Molecular analysis using long cell-free DNA molecules for disease classification

Abstract

The methods and systems described herein include using various features of cell-free DNA molecules to identify characteristics of a biological sample or subject. Such features may include size (e.g., when long cell-free DNA molecules are characterized), methylation, and terminal motifs. The methods include identifying a disease classification and / or predicting a tissue of origin. In some instances, the methods include identifying an amount of long cell-free DNA molecules, and the disease classification may be based on the identified amount. The methods may also include identifying a methylation pattern of the cell-free DNA molecules and comparing the methylation pattern to a reference pattern to predict the tissue origin. In some instances, the methylation pattern analysis includes using a trained machine learning model. The features may also include a relative frequency of sequences with one or more terminal motifs, and the relative frequency in the features may be compared to a reference frequency to identify a disease classification.

Description

[Technical field]

[0001] (CROSS REFERENCE TO RELATED APPLICATIONS) This application claims priority to U.S. Provisional Patent Application No. 63 / 283,190, filed November 24, 2021, and entitled "Molecular Analysis Using Long Cell-Free DNA Molecules for Disease Classification," and U.S. Provisional Patent Application No. 63 / 285,683, filed December 3, 2021, and entitled "Molecular Analysis Using Long Cell-Free DNA Molecules for Disease Classification," the entire contents of which are incorporated by reference into this application for all purposes. [Background technology]

[0002] Many previous studies have focused on the innate / acquired information of circulating cell-free DNA molecules of 400 bp or less in the plasma of cancer patients (Jiang et al. Proc Natl Acad Sci USA. 2015;112:E1317-25; Mouliere et al. PLoS One. 2011;6:e23418; Mouliere et al. Sci Transl Med. 2018;10:eaat4921; Underhill et al. PLoS Genet. 2016;12:e1006162; Chan et al. Proc Natl Acad Sci USA. 2013;110:18761-8). The diagnostic and commercial value of long DNA molecules, e.g., ≥500 bp, ≥600 bp, ≥1 kb, ≥2 kb, ≥3 kb, ≥4 kb, ≥5 kb, ≥10 kb or other combinations, in patients with many other diseases, such as cancer or autoimmune diseases, has yet to be explored.

[0003] Jahr et al. reported the presence of high molecular weight DNA fragments in the plasma of three out of six samples with cancer by using polyacrylamide gel electrophoresis (PAGE) (Jahr et al. Cancer Res. 2001;61:1659-65). Such high molecular weight DNA molecules, present in only 50% of the cancer samples examined in this study, may suggest that such long cell-free DNA molecules may not be consistently detectable among cancer patients. Even in the three samples that showed the presence of high molecular weight cell-free DNA molecules, it is unclear how abundant these DNA molecules are. Furthermore, this study did not provide a comparison with samples from individuals without cancer. These several factors do not seem to suggest that long cell-free DNA molecules are practical biomarkers for cancer detection. Also, polyacrylamide gel electrophoresis (PAGE)-based analysis does not allow for deciphering the actual a priori / acquired information in the sequence. Common genomic analysis tools include short-read massively parallel sequencing, which is designed to analyze short DNA molecules, typically <800bp, or in fact, preferably <600bp. Moreover, the analysis of long cell-free DNA molecules has not yet been explored, as Jahr et al. and others have shown that the detection probability of long cell-free DNA molecules is low. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Jiang et al. Proc Natl Acad Sci USA. 2015;112:E1317-25 [Non-Patent Document 2] Mouliere et al. PLoS One. 2011;6:e23418 [Non-Patent Document 3] Mouliere et al. Sci Transl Med. 2018;10:eaat4921 [Non-Patent Document 4] Underhill et al. PLoS Genet. 2016;12:e1006162 [Non-Patent Document 5] Chan et al. Proc Natl Acad Sci USA. 2013;110:18761-8 Summary of the Invention [Means for solving the problem]

[0005] The techniques described herein can use various features of cell-free DNA molecules to identify characteristics of a biological sample or subject. Such features can include size (e.g., when characterizing long cell-free DNA molecules), methylation, and terminal motifs. For example, some methods, devices, and systems described herein can include analyzing biological samples using long cell-free DNA fragments.

[0006] Various methods may include determining a disease classification and / or predicting a tissue of origin based on one or more features of cell-free DNA molecules (e.g., long cell-free DNA molecules) in the subject's biological sample (e.g., plasma sample). In some instances, the method may include determining an amount of cell-free DNA molecules (e.g., within a size range with an upper limit of 1000 bases), and the disease classification may be based on the determined amount. As a further example, the method may also include identifying a methylation pattern of the cell-free DNA molecules and comparing the methylation pattern of the cell-free DNA molecules to a reference pattern to predict the tissue of origin. The origin of variants on the cell-free DNA molecules can be identified in this manner. As a still further example, the features may also include a relative frequency of sequences with one or more terminal motifs, and the relative frequencies (e.g., a vector of relative frequencies) in the features can be compared to a reference frequency to determine a disease classification.

[0007] In some instances, the methylation pattern analysis can include using a trained machine learning model. The methylation pattern analysis can provide an individual characteristic of a cell-free DNA molecule, e.g., a methylation level determined from a set of sites on the molecule, such as the percentage of sites that are methylated. Such single molecule methylation levels can be used to identify disease states.

[0008] In some embodiments, multiple features of the cell-free DNA molecules (e.g., long cell-free DNA molecules) are combined to identify disease classification and / or predict tissue of origin. For example, methylation patterns of sequence reads with variants compared to a reference sequence may be identified, and such methylation patterns can be used to identify disease classification. As another example, the relative frequency of terminal motifs may be selected from cell-free DNA molecules within a particular size range.

[0009] These and other embodiments of the present disclosure are described in more detail below. For example, other embodiments relate to systems, devices, and computer-readable media associated with the methods described herein.

[0010] A better understanding of the nature and advantages of the embodiments of the present disclosure may be obtained with reference to the following detailed description and the accompanying drawings. Other features and advantages of the present disclosure will be realized by reference to the remaining portions of the specification, including the drawings and claims. Further features and advantages of the present disclosure, as well as the structure and operation of various embodiments of the present disclosure, are described below with reference to the accompanying drawings. In the drawings, like reference numbers may indicate identical or functionally similar elements. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 shows a schematic diagram illustrating an example overview of analyzing long cell-free DNA molecules, according to some embodiments.

[0012] [Diagram 2]1 shows examples of molecules with methylated and / or unmethylated CpG sites sequenced by single molecule real-time sequencing.

[0013] [Diagram 3] FIG. 1 shows a schematic illustrating an example process for identifying kinetic characteristics of cell-free DNA molecules, according to some embodiments.

[0014] [Figure 4] FIG. 1 shows a schematic illustrating another example process for identifying kinetic characteristics of cell-free DNA molecules, according to some embodiments.

[0015] [Diagram 5] 1 shows a graph identifying the percentage of plasma DNA fragments with lengths greater than 500 bp across different sequencing techniques, according to some embodiments.

[0016] [Figure 6] A line graph illustrating the size distribution of one HCC subject and one HBV carrier is shown.

[0017] [Figure 7] 1 shows a bar graph identifying the percentage of cfDNA fragments above a given size in HCC patients with and without vascular invasion.

[0018] [Figure 8] Box plots identifying the percentage of >200 bp long DNA fragments in HCC patients with and without vascular invasion are shown.

[0019] [Figure 9] Box plots identifying the size ratio of HCC patients with and without vascular invasion are shown.

[0020] [Figure 10] FIG. 1 shows a flowchart illustrating an example process for analyzing a subject's biological sample based on the frequency of long cell-free DNA molecules, according to some embodiments.

[0021] [Figure 11] 1 shows a heatmap generated based on hierarchical clustering analysis of 256 4-mer terminal motifs of plasma DNA molecules, according to some embodiments.

[0022] [Figure 12] FIG. 1 shows a heatmap generated using hierarchical clustering analysis of 4-mer terminal motifs of short plasma DNA (<200 bp), according to some embodiments.

[0023] [Figure 13] FIG. 1 shows a heatmap generated using hierarchical clustering analysis of 4-mer terminal motifs of long plasma DNA (>1 kb), according to some embodiments.

[0024] [Figure 14] FIG. 1 shows a heatmap generated using hierarchical clustering analysis of 4-mer terminal motifs in both short plasma DNA (<200 bp) and long plasma DNA (>1 kb), according to some embodiments.

[0025] [Figure 15] 1 shows a heatmap generated using hierarchical clustering analysis of 4-mer terminal motif ratios, according to an embodiment.

[0026] [Figure 16] 1 shows a flowchart illustrating an example process for analyzing a subject's biological sample based on the relative frequency of sequences having one or more terminal motifs, according to some embodiments.

[0027] [Figure 17] A set of graphs is shown identifying motif ranking relationships between short (<600 bp) and long (>1 kb) plasma DNA molecules.

[0028] [Figure 18] Box plots identifying the terminal motif frequency of CCCA in plasma DNA molecules of <200 bp in HCC and non-HCC subjects are shown.

[0029] [Figure 19] A set of box plots identifying the motif frequency of CCCA in plasma DNA molecules is shown.

[0030] [Figure 20] 4 shows the ROC curves identifying the performance of CCCA motif frequency in short and long DNA molecules in distinguishing HCC and non-HCC subjects.

[0031] [Figure 21] Box plots identifying CCCA rates in HCC patients, HBV carriers, and healthy subjects are shown.

[0032] [Figure 22] 4 shows the ROC curve identifying the performance of CCCA rate in distinguishing subjects with HCC from those without HCC.

[0033] [Figure 23] Box plots identifying the terminal motif frequency of CCCA in plasma DNA molecules <200 bp in CRC patients and healthy subjects are shown.

[0034] [Figure 24] Box plots identifying CCCA motif frequencies in plasma DNA molecules longer than 1 kb in CRC patients and healthy subjects.

[0035] [Diagram 25] Box plots identifying CCCA rates in CRC patients and healthy subjects upon SMRT-sequencing are shown.

[0036] [Figure 26] Box plots identifying the terminal motif frequency of CCCA in plasma DNA molecules of <200 bp in HCC patients and HBV carriers are shown.

[0037] [Figure 27] A set of box plots identifying the motif frequency of CCCA in plasma DNA molecules is shown.

[0038] [Figure 28] Box plots are shown identifying the CCCA rate in HCC patients and HBV carriers during nanopore sequencing, which was calculated by dividing the CCCA motif frequency of long DNA molecules (>1 kb) by the CCCA motif frequency of short DNA molecules (<200 bp) in HCC patients and HBV carriers.

[0039] [Figure 29] 1 shows box plots identifying results generated by logistic regression analysis of terminal motif features in short DNA molecules having a size of less than 200 bp.

[0040] [Diagram 30] 4 shows ROC curves identifying the performance of logistic regression using terminal motif features on short DNA molecules (<200 bp) in distinguishing subjects with HCC from those without HCC.

[0041] [Diagram 31] 1 shows box plots identifying results generated from logistic regression analysis of terminal motif features in long DNA molecules having a size greater than 1 kb.

[0042] [Diagram 32] 4 shows ROC curves identifying the performance of logistic regression using terminal motif features on long DNA molecules (>1 kb) in distinguishing subjects with HCC from those without HCC.

[0043] [Diagram 33]Box plots identifying logistic regression analysis using terminal motif features in both long DNA molecules >1 kb and short DNA molecules <200 bp are shown.

[0044] [Diagram 34] 4 shows the ROC curve identifying the performance of logistic regression using a combination of terminal motif features derived from both long (>1 kb) and short (<200 bp) DNA molecules in distinguishing subjects with HCC from those without HCC.

[0045] [Diagram 35] Box plots identifying results generated by logistic regression analysis using motif proportions are shown.

[0046] [Diagram 36] 4 shows the ROC curve identifying the performance of logistic regression using motif proportions in distinguishing subjects with HCC from those without HCC.

[0047] [Figure 37] 4 shows the ROC curve identifying the performance of SVM using terminal motif ratio in distinguishing subjects with HCC from those without HCC.

[0048] [Figure 38] 4 shows the ROC curve identifying the performance of random forest analysis using motif proportions in distinguishing subjects with HCC from those without HCC.

[0049] [Figure 39] 4 shows the ROC curve identifying the performance of LDA analysis using motif ratios in distinguishing subjects with HCC from those without HCC.

[0050] [Diagram 40]1 shows a flowchart illustrating an example process for analyzing a subject's biological sample based on the relative frequency of sequences having one or more terminal motifs, according to some embodiments.

[0051] [Diagram 41] FIG. 1 shows an illustrative example of comparing the methylation pattern of a long cell-free DNA molecule with the methylation pattern of a reference tissue, according to some embodiments.

[0052] [Diagram 42] FIG. 1 shows a technique for analyzing methylation patterns in long cell-free DNA molecules that contain at least one methylation mismatch, according to some embodiments.

[0053] [Diagram 43] A comparison of the prevalence of CpG sites and cancer-derived single nucleotide variants (SNVs) across the genome at 1 kb resolution is shown.

[0054] [Diagram 44] A comparison of the prevalence of CpG sites and cancer-derived SNVs across the genome at 3 kb resolution is shown.

[0055] [Diagram 45] A comparison of the prevalence of CpG sites and cancer-derived SNVs across the genome at 200 bp resolution is shown.

