and a pulse corresponding to an adjacent nucleotide; Including values for; The presence or absence of methylation at a nucleotide at the target position within the window is determined by: For each nucleotide in the window said identification of said nucleotide; the position of the nucleotide relative to the target position; the width of the pulse corresponding to the nucleotide; and the pulse interval, To determine using Including, wherein determining whether said methylation is present at a nucleotide at said target position within said window comprises inputting input data into a model; Here, the model is receiving the input data; applying the optimized parameters of the model to the characteristics of the input data; Outputting a methylation status specifying whether the nucleotide at the target position within each window is methylated or not The method is configured to:
2. Determining whether the methylation is present at the nucleotide at the target position within the window comprises inputting the input data into a model, the model comprising: receiving additional data, the additional data being obtained by measuring pulses of the optical signal corresponding to a plurality of windows of nucleotides, each window of nucleotides corresponding to a first nucleic acid molecule of a plurality of first nucleic acid molecules, the methylation having a known first state at a nucleotide at a target position of each window of each first nucleic acid molecule, the additional data including a value of the characteristic; storing a plurality of first training samples, each including a value for the property of each first nucleic acid molecule and a first label indicating the first state of the nucleotide at the target position; and optimizing parameters of the model using the plurality of first training samples based on an output of the model that matches or does not match a corresponding label of the first label when the values of the features of the first training samples are input to the model, the output of the model identifying whether a nucleotide at a target position in each of the windows has methylation; The method of claim 1 , wherein the training is performed by
3. The method of claim 1 , wherein the model comprises a machine learning model, a principal component analysis, a convolutional neural network, or a logistic regression.
4. The method of claim 1 , wherein the values of the characteristics are arranged in a matrix structure.
5. the sample nucleic acid molecule is a sample nucleic acid molecule of a plurality of sample nucleic acid molecules; the plurality of sample nucleic acid molecules is obtained from a biological sample of a subject; and the input data includes data obtained by measuring pulses of an optical signal corresponding to each window of nucleotides in each sample nucleic acid molecule of the plurality of sample nucleic acid molecules; The method further comprising: For each sample nucleic acid of the plurality of sample nucleic acid molecules, the presence or absence of methylation at the nucleotide at each target position in each of the windows is determined by: For each nucleotide in each of the windows Identification of the nucleotide The position of the nucleotide for each of the target positions the width of the pulse corresponding to the nucleotide; and The pulse interval The method of claim 1 , further comprising determining using:
6. determining that the methylation is present at one or more nucleotides; and 6. The method of claim 5, further comprising: using said presence of said methylation of one or more nucleotides to assign a classification of a disorder in a subject, wherein said sample nucleic acid molecule is obtained from said subject.
7. The method of claim 6 , wherein the disorder comprises cancer.
8. The method of claim 6, wherein the number or sites of methylation are used to determine the classification of the disorder.
9. determining a methylation state, whether said methylation is present at one or more nucleotides; 6. The method of claim 5, further comprising using the methylation state of the one or more nucleotides to determine a clinically relevant DNA fraction, a fetal methylation profile, a maternal methylation profile, the presence of imprinted gene regions, or a tissue of origin.
10. 10. The method of claim 9, wherein: The method includes determining a tissue of origin; The method, wherein determining the tissue of origin comprises determining whether the sample nucleic acid molecule is of fetal or maternal origin.
11. Determining whether the sample nucleic acid molecule is of fetal or maternal origin includes: determining a methylation level of said sample nucleic acid molecule using said methylation state of said one or more nucleotides; and comparing the methylation level of said sample nucleic acid molecule with a reference value; The method of claim 10, comprising:
12. 6. The method of claim 5, further comprising filtering the biological sample for nucleic acid molecules having a size greater than a cutoff size to obtain a plurality of sample nucleic acid molecules.
13. 6. The method of claim 5, further comprising aligning the plurality of sample nucleic acid molecules to a plurality of genomic regions, wherein each genomic region of the plurality of genomic regions has a subread depth greater than a cutoff number.
14. The method of claim 1 , further comprising sequencing the sample nucleic acid molecule.
15. the window of nucleotides comprises a nucleotide on a first strand of the sample nucleic acid molecule and a nucleotide on a second strand of the sample nucleic acid molecule; 2. The method of claim 1, wherein the input data further comprises, for each nucleotide in the window, a value of a strand property, the strand property indicating that the nucleotide is present in either the first strand or the second strand.
16. the sample nucleic acid molecule is a circular DNA molecule, cleaving the double-stranded DNA molecule using a Cas9 complex to form a cleaved double-stranded DNA molecule; 16. The method of claim 15, wherein the cleaved double-stranded DNA molecule is formed by ligating hairpin adaptors to the ends of the cleaved double-stranded DNA molecule.
17. 2. The method of claim 1, wherein the nucleotides within the window are determined using a circular consensus sequence and are determined without aligning the sequenced nucleotides to a reference genome.
18. The method of claim 1, wherein each nucleotide within the window is enriched or filtered.
19. Each nucleotide in the window is or by cleaving the double stranded DNA molecule using a Cas9 complex to form a cleaved double stranded DNA molecule and enriching by ligating a hairpin adaptor to the end of the cleaved double stranded DNA molecule; 20. The method of claim 18, wherein the filtering is performed by selecting double-stranded DNA molecules having a size in a size range.
20. 2. The method of claim 1, wherein the nucleotides within the window are determined without using a circular consensus sequence and without aligning the sequenced nucleotides to a reference genome.
21. The method of claim 1, further comprising using the presence of said methylation to assign a classification of a disorder in the subject, assign a classification or prognosis of a tumor or malignancy in the subject, detect the tissue origin of the sample nucleic acid molecule, or identify chimeric and hybrid DNA, wherein said sample nucleic acid molecule is obtained from said subject.
22. A computer product comprising a non-transitory computer readable medium storing a number of instructions, which when executed controls a computer system to perform the method of any one of claims 1 to 21.
23. A computer product according to claim 22; and one or more processors for executing instructions stored on the computer-readable medium.
Citation Information
Patent Citations
Classification of Nucleic Acid Templates
US20100221716A1