[0056] [Figure 46] FIG. 1 shows a schematic diagram illustrating an example process for predicting whether a cell-free DNA molecule corresponds to tumor DNA depending on its methylation haplotype information.

[0057] [Figure 47] 1 shows box plots identifying the percentage of DNA molecules identified as being of hepatic origin in HCC patients at different stages based on methylation haplotype analysis according to an embodiment of the present disclosure.

[0058] [Figure 48]1 shows a box plot identifying cancer methylation scores of HCC patients across different stages, according to some embodiments.

[0059] [Figure 49] 1 shows a set of survival curves identifying survival analysis in HCC patients according to some embodiments.

[0060] [Figure 50] Box plots identifying HCC methylation scores of HBV carriers and HCC patients calculated using data from SMRT-seq and nanopore sequencing are shown.

[0061] [Figure 51] Graph showing the percentage of liver-derived cfDNA identified by single molecule tissue of origin analysis in plasma samples from HBV carriers and HCC patients using data from SMRT-seq and nanopore sequencing.

[0062] [Figure 52] 1 shows box plots identifying the percentage of plasma DNA molecules classified as of colon origin in 15 healthy subjects, 45 HCC patients, and 4 CRC patients based on embodiments presented in the present disclosure.

[0063] [Figure 53] 1 shows a set of bar plots identifying the percentage of DNA molecules identified as being of HCC tumor origin between HCC patients with and without vascular invasion based on methylation haplotype analysis according to some embodiments.

[0064] [Figure 54] 1 shows a set of bar plots identifying the percentage of DNA molecules identified as being of HCC tumor origin, according to some embodiments.

[0065] [Figure 55]A set of ROC curves are shown that identify the cancer detection accuracy of an analysis of long cell-free DNA single molecule methylation sequence data and another analysis using short cell-free DNA methylation sequence data.

[0066] [Figure 56] A pair of ROC curves are shown identifying the HCC detection accuracy of methylation haplotype-based analysis using long DNA (>1 kb) and plasma DNA tissue mapping analysis using short-read bisulfite sequencing of short plasma DNA molecules (<600 bp).

[0067] [Figure 57] FIG. 1 shows a flowchart illustrating an example process for analyzing a subject's biological sample based on the methylation pattern of long cell-free DNA molecules, according to some embodiments.

[0068] [Figure 58] FIG. 1 shows box plots identifying single molecule methylation levels in different groups of individuals upon single molecule real-time sequencing (SMRT-Seq), according to some embodiments.

[0069] [Figure 59] Box plots are shown identifying single molecule methylation levels in DNA molecules >500 bp in size, containing at least three CpG sites, and with methylation levels <60% in SMRT-Seq.

[0070] [Figure 60] 1 shows ROC curves identifying the performance of single molecule methylation levels in distinguishing HCC and non-HCC subjects in SMRT-Seq and short read sequencing (e.g., Illumina sequencing), according to some embodiments.

[0071] [Figure 61] Box plots identifying single molecule methylation levels in HCC patients with different Barcelona Clinic Liver Cancer (BCLC) stages are shown.

[0072] [Figure 62] 1 shows a flowchart illustrating an example process for identifying disease classification using single molecule methylation levels in DNA molecules, according to some embodiments.

[0073] [Figure 63] FIG. 1 shows an illustrative diagram of pattern recognition of methylation haplotypes using machine learning models, according to some embodiments.

[0074] [Figure 64] 1 shows a set of bar graphs identifying the performance of machine learning models for distinguishing tumor and non-tumor DNA in plasma across various sequencing depths used in the training process.

[0075] [Figure 65] A set of bar graphs are shown identifying the performance of a machine learning model for distinguishing tumor and non-tumor DNA in plasma, where the machine learning model was trained using differentially methylated regions across different sequencing depths.

[0076] [Figure 66] 1 shows a table identifying the performance of machine learning models in distinguishing tumor and non-tumor DNA in the plasma of cancer patients with plasma DNA molecules of different lengths.

[0077] [Figure 67] FIG. 1 shows a flowchart illustrating an example process for identifying tissue-type characteristics based on methylation patterns of long cell-free DNA molecules using a machine learning model, according to some embodiments.

[0078] [Figure 68] FIG. 1 shows a schematic diagram illustrating an example of combined analysis using SNV and CpG methylation haplotype information, according to some embodiments.

[0079] [Figure 69] Characterization of a first group of plasma DNA molecules carrying wild-type alleles and a second group of plasma DNA molecules carrying mutations is shown.

[0080] [Figure 70] 1 shows a table identifying the distribution of the number of CpG sites in a 200 bp or 1 kb region surrounding a somatic mutation.

[0081] [Figure 71] FIG. 1 shows a schematic diagram of a DNA molecule with relative haplotype imbalance with skewed allele ratios and skewed methylation levels that are informative of the presence or absence of cancer.

[0082] [Figure 72] FIG. 1 shows a flowchart illustrating an example process for analyzing a biological sample using variants and methylation patterns to identify tissue of origin based on the methylation patterns of long cell-free DNA molecules, according to some embodiments.

[0083] [Figure 73] FIG. 1 shows a flowchart illustrating an example process for analyzing a biological sample using variants and methylation patterns to identify a cancer classification based on the methylation patterns of long cell-free DNA molecules, according to some embodiments.

[0084] [Figure 74] FIG. 1 shows a schematic illustrating an example process for training a machine learning model to distinguish between patients with and without cancer based on sequence composition, genomic location, fragment information, and epigenetic information present in plasma DNA molecules.

[0085] [Figure 75] FIG. 1 shows a schematic diagram illustrating an example process of applying the trained model to cancer detection using fragment information and a posteriori information present in plasma DNA molecules.

[0086] [Figure 76]FIG. 1 shows a flowchart illustrating an example process of subject biological sample analysis using a machine learning model to identify disease classification based on multiple features of long cell-free DNA molecules, according to some embodiments.

[0087] [Figure 77] 1 shows an example of a set of microsatellite sequences in a DNA molecule.

[0088] [Figure 78] 1 illustrates an example of detecting tumor-derived DNA based on cancer-specific microsatellite markers.

[0089] [Figure 79] 1 illustrates a measurement system according to an embodiment of the present invention;

[0090] [Figure 80] FIG. 1 illustrates a block diagram of an example computer system that can be used with systems and methods according to embodiments of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0091] term A "tissue" corresponds to a group of cells that group together as a functional unit. More than one type of cell may be found within a single tissue. Different types of tissues may consist of different types of cells (e.g., liver cells, alveolar cells, or blood cells), but may also correspond to tissues from different organisms (host vs. virus) or healthy vs. tumor cells. The term "tissue" may generally refer to any group of cells found in the human body (e.g., cardiac tissue, lung tissue, kidney tissue, nasopharyngeal tissue, oropharyngeal tissue). In some embodiments, the term "tissue" or "tissue type" may be used to refer to the tissue from which the cell-free nucleic acid originates. In one example, the viral nucleic acid fragment may be derived from blood tissue, e.g., Epstein-Barr virus (EBV). In another example, the viral nucleic acid fragment may be derived from tumor tissue, e.g., EBV or human papilloma virus infection (HPV).

[0092] The terms "sample", "biological sample", or "patient sample" are meant to include any tissue or material derived from a living or dead subject. A biological sample may be an acellular sample, which may include a mixture of nucleic acid molecules from a subject and possibly from a pathogen, e.g., a virus. A biological sample generally includes nucleic acid (e.g., DNA or RNA) or fragments thereof. The term "nucleic acid" may generally refer to deoxyribonucleic acid (DNA), ribonucleic acid (RNA), or any hybrid or fragment thereof. The nucleic acid in the sample may be acellular nucleic acid. A sample may be a liquid sample or a solid sample (e.g., a cell or tissue sample). A biological sample may be a body fluid such as blood, plasma, serum, urine, vaginal fluid, fluid from a hydrocele (e.g., of the testes), vaginal washing fluid, pleural fluid, peritoneal fluid, cerebrospinal fluid, saliva, sweat, tears, sputum, bronchoalveolar lavage fluid, nipple discharge, aspirated fluid from different parts of the body (e.g., thyroid, breast), etc. A stool sample may also be used. In various embodiments, the majority of the DNA in a biological sample enriched for cell-free DNA (e.g., a plasma sample obtained through a centrifugation protocol) may be cell-free (e.g., more than 50%, 60%, 70%, 80%, 90%, 95%, or 99% of the DNA may be cell-free). In some embodiments, at least 1,000 cell-free DNA molecules are analyzed. In other embodiments, at least 10,000 or 50,000 or 100,000 or 500,000 or 1,000,000 or 5,000,000 cell-free DNA molecules, or more, may be analyzed. At least the same number of sequence reads may be analyzed. The biological sample may be treated to physically disrupt tissue or cellular structures (e.g., by centrifugation and / or cell lysis), thereby releasing intracellular components into a solution that may further contain enzymes, buffers, salts, detergents, etc. to prepare the sample for analysis.

[0093] The term "native genome" (also called CG) is composed of consensus nucleotides at loci in a genome, and thus may be considered as a consensus sequence. A CG can cover the entire genome of a subject (e.g., human genome) or only a portion of the genome. A native genome (CG) can be obtained from cellular DNA and cell-free DNA (e.g., as may be found in plasma). Ideally, the consensus nucleotides should indicate that a locus is homozygous for one allele or heterozygous for two alleles. A heterozygous locus usually contains two alleles that are members of a genetic polymorphism. As an example, a threshold where two alleles each appear at least a certain percentage (e.g., 30% or 40%) in the reads aligned to a locus can be a specific criterion for identifying whether the locus is heterozygous or not. If one type of nucleotide appears at a sufficient percentage (e.g., 70% or more), the locus can be determined to be homozygous in the CG. Although the genome of one healthy cell may differ from that of another healthy cell due to random mutations that occur naturally during cell division, the CG should not change when such a consensus is used. Some cells may have genomes that are rearranged, for example B and T lymphocytes involved in antibody and T cell receptor genes. Such large-scale differences are still a relatively small proportion of the nucleated cell population in blood, and therefore, with sufficient sampling (e.g., sequencing depth) of blood cells, such rearrangements will not affect the identification of the innate genome. Other cell types, including oral cells, skin cells, hair follicles, or biopsies of various normal body tissues, may also serve as CG sources.

[0094] The term "innate DNA" refers to any source of DNA that reflects the genetic makeup that a subject was born with. For a subject, examples of "innate samples" from which innate DNA can be obtained include healthy blood cell DNA, buccal cell DNA, and hair root DNA. The DNA derived from these healthy cells defines the subject's CG. The cells can be identified as healthy in various ways, for example, when a person is known to be cancer-free, or when a sample can be obtained from tissue that is unlikely to contain cancer cells or pre-neoplastic cells (e.g., hair root DNA when liver cancer is suspected). As another example, a plasma sample can be obtained when the patient is cancer-free, and the determined innate DNA is compared to the results obtained from a subsequent plasma sample (e.g., one year or more later). In another embodiment, a single biological sample containing less than 50% tumor DNA can be used for inference of the innate genome and tumor-associated genetic changes. In such a sample, the concentration of tumor-associated single nucleotide mutations is lower than the concentration of each allele of the heterozygous SNP in the CG. Such a sample can be the same biological sample used for the determination of the sample genome described below.

[0095] A "sequence read" refers to a strand of nucleotides sequenced from any part or all of a nucleic acid molecule. For example, a sequence read can be a short strand of nucleotides sequenced from a nucleic acid fragment (e.g., 20-150 nucleotides), a short strand of nucleotides at one or both ends of a nucleic acid fragment, or a concatenation of the entire nucleic acid fragment present in a biological sample. Sequence reads can be obtained in a variety of ways, for example, using sequencing techniques, or using capture probes, for example, as may be used in a hybridization array or a microarray, or using amplification techniques such as polymerase chain reaction (PCR) or linear amplification using a single primer or isothermal amplification. As part of the analysis of a biological sample, at least 1,000 sequence reads may be analyzed. As other examples, at least 10,000 or 50,000 or 100,000 or 500,000 or 1,000,000 or 5,000,000 sequence reads, or more, may be analyzed. Such exemplary numbers of DNA molecules may be analyzed as part of massively parallel sequencing.

[0096] A sequence read may include an "end sequence" associated with an end of a fragment. An end sequence may correspond to the outermost N bases of a fragment, e.g., 1-30 bases at an end of a fragment. If a sequence read corresponds to an entire fragment, the sequence read may include two end sequences. If paired-end sequencing provides two sequence reads corresponding to ends of a fragment, each sequence read may include one end sequence.

[0097] A "sequence motif" may refer to a short repeating pattern of bases in a DNA fragment (e.g., cell-free DNA fragment). A sequence motif may occur at the end of a fragment and therefore may be part of or include a termination sequence. A "terminal motif" may refer to a sequence motif of a termination sequence that occurs preferentially at the end of a DNA fragment, potentially for a particular type of tissue. A terminal motif may also occur just before or just after the end of a fragment and therefore still correspond to a termination sequence. A nuclease may have a particular cleavage preference for a particular terminal motif and a second preferred cleavage preference for a second terminal motif.

[0098] An "end position" or "terminal position" (or simply "end") may refer to the genomic coordinate or genomic identity or nucleotide identity at the outermost base or end of a cell-free DNA molecule, e.g., a plasma DNA molecule. The terminal position may correspond to either end of the DNA molecule. Thus, when the beginning and end of a DNA molecule are referred to, both correspond to end positions. In practice, one terminal position is the genomic coordinate or nucleotide identity of the outermost base on one end of a cell-free DNA molecule detected or identified by analytical methods including, but not limited to, massively parallel or next-generation sequencing, single molecule sequencing, double-stranded or single-stranded DNA sequencing library generation protocols, polymerase chain reaction (PCR), or microarrays. Such in vitro techniques may alter the true in vivo physical end(s) of the cell-free DNA molecule. Thus, each detectable end may represent a biologically true end, or the end may be one or more nucleotides internal, or one or more nucleotides extended from the original end of the molecule, such as, for example, 5'blunting and 3'filling of the overhang of a non-blunt-ended double-stranded DNA molecule with Klenow fragment. The genomic identity or genomic coordinate of the end position may be derived from the result of alignment of the sequence reads to a human reference genome, such as hg19. It may be derived from a catalog of indexes or codes that represent the original coordinates of the human genome. It may refer to a position or nucleotide identity on a cell-free DNA molecule that is read by means including but not limited to target-specific probes, mini-sequencing, DNA amplification.

[0099] A "preferred end" (or "repeat end position") refers to an end that is more represented or more present (e.g., as measured by a rate) in a biological sample having a physiological (e.g., pregnancy) or pathological (disease) condition (e.g., cancer) than in a biological sample not having such a condition, or at a different time or stage of the same pathological or physiological condition, e.g., before or after treatment. A preferred end thus has an increased likelihood or probability of being detected in the relevant physiological or pathological condition compared to other conditions. The increased probability can be compared between pathological and non-pathological conditions, e.g., between patients with and without cancer, and quantified as a likelihood ratio or relative probability. The likelihood ratio can be determined based on the probability of detecting at least a threshold number of preferred ends in a sample being tested, or based on the probability of detecting a preferred end in patients with such a condition than in patients without such a condition. Examples of likelihood ratio thresholds include, but are not limited to, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.8, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5, 5, 6, 8, 10, 20, 40, 60, 80, and 100. Such likelihood ratios can be measured by comparing the relative abundance of samples with and without the relevant condition. Such preferred end positions are found in multiple individuals with the same physiological or disease condition, since there is a higher probability of detecting preferred ends in the relevant physiological or disease condition. With the increased probability, more than one cell-free DNA molecule can be detected as ending at the same preferred end position, even when the number of cell-free DNA molecules analyzed is much less than the size of the genome. Thus, preferred or repeated end positions are also referred to as "frequent end positions". In some embodiments, a quantitative threshold may be used to require that an end point be detected at least multiple times (e.g., 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, or 50 times) within the same sample or same sample aliquot to be considered a preferred end point. Relevant physiological conditions may include conditions when a person is healthy, disease-free, or free of the disease of interest. Similarly, a "preferred end window" corresponds to a set of consecutive preferred end positions.

[0100] The "rate" that a DNA molecule terminates at a position relates to the frequency with which the DNA molecule terminates at that position. Such a rate may be referred to as "end density". The rate may be based on the number of DNA molecules that terminate at the position normalized to the number of DNA molecules analyzed. Normalization may also be based on the representative, median, or total number of ends in the surrounding area. The surrounding area used for normalization may include, but is not limited to, 500, 1000, 3000, 5000 bp, etc., upstream and / or downstream from the position.

[0101] The term "allele" refers to alternative DNA sequences at the same physical genomic locus that may or may not result in different phenotypic characteristics. In any particular diploid organism (except for the sex chromosomes of male subjects) with two copies of each chromosome, the genotype of each gene includes pairs of alleles present at that locus that are the same in homozygotes and different in heterozygotes. A population or species of organisms typically contains multiple alleles at each locus in various individuals. Genomic loci at which more than one allele is found in a population are called polymorphic sites. Allelic variation at a locus can be measured as the number of alleles present (i.e., the degree of polymorphism) or the proportion of heterozygotes in a population (i.e., the heterozygosity rate). As used herein, the term "polymorphism" refers to any inter-individual variation in the human genome, regardless of its frequency. Examples of such variations include, but are not limited to, single nucleotide polymorphisms, simple tandem repeat polymorphisms, insertion-deletion polymorphisms, mutations (which may cause disease), and copy number variations. The term "haplotype" as used herein refers to a combination of alleles at multiple loci that are transmitted together on the same chromosome or chromosomal region. A haplotype can refer to just one pair of loci, or a chromosomal region, or an entire chromosome or chromosomal arm.

[0102] "Relative frequency" (also simply referred to as "frequency") may refer to a proportion (e.g., a percentage, fraction, or concentration). In particular, the relative frequency of a particular terminal motif (e.g., CCGA or only a single base) may provide the proportion of cell-free DNA fragments in a sample that are associated with the terminal motif CCGA, for example, by having a terminal sequence of CCGA. As another example, the relative frequency may be the rank order of occurrence of each motif relative to one another. Such rank orders have the same denominator, so a proportion or raw count may be used.

[0103] A "subread" is a sequence generated from all bases of one strand of a circularized DNA template that has been copied by a DNA polymerase into one continuous strand. For example, a subread may correspond to one strand of DNA of 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 a sequencing error.

[0104] A "site" (also called a "genomic site") corresponds to a single site, which may 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" may correspond to a region that includes multiple sites. A locus may contain only one site making the locus equivalent to the site in that context.

[0105] "Methylation state" refers to the state of methylation at a given site. For example, a site may be methylated, unmethylated, or in some cases, unspecified. A sequence read may include one or more sites at which the methylation state of the corresponding cell-free DNA molecule can be specified. Each site of the one or more sites may be associated with a methylation state. For example, the one or more sites may be CpG sites, and each site may be a CpG site with a specified methylation state. In some instances, each one or more sites of a sequence read includes at least N sites. For example, a given sequence read may include at least 3 CpG sites. Other numbers may be considered, including but not limited to at least 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 30, 35, 40, 45, 50, or more than 50 sites. Additionally or alternatively, the sequence reads can correspond to long cell-free DNA molecules having a size within a first size range (e.g., greater than 500 base pairs (bp)) and can include at least N sites (e.g., three CpG sites). As used herein, a "set of sites" can correspond to N sites.

[0106] The "methylation index" of each genomic site (e.g., CpG site) may refer to the proportion of DNA fragments (e.g., as determined from sequence reads or probes) that exhibit methylation at that site relative to the total number of reads that cover that site. A "read" may correspond to information obtained from a DNA fragment (e.g., the methylation state of a site). A read may be obtained using reagents (e.g., primers or probes) that preferentially hybridize to DNA fragments of a particular methylation state at one or more sites. Typically, such reagents are applied after treatment with processes that differentially modify or recognize DNA molecules 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 techniques that recognize methylcytosine and hydroxymethylcytosine (e.g., single molecule real-time sequencing and nanopore sequencing (e.g., by Oxford Nanopore Technologies). The methylation index may be converted to a binary value (0 or 1). For example, the methylation index may be recoded to 0 if the actual methylation index is ≦0.5, and the methylation index may be recoded to 1 if the actual methylation index is >0.5. The methylation index is a binary value when referring to methylation across individual CpG sites in a single DNA molecule.

[0107] The "methylation density" of a region may refer to the number of reads showing methylation at a site in this region divided by the total number of reads covering the site in the region. The site may have a specific characteristic, for example, a CpG site. Thus, the "CpG methylation density" of a region may refer to the number of reads showing CpG methylation divided by the total number of reads covering the CpG sites in this region (e.g., a specific CpG site, a CpG site in 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 after bisulfite treatment (corresponding to methylated cytosines) at the CpG sites 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. The region may be the whole 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 the region if the region contains only that CpG site. "Percentage of methylated cytosines" may refer to the number of cytosine sites "C" that are shown to be methylated (e.g., unconverted after bisulfite conversion) in this region relative to the total number of cytosine residues analyzed, i.e., including cytosines outside the context of CpG. Examples of "methylation levels" include methylation index, methylation density, the number of molecules that are methylated at one or more sites, and the percentage of molecules (e.g., cytosines) that are 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 status 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)).

[0108] 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 or portions of the genome.

[0109] A "methylation profile" includes information about DNA or RNA methylation at multiple sites or regions. Information about DNA methylation may include, but is not limited to, the methylation index of CpG sites, the methylation density (abbreviated as MD) of CpG sites in a region, the distribution of CpG sites over a contiguous region, the methylation pattern or level of each individual CpG site in a region that contains more than one CpG site, and non-CpG methylation. In some embodiments, a methylation profile may include the pattern of methylation or non-methylation of more than one type of base (e.g., cytosine or adenine). The methylation profile of a substantial portion of a genome may 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 a cytosine residue in a CpG dinucleotide (i.e., 5-methylcytosine). DNA methylation may occur to cytosines in other contexts, such as CHG and CHH, where H is adenine, cytosine, or thymine. Cytosine methylation may be in the form of 5-hydroxymethylcytosine.6 -Methyladenine has also been reported.

[0110] "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. When extending this concept to base modifications that are not limited to methylation, the term "modification pattern" is used, which refers 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, where "U" indicates an unmodified site and "M" indicates a modified site. An example of a non-methylation-based base modification is an oxidative change such as 8-oxoguanine.

[0111] The terms "hypermethylated" and "hypomethylated" may refer to the methylation density of a single DNA molecule 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 methylatable bases or nucleotides in that molecule. A hypermethylated molecule is one in which the methylation level of a single molecule is equal to or greater than a threshold value, which may be defined for each application. This threshold value may be 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 95%. A hypomethylated molecule is one in which the methylation level of a single molecule is equal to or less than a threshold value, which may be defined for each application and may vary for each application. This threshold value may be 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 95%.

[0112] The terms "hypermethylated" and "hypomethylated" may also refer to the methylation level of a population of DNA molecules as measured by the multi-molecule methylation level of these molecules. A hypermethylated population of molecules is a population where the multi-molecule methylation level is equal to or greater than 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 where the multi-molecule methylation level is equal to or less than a threshold, which may be defined for each application. This threshold may be 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 95%. In some embodiments, the population of molecules may be aligned to one or more selected genomic regions. In some embodiments, the selected genomic region(s) may be associated with a disease such as 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), 100nt, 200nt, 300nt, 500nt, 1000nt, 2knt, 5knt, 10knt, 20knt, 30knt, 40knt, 50knt, 60knt, 70knt, 80knt, 90knt, 100knt, 200knt, 300knt, 400knt, 500knt, or 1Mnt.

[0113] "Methylation-aware sequencing" refers to any sequencing method that allows the methylation state of a DNA molecule to be ascertained during the sequencing process, including, but not limited to, bisulfite sequencing, or sequencing preceded by methylation-sensitive restriction enzyme digestion, immunoprecipitation with anti-methylcytosine antibodies or methylation-binding proteins, or single molecule sequencing that allows elucidation of the methylation state (e.g., without bisulfite sequencing). Any such sequencing described herein may be massively parallel sequencing. "Methylation-aware assays" or "methylation-sensitive assays" may include both sequencing and non-sequencing-based methods, such as MSP, probe-based interrogations, hybridization, restriction enzyme digestion followed by densitometry, anti-methylcytosine immunoassays, mass spectrometry interrogations of the proportion of methylated cytosines or hydroxymethylcytosines, immunoprecipitation not followed by sequencing, etc.

[0114] 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 an entire genome. Sequencing depth may be expressed as 50x, 100x, etc., where "x" refers to the number of times a locus is covered by sequence reads. Sequencing depth may also be applied to multiple loci or an entire genome, in which case x may refer to the average number of times a locus or haploid genome or entire genome is sequenced, respectively. Ultra-deep sequencing may refer to a sequencing depth of at least 100x.

[0115] The term "level of cancer" may refer to whether cancer is present, the stage of cancer, the size of the tumor, whether there is metastasis, the total tumor burden in the body, and / or other measures of the severity of cancer. The level of cancer may be a number or other indicia such as a symbol, alphabetic letter, and color. The level may be zero. The level of cancer also includes pre-malignant or pre-cancerous conditions (conditions) that are related to the number of mutations or mutations. The level of cancer may be used in various ways. For example, screening may check whether cancer is present in a person who was not previously known to have cancer. Evaluation may test a person who has been diagnosed with cancer to monitor the progression of the cancer over time, study the effectiveness of a therapy, or determine a prognosis. In some embodiments, the prognosis may be expressed as the probability that the patient will die from cancer, or the probability that the cancer will progress or metastasize after a certain duration or time. Detection may mean "screening" or checking whether a person with suggestive characteristics of cancer (e.g., symptoms or other positive test results) has cancer.

[0116] "Level of pathology" (or level of injury) may refer to the amount, extent, or severity of pathology associated with an organism, and the level may be as described above for cancer. Another example of a pathology is the rejection of a transplanted organ. Other examples of pathologies may include genetic imprinting disorders, autoimmune attacks (e.g., lupus nephritis or multiple sclerosis, which 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 may be considered as a pathology-free classification.

[0117] A "separation value" corresponds to a difference or ratio involving two values, for example, two fractional contributions or two methylation levels. A separation value can be a simple difference or ratio. A separation value can include other factors, for example, multiplicative factors. As another example, a difference or ratio of a function of values ​​can be used, for example, the difference of the natural logarithms (ln) of two values.

[0118] "Separation values" and "aggregation values" (e.g., relative frequencies) are two examples of parameters (also called metrics) that provide a measure of the variation of a sample between different classifications (states) and can therefore be used to identify different classifications. An aggregation value can be a separation value, for example, when the difference is taken between a set of relative frequencies of a sample and a set of reference relative frequencies, as is done in clustering.

[0119] "Relative abundance" may correspond to a type of separation value that associates an amount (one value) of cell-free DNA molecules that end within one window of genomic locations with an amount (another value) of cell-free DNA molecules that end within another window of genomic locations. The two windows may overlap but have different sizes. In other embodiments, the two windows do not overlap. Furthermore, the width of the window may be one nucleotide, and therefore may correspond to one genomic location. End density is a type of relative abundance.

[0120] The term "classification" as used herein refers to any number(s) or other letter(s) associated with a particular characteristic of a sample. For example, a "+" sign (or the word "positive") may represent that the sample is classified as having a deletion or amplification. The classification may be a binary classification (e.g., positive or negative) or may have more classification levels (e.g., a scale of 1-10 or 0-1). The terms "cutoff" and "threshold" refer to a predetermined number used in an operation. For example, a cutoff size may refer to a size above which a fragment is excluded. A threshold may be a value above or below which a particular classification applies. Any of these terms may be used in any of these contexts.

[0121] The term "parameter" as used herein refers to a numerical value that characterizes a quantitative data set and / or a numerical relationship between quantitative data sets. For example, the ratio (or a function of the ratio) of a first amount of a first nucleic acid sequence to a second amount of a second nucleic acid sequence is a parameter.

[0122] The term "size profile" generally relates to the size of DNA fragments in a biological sample. A size profile can be a histogram that provides the distribution of a certain amount of DNA fragments of different sizes. Various statistical parameters (also called size parameters or simply parameters) can be used to distinguish one size profile from another. One parameter is the percentage of DNA fragments of a certain size or size range relative to all DNA fragments or DNA fragments of other sizes or ranges.

[0123] The terms "cutoff" and "threshold" refer to a predefined number used in an operation. For example, a cutoff size may refer to a size above which fragments are excluded. A threshold may be a value above or below which a particular classification is applied. Any of these terms may be used in any of these contexts. A cutoff or threshold may be a "reference value" or may be derived from a reference value that represents a particular classification or distinguishes between two or more classifications. Such a reference value may be determined in a variety of ways, as will be understood by those skilled in the art. For example, a metric may be determined for two different cohorts of subjects with different known classifications, and a reference value may 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, a reference value may be determined based on statistical analysis or simulation of samples. A particular value of a cutoff, threshold, reference, etc. may be determined based on a desired degree of accuracy (e.g., sensitivity and specificity).

[0124] The abbreviation "bp" refers to base pairs. In some cases, "bp" can be used to indicate the length of a DNA fragment that is single stranded and does not contain base pairs. In the context of single stranded DNA, "bp" may be understood to provide the length in nucleotides.

[0125] The abbreviation "nt" refers to nucleotides. In some cases, "nt" can be used to indicate the length of a single strand of DNA in bases. "nt" can also be used to indicate a relative location, such as upstream or downstream, of the analyzed locus. For double-stranded DNA, "nt" can also refer to the length of a single strand, not the total number of nucleotides in the two strands, unless the context clearly indicates otherwise. In some contexts related to technical conceptualization, data presentation, processing, and analysis, "nt" and "bp" may be used interchangeably.

[0126] The term "kinetic features" may 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 configurations, strand information, pulse intervals, pulse widths, and pulse strengths. Single molecule real-time sequencing continuously monitors the effect of polymerase activity on a DNA template. Thus, measurements generated from such sequencing can be considered kinetic features, e.g., nucleotide sequences.

[0127] The term "machine learning model" may include models based on predicting test data using sample data (e.g., training data), and therefore may include supervised learning. Machine learning models are often developed using a computer or processor. Machine learning models may include statistical models.

[0128] 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.

[0129] The term "real-time sequencing" can refer to techniques that involve data collection or monitoring during the progression of the reactions involved in sequencing. For example, real-time sequencing can involve optical monitoring or imaging of DNA polymerase incorporating new bases.

[0130] The term "about" or "approximately" may 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 limitations of the method of measuring or determining the value, i.e., the measurement system. For example, "about" may mean within one or more standard deviations, as is customary in the art. Alternatively, "about" may 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" may mean within an order of magnitude, within 5-fold, and more preferably within 2-fold of a value. When a particular value is described in the present application and claims, unless otherwise indicated, it should be assumed that the term "about" means within an acceptable error range of the particular value. The term "about" may have the meaning commonly understood by one of ordinary skill in the art. The term "about" may refer to ±10%. The term "about" may refer to ±5%.

[0131] Where a range of values ​​is provided, unless the context clearly indicates otherwise, it is understood that each intervening value between the upper and lower limits of that range is also specifically disclosed, down to one-tenth of the lower limit. Each smaller range between any stated or intervening value in a stated range and any other stated or intervening value in that stated range is encompassed within an embodiment of the disclosure. The upper and lower limits of these smaller ranges may be independently included or excluded in the range, and each range in which either, both, or neither limits are included in the smaller range is also encompassed within the disclosure, subject to any limit specifically excluded in the stated range. When a stated range includes one or both limits, ranges excluding either or both of those limits are also included in the disclosure.

[0132] Standard abbreviations may be used, e.g., bp is base pairs, kb is kilobase, pi is picoliters, s or sec is seconds, min is minutes, h or hr is hours, aa is amino acid, nt is nucleotide, etc.

[0133] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. Although any methods and materials similar or equivalent to those described herein can be used in the practice or testing of embodiments of the present disclosure, some possible exemplary methods and materials are described herein.

[0134] The technology involves analyzing the presence, abundance, and sequence characteristics of long cell-free DNA molecules in plasma samples from subjects with and without cancer. These characteristics can then be used to identify disease classification of subjects. The use of these long cell-free DNA molecules allows for analyses that are not intended or possible with shorter cell-free DNA fragments. For example, methylated CpG sites and single nucleotide polymorphism (SNP) status are often used to analyze DNA fragments in biological samples. CpG sites and SNPs are typically hundreds or thousands of base pairs away from the nearest CpG site or SNP. The length of most cell-free DNA fragments in biological samples is usually less than 200 bp. As a result, it is nearly impossible or impossible to find two or more consecutive CpG sites or SNPs on most cell-free DNA fragments. Cell-free DNA fragments longer than 200 bp, including those longer than 600 bp or 1 kb, may contain multiple CpG sites and / or SNPs. The presence of multiple CpG sites and / or SNPs on long cell-free DNA fragments allows for more efficient and / or accurate analysis than short cell-free DNA fragments alone.

[0135] In some embodiments, the methylation pattern of cell-free DNA molecules is used to identify the classification of a subject's disease. The methylation pattern of cell-free DNA molecules can include the methylation status of a set of sites (e.g., at least three CpG sites). The methylation status can indicate whether the corresponding sites are methylated or not. To identify the methylation pattern, methylation-aware sequencing (e.g., single molecule sequencing, nanopore sequencing) can be used to sequence the biological sample to obtain sequence reads, where each sequence read includes a respective methylation pattern. Long cell-free DNA molecules (e.g., larger than 600 bp in size) can be used because they can include a relatively large number of CpG sites (e.g., at least three CpG sites).

[0136] For each of the sequence reads, the methylation pattern of the sequence read is compared to one or more reference methylation patterns. Each of the one or more reference methylation patterns may be associated with a tissue type of the plurality of tissue types. In some instances, the reference methylation pattern of the one or more reference methylation patterns is associated with a disease with a known classification. For example, the comparison may include (i) determining, for each site of the set of sites, a similarity metric between the methylation state of the CpG site of the sequence read and the methylation index of the reference methylation pattern of the corresponding CpG site, and (ii) generating an aggregate value (e.g., sum) for the sequence read based on the similarity metric. Based on the comparison, a tissue classification of the sequence read (e.g., liver) may be identified based on the reference methylation pattern that most closely matches the methylation pattern of the sequence read. Continuing with this example, the reference methylation pattern that most closely matches the methylation pattern may be determined if the aggregate value of the reference methylation pattern is greater than one or more other aggregate values ​​of other reference methylation patterns. The tissue classification process may be repeated for each sequence read until a tissue classification of the sequence read is identified. A disease classification can then be identified based on the tissue classification, for example, a disease classification can be identified based on the amount of sequence reads classified as having a particular tissue classification (e.g., liver, lung, colon).

[0137] In some instances, the methylation pattern of each sequence read is input into a machine learning model to generate an output indicative of the tissue classification of the sequence read. The classification can be used to identify a characteristic of the tissue type (e.g., the amount of sequence reads classified as originating from that tissue type). The characteristic of the tissue type can also identify a disease state of a disease associated with that tissue type.

[0138] In some instances, the methylation pattern and one or more variants (e.g., polymorphisms) detected in the cell-free DNA molecules are used to identify the tissue of origin. For example, some plasma DNA molecules may have mutations that are not present in white blood cells. However, based on their respective methylation patterns, these plasma DNA molecules can be identified as associated with liver tissue. In some instances, the variants and methylation patterns are input into a machine learning model to generate an output, where the output is used to identify the tissue of origin of the cell-free DNA molecules.

[0139] The methylation pattern and one or more variants (e.g., polymorphisms) detected in cell-free DNA molecules can be used together to identify a cancer classification. For example, variants (e.g., single base variants) in plasma DNA molecules and the respective methylation patterns (e.g., multiple unmethylated states) of the sequences surrounding these variants can be used together to identify a hepatocellular carcinoma (HCC) classification.

[0140] In some embodiments, the amount of long cell-free DNA molecules is used to identify the classification of cancer in a subject. For example, the size of each cell-free DNA molecule is measured. The amount of cell-free DNA molecules having a size within a size range (e.g., a size greater than 1000 bp) may be identified. A normalized parameter can be determined from the identified amount of cell-free DNA molecules. For example, the normalized parameter can be determined by normalizing a first amount with a second amount of cell-free DNA molecules within a second size range (e.g., a size less than 150 bp). In some instances, the normalized parameter is a ratio value of the first amount and the second amount. The normalized parameter can then be used to identify the level of cancer.

[0141] In some embodiments, the frequency of the end motif of cell-free DNA molecules is used to identify disease classification. A biological sample is sequenced to obtain sequence reads. For each of the sequence reads, each sequence motif (e.g., CCCA) of the end sequence of the sequence read may be identified. Then, for each sequence motif of the set of N sequence motifs, a relative frequency may be determined. For example, the relative frequency of a sequence motif can be determined based on the proportion of cell-free DNA molecules having an end sequence corresponding to a sequence motif relative to the number of cell-free DNA molecules having an end sequence corresponding to other sequence motifs of the set of N sequence motifs.

[0142] The N frequency vector may be identified using the relative frequencies of a set of N motifs, where each of the N frequencies is normalized to each other or to other frequencies of sequence motifs in a group of reference samples. The vector may be compared to a plurality of reference vectors. The comparison may include determining the distance between the vector and a reference vector of the plurality of reference vectors. Each of the plurality of reference vectors is determined using a reference sample with a known disease classification. Based on the comparison, the disease classification of the subject can be identified. For example, the classification may include selecting the disease classification of the particular reference vector identified as having the shortest distance to the N frequency vector.

[0143] In some embodiments, the terminal motif frequency of cell-free DNA molecules with different size ranges is used to identify a disease classification. For one or more sequence motifs (e.g., CCCA), a first motif frequency may be determined for cell-free DNA molecules in a first size range (e.g., size greater than 1 kb), and a second motif frequency may be determined for cell-free DNA molecules in a second size range (e.g., size less than 200 bp). A separation value (e.g., ratio value) may be determined based on the first motif frequency and the second motif frequency. The separation value can be used to identify a disease classification. For example, the separation value may be compared to a cutoff value determined using a reference sample with a known disease classification. In another example, the separation value may be processed using a machine learning model to identify a disease classification, where the machine learning model (e.g., logistic regression, support vector machine) is trained using a training sample with a known disease classification.

[0144] In some embodiments, the machine learning model is trained using various features of the training dataset to distinguish between reads from the first tissue and other tissues. Based on the distinction, a cancer classification can be identified. The sequence reads may be obtained from a plasma DNA sample. In some instances, at least some of the sequence reads have a length greater than a threshold size (e.g., 600 bp). For each sequence read, one or more features are identified. The one or more features may include, for the sequence read, the location of the ends in the reference genome, the sequence composition, the size, the sequence motif at one or more ends, or the DNA methylation pattern. The features may be input to the trained machine learning model. The machine learning model can generate an output that can be used to identify a classification of the sequence read. The classification can identify whether the sequence read is from the first tissue type or another tissue type. The classification of the sequence reads may then be used to identify a disease classification. For example, the amount of sequence reads classified as being from the first tissue type may be identified, and the amount may be used to identify a disease classification.

[0145] In some embodiments, the single molecule methylation level of cell-free DNA molecules is used to identify the level of a pathology of a subject. For example, the percentage of methylation sites is determined for each DNA molecule of the plurality of cell-free DNA molecules. In some examples, the plurality of cell-free DNA molecules has a size above a threshold value (e.g., 500bp). The determined percentage of methylation sites of the plurality of cell-free DNA molecules can be used to determine a statistical value (e.g., a mean value, a median value). The statistical value can be compared to a reference to identify a pathology.

[0146] Analysis of long cfDNA molecules can provide a previously unexplored added value in cancer detection and evaluation. Longer cfDNA molecules have more CpG sites and therefore can have more specific methylation patterns, profiles or haplotypes than shorter molecules. Thus, the permutations of methylated and unmethylated sites are much larger. This can improve the identification of DNA molecules originating from any particular tissue, i.e. tissue of origin analysis.

[0147] Such tissue-of-origin analysis is distinct from previously known techniques that use short-read sequencing to analyze short cell-free DNA molecules. Due to the limited number of CpG sites on short cell-free DNA molecules, previous methods used population statistics on a population / multiple short cell-free DNA molecules to assemble methylation profiles of cell-free DNA content in plasma samples. This approach could only infer the relative contributions of cell-free DNA molecules originating from different tissues or organs. Due to the DNA methylation pattern specificity conferred by the higher number of CpG sites on long cell-free DNA molecules, it is believed that identifying the tissue of origin of such individual long cell-free DNA molecules is feasible. In other words, individual molecules can be assigned to a tissue of origin or organ of origin.

[0148] Another expected advantage of analyzing long cell-free DNA molecules is the potential to link sequence variants on the molecule with methylation information of adjacent CpGs on the same molecule. Indeed, analysis of long cell-free DNA molecules allows the analysis of two or more molecular (e.g., congenital or acquired) features on such molecules. Examples include (i) two or more sequence variations (e.g., point mutations, microsatellite variations, etc.), (ii) two or more acquired variations (e.g., two or more hyper- or hypo-methylated CpG sites), and (iii) different combinations of congenital and acquired changes. In addition, malignant tumors are known to have a higher rate of cell death, so the abundance of long cell-free DNA molecules released from tumors may differ from non-tumor tissues. In the present disclosure, several approaches are invented to analyze long cell-free DNA fragments to enable detection and monitoring of cancer and many other diseases, including but not limited to autoimmune diseases, organ transplant rejection, trauma, ischemia, necrosis, etc. In some embodiments, the approaches presented in this disclosure can be used for prognosis, risk stratification, treatment guidance, etc. I. Techniques for analyzing long cell-free DNA molecules

[0149] Cell-free DNA was collected from plasma samples of patients with cancer and subjects without cancer. Such cell-free DNA was subjected to single molecule sequencing for various analyses, including but not limited to methylation haplotype analysis, tissue of origin of individual plasma DNA molecules, fragment size profiling, plasma DNA end analysis, ragged end analysis, microsatellite instability, etc. Information regarding techniques for identifying various features (e.g., methylation status, ragged ends) of long cell-free DNA molecules is further described in U.S. Patent Application No. 16 / 995,607, the entire contents of which are incorporated herein by reference for all purposes. A. Detection of long cell-free DNA molecules in biological samples

[0150] FIG. 1 shows a schematic diagram 100 illustrating an example overview of analyzing long cell-free DNA molecules, according to some embodiments. In one example, the analysis may include sequencing, e.g., single molecule sequencing. Single molecule sequencing may include, but is not limited to, single molecule real-time sequencing (i.e., SMRT-seq) (e.g., by Pacific Biosciences, PacBio SMRT-seq) and nanopore sequencing (e.g., by Oxford Nanopore Technologies). Nucleotides of each sequenced DNA molecule can be identified according to electrical or optical signals generated during the sequencing process. The identified nucleotides can be used for subsequent analysis of the corresponding long cell-free DNA molecule. Additionally or alternatively, other sequencing techniques may be used to detect long cell-free DNA. For example, cluster-based sequencing may include sequencing each end of a given fragment (e.g., 200 bp or more) to generate sequence reads for the identified nucleotide sequence (e.g., 400 bp or more).

[0151] For example, the length of plasma DNA can be determined by counting the number of nucleotides present in the sequence. 4-mer terminal motifs in plasma DNA can be identified by analyzing 4 nucleotides at their ends. Similarly, in some embodiments, other types of terminal motifs can be used, including but not limited to 1-mer, 2-mer, 3-mer, 5-mer, 6-mer, 7-mer, 8-mer, 9-mer, 10-mer, 15-mer, 20-mer, or other combinations.

[0152] In some embodiments, analysis of plasma DNA molecules between cancer and non-cancer subjects may also include ragged ends (i.e., the original double strand with a single stranded overhang(s)) and microsatellite instability, which refers to genomic alterations in which microsatellites, usually 1-6 nucleotide repeats, accumulate mutations corresponding to deletions / insertions of one or more nucleotides.

[0153] In some embodiments, PacBio SMRT-seq can be used to identify methylation status across a range of CpG sites in plasma DNA molecules by analyzing DNA polymerase kinetic signals in a measurement window following, but not limited to, a previously published approach (Tse et al. Proc Natl Acad Sci USA. 2021;118: e2019768118). Additionally or alternatively, nanopore sequencing can be used to determine the methylation state across a range of CpG sites in plasma DNA molecules by analyzing electrical signals in response to DNA molecules passing through a nanopore, according to the tools presented in U.S. Application No. 63 / 173,728, published approaches, such as but not limited to the open source software Nanopolish (Simpson et al. Nat Methods. 2017;14:407-410), DeepMod (Liu et al. Nat Commun. 2019;10:2449), Tombo (Stoiber et al. BioRxiv. 2017:p.094672), DeepSignal (Ni et al. Bioinformatics. 2019;35:4586-4595), Guppy (github.com / nanoporetech), Megalodon (github.com / nanoporetech / megalodon), etc. In some embodiments, methylation patterns can be obtained by chemical (e.g., bisulfite) or enzymatic (e.g., TET2 and APOBEC) conversion followed by PacBio SMRT-seq and / or nanopore sequencing. Enzymatic conversion converts unmethylated cytosines to uracils that are amplified and sequenced as thymines, while methylated cytosines remain unchanged. Thus, methylation status can be determined by detecting thymines (unmethylated signals) or cytosines (methylated signals) throughout CpG sites in a reference genome.

[0154] In some embodiments, sequenced reads are aligned to the human reference genome using Minimap2 (Li H. Bioinformatics. 2018;34(18):3094-3100). In some embodiments, BLASR (Mark J Chaisson et al. BMC Bioinformatics. 2012; 13: 238), 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), and LAST (Kielbasa SM et al. Genome Res. 2011;21(3):487-493) are used to align sequenced reads to a reference genome. In some embodiments, alignment of sequenced reads is not used. B. Identifying the methylation status of long cell-free DNA molecules

[0155] As described herein, the methylation status across CpG sites can be obtained by analyzing the kinetic signatures generated during SMRT sequencing. For example, taking Pacific Biosciences SMRT sequencing as an example of single molecule real-time sequencing for illustrative purposes, a DNA polymerase molecule is placed at the bottom of a well that acts as a zero mode waveguide (ZMW). A ZMW is a nano-optical device for confining light to a small observable volume, which can be a hole with a very small diameter, and does not allow the propagation of light in the wavelength range used for detection, so that only the emission of optical signals 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.

[0156] FIG. 2 shows an example of a molecule 200 with methylated and / or unmethylated CpG sites sequenced by single molecule real-time sequencing. A DNA molecule is first ligated to a hairpin adaptor to form a circularized molecule that binds to an immobilized DNA polymerase to initiate DNA synthesis. In FIG. 2, a DNA molecule 202 is ligated to a hairpin adaptor to form a linked molecule 204. Linked molecule 204 then forms a circularized molecule 206. Molecules without CpG sites can also be sequenced. Circularized molecule 206 contains an unmethylated CpG site 208 that can still be sequenced.

[0157] Once the methylation status across CpG sites in a plasma DNA molecule (herein referred to as methylation haplotype) has been identified, the methylation haplotype of the plasma DNA molecule can be compared with the methylation haplotypes of various tissues to identify the tissue of origin of the plasma DNA molecule. In other words, a methylation haplotype was defined as a methylation pattern across one or more CpG sites in a single DNA molecule. For example, "-MUMMM-" represented a methylation haplotype, which indicated a methylated CpG followed by an unmethylated CpG followed by three consecutive methylated CpG sites. The methylation haplotype information of "-MUMMM-" and "-MUMMU-" was different. The aforementioned tissues may include, but are not limited to, neutrophils, T cells, B cells, megakaryocytes, erythrocytes, monocytes, NK cells, liver, lung, esophagus, heart, pancreas, colon, small intestine, adipose tissue, adrenal gland, brain, breast, kidney, bladder, thyroid, prostate, uterus, etc. The tissue may be involved in cancers such as, but not limited to, bladder cancer, breast cancer, colon and rectal cancer, endometrial cancer, kidney cancer, leukemia, liver cancer, lung cancer, melanoma, non-Hodgkin's lymphoma, pancreatic cancer, prostate cancer, thyroid cancer, and the like. 1. Dynamic characteristics

[0158] Some embodiments of the methods disclosed herein are based on measuring and utilizing the interpulse distance (IPD), pulse width (PW), and sequence configuration of every base within a measurement window. We speculate that if a combination of multiple metrics could be used, e.g., features such as upstream and downstream sequence configuration, strand information, IPD, pulse width, and pulse strength could be simultaneously utilized, precise measurement of base modifications (e.g., mC detection) at single base resolution may be achieved. Sequence configuration 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 that is the subject or target of base modification analysis. In one embodiment, the stretch of DNA may be in close proximity to the base that is the subject of base modification analysis. In another embodiment, the stretch of DNA may be far away from the base that is the subject of base modification analysis. The stretch of DNA may be upstream and / or downstream of the base that is the subject of base modification analysis.

[0159] In one embodiment, the upstream and downstream sequence composition, strand information, IPD, pulse width and pulse strength features for use in base modification analysis are referred to as kinetic features.

[0160] There is a need for techniques to detect modifications in bases without enzymatically or chemically converting the modifications and / or bases. As described herein, modifications in target bases can be detected using kinetic feature data obtained from single molecule real-time sequencing of bases surrounding the target base. Kinetic features can include pulse intervals, pulse widths, and sequence configurations. These kinetic features can be obtained for a measurement window of a certain number of nucleotides upstream and downstream of the target base. These features (e.g., at specific positions in the measurement window) can be used to train machine learning models. As an example of sample preparation, two strands of a DNA molecule can be joined to hairpin adaptors to form a circular DNA molecule. The circular DNA molecule allows for the acquisition of kinetic features of either or both of the Watson and Crick strands. A data analysis framework can be developed based on the kinetic features in the measurement window. This data analysis framework can be used to detect modifications, including methylation. This section describes various techniques for detecting modifications.

[0161] FIG. 3 shows a schematic diagram 300 illustrating an example process for identifying kinetic characteristics of cell-free DNA molecules, according to some embodiments. As shown in FIG. 3, as an example, a sub-read of a Watson strand was obtained from Pacific Biosciences SMRT sequencing to analyze one specific base for base modification status. In FIG. 3, three bases from each side of the base subjected to base modification analysis are defined as a measurement window 300. In one embodiment, the sequence composition, IPD, and PW of these seven bases (i.e., three nucleotides (nt) upstream and downstream sequences and one nucleotide for base modification analysis) were compiled into a two-dimensional (i.e., 2D) matrix as a measurement window. In the example shown, the measurement window 300 is used for one sub-read of a Watson strand. Other variations are described herein.

[0162] The first row 302 of the matrix indicated the sequences that were tested. In the second row 304 of the matrix, the 0 position represented the base for base modification analysis. The relative positions of -1, -2, and -3 indicated positions 1-nt, 2-nt, and 3-nt upstream of the base that was subjected to base modification analysis, respectively. The relative positions of +1, +2, and +3 indicated positions 1-nt, 2-nt, and 3-nt downstream of the base that was subjected to base modification analysis, respectively. Each position included two columns containing the corresponding IPD and PW values. The next four rows (rows 308, 312, 316, and 320) corresponded to the four types (A, C, G, and T) of nucleotides in the strand (e.g., Watson strand), respectively. The presence of IPD and PW values ​​in the matrix depended on which corresponding nucleotide type was sequenced at a particular position. As shown in Figure 3, at the relative position of 0, the IPD and PW values ​​are displayed in the row showing "G" in the Watson strand, which indicates that a guanine was read in the sequence result at that position. Other cells in the column that do not correspond to the sequenced base are coded as "0". As an example, the sequence information corresponding to the 2D digital matrix (Figure 3) is 5'-GATGACT-3' for the Watson strand.

[0163] FIG. 4 shows a schematic diagram 400 illustrating another example process for identifying dynamic characteristics of cell-free DNA molecules, according to some embodiments. As shown in one embodiment depicted in FIG. 4, a measurement window may be applied to data from a Crick strand. Subreads of the Crick strand were obtained from single molecule real-time sequencing to analyze one specific base for base modification status. In FIG. 4, three bases from each side of the base subjected to base modification analysis and the base subjected to base modification analysis are defined as a measurement window. In one embodiment, the sequence composition, IPD, and PW of these seven bases (i.e., three nucleotides (nt) upstream and downstream sequences and one nucleotide for base modification analysis) were compiled into a two-dimensional (i.e., 2D) matrix as a measurement window.

[0164] The first row of the matrix indicated the sequence that was tested. In the second row of the matrix, the position 0 represented the base for base modification analysis. The relative positions of -1, -2, and -3 indicated the positions 1-nt, 2-nt, and 3-nt upstream of the base that was subjected to base modification analysis, respectively. The relative positions of +1, +2, and +3 indicated the positions 1-nt, 2-nt, and 3-nt downstream of the base that was subjected to base modification analysis, respectively. Each position included two columns containing the corresponding IPD and PW values. The next four rows corresponded to the four types of nucleotides (A, C, G, and T) in this strand (e.g., the Crick strand). The presence of IPD and PW values ​​in the matrix depended on which corresponding nucleotide type was sequenced at a particular position. As shown in Figure 4, at the relative position of 0, the IPD and PW values ​​were displayed in the row indicating "T" in the Crick strand, which suggests that thymine was read in the sequence results at that position. Other lattices in the column that do not correspond to a sequenced base are coded as "0." As an example, the sequence information corresponding to the 2D digital matrix (FIG. 4) is 5'-ACTTAGC-3' for the Crick strand. 2. Machine learning models

[0165] For the machine learning model, the input data structures of the sub-reads can be used for training. The input data structures can correspond to windows of nucleotides sequenced in the sample nucleic acid molecule. The training set can have sites with known methylation states. Each training sample can include one of the first plurality of first data structures and a label indicating a first state of modification (e.g., methylation) of the nucleotide at the target position. Training is performed by optimizing parameters of the model based on model outputs that match or do not match corresponding labels among the first label and optionally the second label when the first plurality of first data structures and optionally the second plurality of second data structures are input to the model. The model output specifies whether the nucleotide at the target position in each window has a modification or not. In some embodiments, the model output can include a probability of being in each of the plurality of states. The state with the highest probability can be taken as the state.

[0166] The model may include a convolutional neural network (CNN). The CNN may include a set of convolutional filters configured to filter the first and optionally the second plurality of data structures. The filters may be any filter described herein. The number of filters in each layer may be 10-20, 20-30, 30-40, 40-50, 50-60, 60-70, 70-80, 80-90, 90-100, 100-150, 150-200, or more. The kernel size of the filters may be 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 15-20, 20-30, 30-40, or more. The CNN may include an input layer configured to receive the filtered first and optionally the filtered second plurality of data structures. The CNN may also include multiple hidden layers including multiple nodes. A first layer of the multiple hidden layers is coupled to the input layer. The CNN may further include an output layer coupled to a last layer of the multiple hidden layers and configured to output an output data structure. The output data structure may include the characteristics.

[0167] The models may include supervised learning models, which may include a variety of approaches and algorithms, including analytical learning, artificial neural networks, backpropagation, boosting (meta-algorithms), Bayesian statistics, case-based reasoning, decision tree learning, inductive logic programming, Gaussian process regression, genetic programming, group methods of data processing, kernel estimators, learning automata, learning classifier systems, minimum message length (decision trees, decision graphs, etc.), multiple linear subspace learning, naive Bayes classifiers, maximum entropy classifiers, conditional random fields, nearest neighbor algorithms, probabilistic approximately correct learning (PAC) learning, ripple down rules, knowledge acquisition methods, symbolic machine learning algorithms, sub-symbolic machine learning algorithms, support vector machines, minimal complexity machines (MCM), random forests, ensembles of classifiers, regular classification, data pre-processing, processing of imbalanced data sets, statistical relationship learning, or multi-criteria classification algorithm Proaftn. The model can be a linear regression, a logistic regression, a deep recurrent neural network (e.g., long short-term memory, LSTM), a Bayesian classifier, a hidden Markov model (HMM), a linear discriminant analysis (LDA), K-means clustering, density-based spatial clustering of applications with noise (DBSCAN), a random forest algorithm, a support vector machine (SVM), or any model described herein. II. Frequency-based analysis of long cell-free DNA molecules

[0168] The amount of long cell-free DNA molecules present in plasma may depend on the disease state of a particular subject.For example, the first amount of long cell-free DNA molecules present in the biological sample of a subject with hepatocellular carcinoma (HCC) may be smaller than the second amount of long cell-free DNA molecules present in the biological sample of another subject who is a Hepatitis B Virus (HBV) carrier.Therefore, the long cell-free DNA molecules of HCC patients and HBV carriers can be sequenced using single molecule real-time sequencing (e.g., by PacBio sequencer) to identify these amount-based features.

[0169] In some embodiments, a long DNA molecule is defined as a DNA molecule having a length of 500bp, 600bp, 1kb, 2kb, 3kb, 4kb, 5kb, 10kb or more, or more than 10kb. In some instances, a long DNA molecule is defined as having a size within a size range. A size range may include a lower limit and an upper limit. The lower limit identifies the minimum size of a cell-free DNA molecule that is considered to be a long DNA molecule. For example, the lower limit of a size range includes at least 200bp, at least 300bp, at least 400bp, at least 500bp, at least 600bp, at least 700bp, at least 800bp. Conversely, the upper limit identifies the maximum size of a cell-free DNA molecule that is considered to be a long DNA molecule. For example, the upper limit of a size range includes at least 500bp, 600bp, 1kb, 2kb, 3kb, 4kb, 5kb, 10kb, or more than 10kb. In some instances, a size range specifies only a lower limit and not an upper limit. The above lengths are non-limiting and other types of lengths may be considered. A. Comparison of frequency-based analysis using long vs. short cell-free DNA molecules

[0170] Plasma DNA samples from 5 patients with chronic hepatitis B infection (HBV carriers) and 19 patients with HCC were subjected to single molecule real-time (SMRT) sequencing template construction using SMRTbell Express Template Prep Kit 2.0 (Pacific Biosciences). DNA was purified with 1.8× AMPure PB beads, and library size was estimated using a TapeStation instrument (Agilent). Sequencing primer annealing and polymerase binding conditions were calculated with SMRT Link v10.1 software (Pacific Biosciences). Briefly, sequencing primer v4 was annealed to the sequencing template, and then polymerase was bound to the template using Sequel II Binding Kit 2.1 and Internal Control Kit 1.0 (Pacific Biosciences). Sequencing was performed on a SMRT Cell 8M. Sequencing movies were collected for 30 h using the Sequel II Sequencing 2.0 Kit (Pacific Biosciences). A median of 314,477 sequenced reads (interquartile range (IQR): 128,791-561,018) were obtained. DNA methylation status across CpG sites in plasma DNA molecules was identified according to the HK model (Tse et al. Proc Natl Acad Sci USA. 2021;118;e2019768118). For comparison, short-read sequencing (e.g., Illumina sequencing) was performed on the same plasma DNA samples. The length of each sequence read was identified for sequence reads corresponding to HCC samples. The size distribution of sequence reads with a length greater than 500 bp was identified for each Illumina sequencer result and SMRT sequence result.

[0171] FIG. 5 shows a graph 500 identifying the percentage of plasma DNA fragments with lengths greater than 500 bp across various sequencing techniques, according to some embodiments. FIG. 5 shows that the percentage of plasma DNA fragments >500 bp in patients with HCC is much higher in single molecule real-time sequencing (SMRT-seq) results (median: 22.88%, range: 11.64%-40.46%) than in Illumina sequencing results (median: 0.68%, range: 0.34%-1.24%) (P-value<0.0001, Mann-Whitney U test). The data suggested that there was a substantial amount of long plasma DNA molecules that could not be investigated based on Illumina sequencing.

[0172] FIG. 6 shows a line graph 600 illustrating the size distribution of one HCC subject 602 and one HBV carrier 604. SMRT-seq was used to generate sequence reads for each of the samples. The Y-axis corresponds to frequency values ​​shown on a logarithmic scale (e.g., a normalized parameter of the size distribution). Both size profiles displayed peaks of nucleosome sizes at 166bp, 333bp, 500bp, 663bp, 830bp, 994bp, etc. The frequency of DNA lengths longer than 1 kb appeared to decline faster in patients with HCC than in HBV carriers. In some embodiments, changes in the size profile can be used to distinguish patients with and without HCC.

[0173] The above results show that the size distribution of tumor samples can be used for disease classification.In some embodiments, a cutoff is determined to classify whether a sample contains cancer or not.The cutoff can correspond to a normalized parameter that represents a certain amount or frequency of cell-free DNA molecules that have a certain length (e.g., 600bp) or more. B. Prediction of histological status of disease based on frequency of long DNA molecules

[0174] Vascular invasion in HCC is a prerequisite for systemic tumor dissemination and is the best predictor of tumor recurrence after transplantation or tumor resection (Thuluvath. J. Clin. Gastroenterol. 2009;43:101-2). Although several studies have suggested that circulating plasma DNA concentration correlates with vascular invasion status (Huang et al. Pathol. Oncol. Res. 2012;18: 271-276) and tumor-associated mutations (Oversoe et al. Scand. J. Gastroenterol. 2020;55:1433-1440; Liao et al. Oncotarget. 2016; 7:40481-40490), it is unclear whether the size characteristics of cfDNA are related to vascular invasion.

[0175] To explore the size characteristics associated with vascular invasion in DNA molecules, plasma DNA in HCC patients was examined using single molecule real-time sequencing. In the cohort, 18 patients had vascular invasion, while 27 patients did not. Figure 7 shows a bar graph 700 that identifies the percentage of cfDNA fragments above a given size for HCC patients 702 with vascular invasion and HCC patients 704 without vascular invasion. Red bars indicate HCC cases with vascular invasion, and teal bars indicate HCC cases without vascular invasion. In addition, the x-axis shows the percentage of DNA molecules longer than a given size cutoff (e.g., 200bp, 500bp, 2kbp). As shown in Figure 7, plasma DNA from subjects with vascular invasion has a shorter size distribution than subjects without vascular invasion, and this difference is evident up to a size of 2kb, which could not be revealed by previous sequencing methods such as Illumina sequencing.

[0176] In some embodiments, the percentage of DNA fragments larger than a certain size (e.g., ≧200bp, ≧500bp, ≧600bp, ≧1kb, ≧2kb, ≧3kb, ≧4kb, ≧5kb, ≧10kb, or other combinations) can be used to non-invasively predict the vascular invasion status of cancer patients. FIG. 8 shows a box plot 900 identifying the percentage of long DNA fragments >200bp in HCC patients with and without vascular invasion. SMRT sequencing was used to identify the long DNA molecules. As shown in FIG. 8, HCC patients with vascular invasion had a significantly lower percentage of long DNA fragments >200bp (P value: 0.015, Mann-Whitney U test), suggesting a potential use in predicting the vascular invasion status of HCC patients. Furthermore, a potential use in predicting the vascular invasion status can enable assessment and prognosis of recurrence risk in a non-invasive manner.

[0177] Additionally or alternatively, the size ratio of long DNA fragments to short DNA fragments can be used to non-invasively predict vascular invasion status in cancer patients. FIG. 9 shows a box plot 900 identifying the size ratio of HCC patients with and without vascular invasion. The size ratio of HCC patients with and without vascular invasion was calculated by dividing the proportion of long DNA fragments (>500bp) by short DNA fragments (<150bp). As shown in FIG. 9, HCC patients with vascular invasion have a significantly lower size ratio than HCC patients without vascular invasion (P value: 0.004, Mann-Whitney U test). The results in FIG. 9 show its potential use in predicting vascular invasion status in HCC patients and in realizing recurrence risk assessment and prognosis in a non-invasive manner. C. Methods for Frequency-Based Analysis of Long Cell-Free DNA Molecules

[0178] 10 shows a flowchart 1000 illustrating an example process for analyzing a subject's biological sample based on the frequency of long cell-free DNA molecules, according to some embodiments. The biological sample may contain DNA originating from normal cells and possibly cells associated with cancer. In addition, at least a portion of the DNA is cell-free in the biological sample.

[0179] In step 1002, the size of a plurality of cell-free DNA molecules from a biological sample may be measured. For example, single molecule real-time sequencing (i.e., SMRT-seq) (e.g., by Pacific Biosciences, PacBio SMRT-seq) and nanopore sequencing (e.g., by Oxford Nanopore Technologies) may be used to identify and count the nucleotides of the cell-free DNA molecules. The number of nucleotides may be counted to determine the size of the cell-free DNA molecules. In other embodiments, sequence reads at each end of the DNA molecules may be sequenced, and pairs of reads may be aligned to a reference genome to determine the size of the DNA molecules.

[0180] In step 1004, a first amount of cell-free DNA molecules having a size within a first size range may be measured. The first size range includes an upper limit of at least 1,000 bases, at least 3,000 bases, or more. In some instances, the first size range includes a lower limit greater than zero. The lower limit may be selected from one of at least 300 bases, at least 400 bases, at least 500 bases, at least 600 bases, or at least 800 bases. Thus, some examples of size ranges are 300-1000 bp, 300-3000 bp, 300-800 bp, 400-800 bp, 400-1500 bp, and 500-3000 bp.

[0181] Additionally or alternatively, the first amount of cell-free DNA molecules may have end sequences corresponding to one or more sequence motifs (e.g., CCCA). To identify cell-free DNA molecules having one or more sequence motifs, sequence reads are obtained from sequencing a plurality of cell-free DNA molecules from a biological sample. For each of the sequence reads, a sequence motif is identified for each of the one or more end sequences of the corresponding cell-free DNA molecule. A group of a plurality of cell-free DNA molecules having at least one of the set of one or more sequence motifs in the end sequence may be identified. And the first amount is the amount of a subgroup of the group of a plurality of cell-free DNA molecules having a first size range.

[0182] In step 1006, the first amount may be used to generate a value of a normalized parameter. The normalized parameter may be the frequency of cell-free DNA molecules in the biological sample having a size within a first size range. In some instances, the normalized parameter may be the frequency of cell-free DNA molecules in the biological sample having a size within the first size range, normalized on a logarithmic scale. In other instances, a second amount of cell-free DNA molecules in a second size range may be used to normalize the first amount. The second size range may be different from the first size range. For example, the second size range may be less than the first size range (e.g., 1-150 bp).

[0183] In step 1008, the normalized parameters may be used to identify a classification of the level of cancer. For example, the normalized parameters may be compared to a cutoff value. In some instances, a cutoff is determined to classify a sample as containing cancer or not. The cutoff may correspond to a normalized parameter that represents a certain amount or frequency of cell-free DNA molecules of a reference sample having a length equal to or greater than a certain length (e.g., 600 bp), where the reference sample is associated with a cancer of known level classification. The cutoff value or comparison may be determined using machine learning with a training dataset, e.g., the training sample from FIG. 6.

[0184] Cut-off values ​​and comparisons for other methods may also be determined using machine learning with training data sets. The comparison of the normalized parameters and cut-offs (references) may involve, for example, machine learning models trained using supervised learning. In some instances, the cut-off values ​​are determined using one or more training data sets that include reference samples with cancers of known classification levels. For example, the normalized parameters or separation values ​​(and possibly other criteria, such as copy number and methylation levels) and training subjects of known classification from which the training samples were taken may form the training data set. The parameters of the machine learning model may be optimized based on the training set to provide an optimized accuracy of classifying the level of cancer. Examples of machine learning models include neural networks, decision trees, clustering, and support vector machines.

[0185] The level of cancer may include no cancer, early stage, intermediate stage, or advanced stage. And the classification may be selected from one of the levels. Thus, the classification may be identified from multiple levels of cancer, including multiple stages of cancer. For example, the cancer may be hepatocellular carcinoma, lung cancer, breast cancer, gastric cancer, glioblastoma multiforme, pancreatic cancer, colorectal cancer, nasopharyngeal carcinoma, and head and neck squamous cell carcinoma. Identification of the disease classification may include the histological state of the cancer, for example, the presence or absence of vascular invasion. III. Terminal motif analysis of long and short cfDNA molecules

[0186] Terminal motifs of cell-free DNA molecules of biological samples can be identified and used for disease classification. In addition to previous studies (Jiang et al. Cancer Discov. 2020; 10:664-673) in which it was feasible to use terminal motif signatures for cancer diagnosis in cfDNA molecules of <600bp based on short-read sequencing (Illumina), terminal motif features in long cfDNA molecules can also be used for cancer diagnosis. In particular, analysis of terminal motifs, including but not limited to 1-mer, 2-mer, 3-mer, 5-mer, 6-mer, 7-mer, 8-mer, 9-mer, 10-mer, 15-mer, 20-mer, or other combinations, can be used to distinguish subjects with and without cancer.

[0187] For each terminal motif of the set of terminal motifs, the relative frequency of sequences with the terminal motif in the biological sample may be determined. In some instances, the relative frequency of sequences with the terminal motif is determined based on the other frequencies of sequence motifs in a group of reference samples. The relative frequencies of sequences in the set of terminal motifs can thereby form a vector of N frequencies of biological samples, where N corresponds to the number of terminal motifs in the set of terminal motifs. The vector of N frequencies of biological samples may be compared to a plurality of reference vectors determined from a group of reference samples with known disease classification (e.g., HCC). Based on the comparison, the disease classification of the biological sample can be determined. A. Cluster-Based Analysis

[0188] Based on hierarchical clustering analysis using 5'4-mer terminal motifs of plasma DNA molecules, patients with and without HCC are often grouped into different clusters. Plasma DNA molecules may be sequenced using single molecule sequencing (e.g., SMART-seq), such that sequence reads include long cell-free DNA molecules. Terminal motifs may be identified for each sequence read, and the relative frequency of sequence reads may be identified for each type of terminal motif (e.g., CCGC). Biological samples corresponding to a disease classification have similar relative frequencies of sequence reads across different motifs and can be grouped together to form a cluster. Such similar relative frequencies may suggest that terminal motifs inferred from single molecule sequencing of plasma DNA can provide information about the presence or absence of cancer.

[0189] FIG. 11 shows a heatmap 1100 generated based on a hierarchical clustering analysis of 256 4-mer terminal motifs of plasma DNA molecules, according to some embodiments. For example, the mean and standard deviation of sequence frequency across biological samples (e.g., HCC samples, HBV carrier samples) may be determined for the terminal motif representing the row. A relative frequency of sequence reads with the terminal motif can then be generated for the biological sample, where the relative frequency may be based on the terminal motif frequency of sequence reads with the terminal motif subtracted from the calculated mean and then divided by the standard deviation. The result of the relative frequency of the terminal motif can then be shown as a color-coded value in the corresponding row of the column representing the biological sample in the heatmap 1100 (e.g., HCC04). The process may continue for other terminal motifs, whereby a color-coded value for the entire column of the heatmap can be identified for a given sample.

[0190] In some instances, z-scores are used to indicate the relative frequency of terminal motifs in sequences from sequencing cell-free DNA molecules. The z-score can be the difference between a particular terminal motif and the average frequency (e.g., across samples for that given terminal motif) divided by the variation in frequency (e.g., across samples for that given terminal motif). In FIG. 11, each row in the heatmap represented the z-score value of the frequency of a particular terminal motif across different training samples (e.g., HCC samples, HBV carrier samples). The Z-score of a particular terminal motif can be calculated using the mean and standard deviation of the particular terminal motif among the training samples. Using z-scores, the terminal motif frequencies can be visualized in different colors for a clearer comparison.

[0191] Biological samples may be grouped based on the similarity of their relative terminal motif frequencies. As shown in FIG. 11, two clusters "A" and "B" may be formed. Subgroups A and B were associated with low and high incidence of the histological condition of vascular invasion. In particular, the "A" cluster identifies 55.6% of HCC samples involved in vascular invasion, and the "B" cluster identifies 87.5% of HCC samples involved in vascular invasion. Vascular invasion refers to a disease state in which tumor cells (e.g., ctDNA) are present in the lumen of blood vessels and / or lymphatic vessels. Vascular invasion may also include extramural vascular invasion (EMVI), which involves direct invasion of blood vessels (usually veins) by the tumor. Vascular invasion may indicate a relatively severe cancer. This was determined by examining anatomic pathology reports.

[0192] Using these clusters as frequencies in a reference sample, the terminal motif frequency of a particular biological sample can be compared to the reference samples to identify disease classification.

[0193] These results suggested that it may be possible to non-invasively classify histological status based on the use of plasma DNA terminal motifs derived from single molecule sequencing. Furthermore, classification of vascular invasion may be clinically relevant for patient prognosis, especially since vascular invasion is involved in more severe forms of the corresponding disease. B. Cluster-based analysis of terminal motif frequencies in long DNA molecules

[0194] Based on hierarchical clustering analysis using 5'4-mer terminal motifs, short DNA molecules (e.g., <200bp) and long DNA molecules (e.g., >1kb) were analyzed. In some embodiments, the combined analysis of short and long DNA molecules can link a first vector containing the frequency of 256 motifs from long DNA molecules (e.g., >1kb) and a second vector containing the frequency of 256 motifs from short DNA molecules (e.g., <200bp) into a new vector with 512 dimensions. Additionally or alternatively, the combined analysis of short and long DNA molecules can be the ratio of a first vector containing the frequency of 256 motifs from long DNA molecules (e.g., >1kb) to a second vector containing the frequency of 256 motifs from short DNA molecules (e.g., <200bp). In some embodiments, short and long DNA molecules are defined with different cutoffs. For example, short DNA molecules may be defined by cutoffs including but not limited to 50bp, 60bp, 70bp, 80bp, 90bp, 100bp, 110bp, 120bp, 130bp, 140bp, 150bp, 160bp, 170bp, 180bp, 190bp, 200bp, 250bp, 300bp, 400bp, 500bp, less than 600bp, etc. Long DNA molecules may be defined by cutoffs including but not limited to greater than 600bp, 700bp, 800bp, 1kb, 2kb, 3kb, 4kb, 5kb, 10kb, 15kb, 20kb, 30kb, 40kb, 50kb, etc.

[0195] FIG. 12 shows a heat map 1200 generated using hierarchical clustering analysis of 4-mer terminal motifs of short plasma DNA (<200 bp), according to some embodiments. Additionally, FIG. 13 shows a heat map 1300 generated using hierarchical clustering analysis of 4-mer terminal motifs of long plasma DNA (>1 kb), according to some embodiments. FIG. 14 shows a heat map 1400 generated using hierarchical clustering analysis of 4-mer terminal motifs of both short plasma DNA (<200 bp) and long plasma DNA (>1 kb), according to some embodiments. FIG. 15 shows a heat map 1500 generated using hierarchical clustering analysis of 4-mer terminal motif ratios, according to some embodiments. As shown in FIGS. 12-15, each of the percentages shown in parentheses indicates the percentage of HCC patients identified from the corresponding patient group.

[0196] Compared to the analysis based on short DNA molecules (Figure 12), the ability to distinguish between HCC and non-HCC groups is improved when using long DNA molecules (>1 kb) (Figure 13) and is further strengthened in the combined analysis (Figures 14 and 15). The improvement can be demonstrated by the fact that a clearer separation between patients with and without HCC was observed from Figures 13, 14, and 15 compared to Figure 12. For example, as shown in Figure 12, when dividing the hierarchical clusters into two main groups, the percentage of HCC patients between the groups could not be distinguished when using terminal motifs derived from short cfDNA molecules (e.g., <200 bp) (62.07% vs. 61.36%) (P value: 1, Fisher's exact test). In contrast, as shown in Figure 13, when using terminal motifs derived from long cfDNA molecules (e.g., >1 kb), the percentage of HCC patients between the groups became significantly different (85.71% vs. 29.03%) (P value: 1.577 x 10 -6 , Fisher's exact test). Furthermore, when combining the terminal motifs of short and long cfDNA molecules, the percentage of HCC patients in the group was significantly higher than that in the control group (85.11% vs. 19.23%) (P value: 3.51 x 10 -8, Fisher's exact test) and Figure 15 (92.31% vs. 26.47%) (P value: 5.121 x 10 -9 The differences appeared to be more discriminative, as shown by the Fisher's exact test. C. Cluster-Based Analysis Methods

[0197] 16 shows a flowchart 1600 illustrating an example process for analyzing a subject's biological sample based on the relative frequency of sequences having one or more terminal motifs, according to some embodiments. The biological sample may contain DNA originating from normal cells and possibly cells associated with disease (e.g., cancer). In addition, at least a portion of the DNA is acellular in the biological sample.

[0198] At step 1602, sequence reads obtained from sequencing of cell-free DNA molecules may be received. For example, single molecule real-time sequencing (i.e., SMRT-seq) (e.g., by Pacific Biosciences, PacBio SMRT-seq) and nanopore sequencing (e.g., by Oxford Nanopore Technologies) may be used to obtain sequence reads from biological samples. Other sequencing technologies may also be used, for example, as described herein.

[0199] In some instances, the sequence reads correspond to long cell-free DNA molecules having a size within a first size range, which may include a lower limit and an upper limit. By way of example, the first size range may include an upper limit of at least 1,000 bases, at least 3,000 bases, or more. In some instances, the lower limit may be selected from one of at least 300 bases, at least 400 bases, at least 500 bases, at least 600 bases, or at least 800 bases.

[0200] Additionally or alternatively, a first set of sequence reads may be selected from the sequence reads. The first set of sequence reads may include sizes within a first size range. Then, a second set of cell-free DNA molecules may be selected from the sequence reads, where the second set of sequence reads may include sizes within a second size range. In some instances, the second size range has an upper limit that is greater than the upper limit of the first size range. For example, the first size range may be less than 600 bp and the second size range may be greater than 1000 bases. In some instances, the two size ranges may overlap, for example, the first size range may be less than 800 bp and the second size range may be between 700 bp and 2000 bp.

[0201] In step 1604, for each of the sequence reads, a sequence motif for each of the one or more end sequences of the corresponding cell-free DNA molecule may be identified. For example, a 4-mer end motif for a sequence read may be identified by analyzing four nucleotides at its end. Similarly, in some embodiments, other types of end motifs may be used, including but not limited to 1-mer, 2-mer, 3-mer, 5-mer, 6-mer, 7-mer, 8-mer, 9-mer, 10-mer, 15-mer, 20-mer, or other combinations.

[0202] At step 1606, for each of the set of N sequence motifs, a relative frequency of the sequence motif may be determined. The N relative frequencies may be determined. The relative frequency of a terminal motif may be the percentage of DNA molecules that have that particular terminal motif. As another example, the relative frequency may be a rank of the sequence motif, e.g., a rank of the raw number of DNA molecules (fragments) that have the terminal motif. In some instances, the normalized frequency is a z-score, e.g., as described above with respect to FIG. 11. By way of example, N may be an integer equal to 2, 3, 4, 5, 8, 10, 15, 16, 20, 50, 64, 100, 128, 200, 256, or more, depending on, e.g., the k-mer size of the terminal motif used.

[0203] In some instances, the relative frequency can be determined based on the proportion of cell-free DNA molecules with end sequences corresponding to the sequence motif relative to the cell-free DNA molecules from the biological sample.Alternatively, the relative frequency can be determined based on the proportion of cell-free DNA molecules with end sequences corresponding to the sequence motif relative to the number of cell-free DNA molecules with end sequences corresponding to other sequence motifs of a set of N sequence motifs.

[0204] In step 1608, the N relative frequencies may be used to generate a vector of N frequencies corresponding to the set of N sequence motifs. Each of the N frequencies in the vector may be normalized to each other (e.g., by rank) or to other frequencies of the sequence motif in the group of reference samples (e.g., as described above for z-score). Normalization of each frequency in the group of reference samples may also be performed using rank. For example, the vector of N frequencies may be generated by normalizing the relative frequency of the sequence motif using other frequencies of the sequence motif in the group of reference samples. In some instances, each frequency in the vector of N frequencies is determined by comparing the relative frequency to a representative frequency of the sequence motif in the group of reference samples to, for example, determine a z-score.

[0205] In some instances, a vector of N frequencies may be generated based on (i) a first vector corresponding to the N relative frequencies of a first set of sequence reads within a first size range, and (ii) a second vector corresponding to the N relative frequencies of sequence reads in a second size range. Thus, the vector of N frequencies may be a value that identifies a correlation between short DNA molecules (e.g., the first set of sequence reads) and long DNA molecules (e.g., the second set of sequence reads).

[0206] In step 1610, the N frequency vectors may be compared to a number of reference vectors determined from a group of reference samples with known disease classification. The comparison may include determining the distance between the vector and the reference vector. By way of example, the reference vector may be one of the reference samples or one representative of a group (cluster) of reference samples, e.g., a statistic (median, mean, or centroid) of the vectors of the group of reference samples.

[0207] At step 1612, a disease classification of the biological sample may be identified based on the comparison. In some instances, the classification may be identified using hierarchical clustering and / or heat map clustering. Other machine learning techniques, such as neural networks, decision trees, and support vector machines, may also be used.

[0208] In some instances, identifying the classification of the disease includes identifying a classification associated with a cluster of reference vectors that is closest to the N frequency vector. For example, a first distance between the N frequency vector and a closest reference vector of a first cluster of reference vectors of the set of clusters may be identified. The first cluster of reference vectors represents a first subgroup of the group of reference samples classified as having the disease. A second distance between the N frequency vector and a closest reference vector of a second cluster of reference vectors of the set of clusters may also be identified. The second cluster of reference vectors represents a second subgroup of the group of reference samples classified as not having the disease. The first and second distances may then be compared. If the first distance is greater than the second distance, the subject may be identified as not having the disease. If the first distance is less than the second distance, the subject may be identified as having the disease. D. Ranking and separation values ​​of short and long cfDNA molecules

[0209] In addition to using the relative frequency of various terminal motifs to identify disease classification of biological samples, the frequency of sequences with a particular terminal motif may be identified for each size range of plasma DNA molecules of the biological sample. In some instances, the relative frequency of a terminal motif is identified based on the number of sequences with a terminal motif compared to the number of sequences with other terminal motifs that can be found in plasma DNA molecules. Additionally or alternatively, the relative frequency can be the percentage of sequences with a terminal motif relative to the total sequences of plasma DNA molecules. The frequency of sequences in each size range of DNA molecules can be used to identify a separation value. The separation value can then be used to identify a disease classification.

[0210] FIG. 17 shows a set of graphs 1700 identifying the motif ranking relationship between short plasma DNA molecules (<600 bp) and long plasma DNA molecules (>1 kb). In FIG. 17, each circle in the graphs represents a 4-mer terminal motif. Graph "A" identifies the motif ranking of subjects with chronic HBV infection, and graph "B" identifies the motif ranking of subjects with HCC. In particular, in graph 1702, the ranking of 256 terminal motifs of plasma DNA molecules (<600 bp) for patients with chronic HBV infection is plotted against their counterparts of long plasma DNA molecules (>1 kb). The pink region 1806A in graph 1702 was ranked in the top 10 for plasma DNA molecules, but only 11 for long plasma DNA molecules. th The ranked motifs are identified below. Conversely, the yellow region 1808A ranked in the top 10 for long plasma DNA molecules, but 11 for short plasma DNA molecules. th The ranked motifs are highlighted below. The motif patterns between short and long DNA molecules were found to be different. For example, the ranks of GCTT, ACTT, and GTTT were increased in long plasma DNA compared to short plasma DNA, whereas the ranks of CCAG, CCTG, and CCAA were decreased.

[0211] For patients with HCC identified in graph 1704, the relative frequencies reflected in the rank order of the 256 terminal motifs in plasma DNA differed from the relative frequencies of the terminal motifs in patients with chronic HBV infection (graph 1702). In particular, graphs 1702 and 1704 show the terminal motif frequencies of plasma DNA molecules. Similar to graph 1702, the pink region 1806B in graph 1704 ranked in the top 10 for plasma DNA molecules, but was ranked in the top 11 for long plasma DNA molecules. th The following motifs are ranked: Conversely, the yellow region 1808B ranked in the top 10 for long plasma DNA molecules, but 11 for short plasma DNA molecules. th The ranked motifs are highlighted below. For example, the motif rankings for CCAG, CCTG, and CCAA motifs in plasma DNA of patients with HCC were different from those with HBV infection. Thus, these data suggest that analysis of terminal motifs between short and long plasma DNA molecules can provide valuable information for cancer detection. For example, the CCAG separation value of a biological sample may be compared with a cutoff value to identify disease classification. However, such analysis involving plasma DNA molecules with sizes >1 kb could not be obtained with short-read sequencing technologies (e.g., Illumina sequencing platform) due to the lack of ability to sequence long DNA molecules, e.g., DNA molecules >600 bp. E. Terminal motif frequency of long DNA molecules

[0212] In some embodiments, the terminal motif analysis corresponds to the analysis of one particular 4-mer terminal motif. For example, the terminal motif frequency of CCCA was calculated for short plasma DNA molecules of <200 bp, long plasma DNA molecules of >600 bp, and long plasma DNA molecules of >1 kb. 1.Hepatocellular carcinoma

[0213] FIG. 18 shows a boxplot 1800 identifying the CCCA terminal motif frequency in plasma DNA molecules of <200 bp in HCC and non-HCC subjects. As shown in FIG. 18, the reduction in CCCA terminal motifs in short cfDNA molecules in the HCC group is consistent with previous findings revealed by the Illumina platform (Jiang et al. Cancer Discov. 2020; 10:664-673), where a reduction in HCC was observed. The statistically significant reduction in CCCA motif frequency in HCC overlaps with other classifications. We explored whether using longer DNA fragments would produce better results.

[0214] FIG. 19 shows a pair of box plots 1900 identifying CCCA motif frequencies in plasma DNA molecules. Box plot 1902 shows CCCA frequencies in plasma DNA molecules longer than 600 bp in HCC and non-HCC subjects, and box plot 1904 shows CCCA frequencies in plasma DNA molecules >1 kb in HCC and non-HCC subjects. In contrast to FIG. 18, when long DNA molecules were identified and analyzed using SMRT sequencing in the cohort of the present disclosure, it was surprisingly discovered that a higher (but not a lower) motif frequency of CCCA was observed in long cfDNA molecules in HCC patients compared to non-HCC subjects. In addition, the separation of HCC from other classifications was greater for long cfDNA molecules than for short cfDNA molecules.

[0215] To evaluate the terminal motif analysis using long DNA molecules, Figure 20 shows the ROC curve 2000 that identifies the performance of CCCA motif frequency in short DNA molecules 2002 and long DNA molecules 2004 in distinguishing HCC and non-HCC subjects. As shown in Figure 20, the AUC based on the terminal motif of short cfDNA is 0.69, in contrast to the AUC of long cfDNA is 0.88 (P value: 0.0065, bootstrap test). The use of the terminal motif CCCA derived from long cfDNA molecules has significantly higher ability to distinguish HCC patients from non-HCC patients compared to the use of short cfDNA molecules.

[0216] In some embodiments, information from both long and short DNA molecules is integrated into one analytical model to enhance the ability of disease classification. FIG. 21 shows a box plot 2100 identifying CCCA rates in HCC patients, HBV carriers, and healthy subjects. CCCA rates were calculated by dividing the CCCA motif frequency of long DNA molecules (>1 kb) by the CCCA motif frequency of short DNA molecules (<200 bp) in HCC patients, HBV carriers, and healthy subjects. HCC patients displayed a significantly higher CCCA rate than non-HCC subjects (P value: 3.919 x 10 -10 , Mann-Whitney U test). The ability to distinguish HCC from non-HCC subjects was significantly enhanced when using CCCA percentages compared to either short (FIG. 18) or long (FIG. 19) DNA molecules.

[0217] As an example, Figure 22 shows a ROC curve 2200 that identifies the performance of CCCA ratio in distinguishing subjects with HCC from those without HCC. Figure 22 shows a long vs. short CCCA ratio analysis with an AUC of 0.9. In some embodiments, different terminal motif ratios (e.g., CCCT, CCCC, CCCG, TTTA) may be used for cancer detection, or multiple terminal motif ratios may be used together for cancer detection. These results are surprising because short-read sequencing technology cannot analyze long DNA molecules (e.g., >600bp), and such terminal motif ratio analysis was considered to be infeasible in conventional tests. 2.Colorectal cancer

[0218] SMRT sequencing was performed on plasma DNA molecules from 4 colorectal cancer (CRC) patients and 15 healthy subjects. Figure 23 shows a boxplot 2300 that identifies the CCCA terminal motif frequency in plasma DNA molecules of <200bp in CRC patients and healthy subjects. Plasma DNA molecules were sequenced using SMRT-sequencing. Figure 23 shows that the CCCA frequency is significantly reduced in CRC patients compared to healthy subjects (P value: <0.01, Mann-Whitney U test).

[0219] Figure 24 shows a box plot 2400 identifying CCCA motif frequency in plasma DNA molecules longer than 1 kb in CRC patients and healthy subjects. Plasma DNA molecules were sequenced using SMRT sequencing. Figure 24 shows that CCCA frequency is significantly elevated in CRC patients when compared with healthy subjects (P value: 0.01, Mann-Whitney U test). Such an increase in CCCA frequency demonstrates that the long cfDNA terminal motif feature can be applied to the detection of multiple cancer types, not limited to colorectal cancer and hepatocellular carcinoma, as presented in this disclosure. This is also surprising, since conventional sequencing methods (e.g., Illumina sequencing) cannot identify long DNA molecules (e.g., plasma DNA molecules with a size larger than 600 bp).

[0220] In some embodiments, information from both long and short DNA molecules is integrated into one analysis model to enhance the ability of disease classification in colorectal cancer patients. Figure 25 shows a box plot 2500 that identifies the CCCA rate in CRC patients and healthy subjects during SMRT sequencing. As shown in the box plot 2500, the CCCA rate is calculated by dividing the CCCA motif frequency of long DNA molecules (>1 kb) by the CCCA motif frequency of short DNA molecules (<200 bp) in CRC patients and healthy subjects. CRC patients displayed significantly higher CCCA rate than healthy subjects (P value: 0.004, Mann-Whitney U test). F. Terminal motif analysis by nanopore sequencing

[0221] In some embodiments, nanopore sequencing by Oxford Nanopore Technologies (ONT) is used in nucleic acid terminal motif analysis. To demonstrate the effectiveness of nanopore sequencing, plasma DNA molecules from eight HCC patients and six HBV carriers were sequenced by nanopore sequencing.

[0222] Figure 26 shows a boxplot 2600 identifying the CCCA terminal motif frequency in plasma DNA molecules <200bp in HCC patients and HBV carriers. As shown in Figure 26, the decline in CCCA terminal motifs in short cfDNA molecules in the HCC group was consistent with previous findings revealed by the Illumina platform (Jiang et al. Cancer Discov. 2020; 10:664-673). However, the utility of long cfDNA molecules in terminal motif analysis was unknown, since long DNA molecules were barely detectable by the Illumina sequencing platform (0% of the molecules were >600bp).

[0223] FIG. 27 shows a pair of box plots 2700 identifying motif frequencies of CCCA in plasma DNA molecules. As shown in FIG. 27, box plot 2702 shows CCCA frequency of plasma DNA molecules longer than 600 bp in HCC and HBV carriers, and box plot 2704 shows CCCA frequency of plasma DNA molecules longer than 1 kb in HCC and HBV carriers. In contrast to FIG. 26, it was found that when long DNA molecules were identified and analyzed using nanopore sequencing in the cohort of the present disclosure, a higher motif frequency of CCCA in long cfDNA molecules was observed in HCC patients compared to non-HCC subjects, which is consistent with the data of the present disclosure generated from the SMRT sequencing platform described in the present disclosure.

[0224] In some embodiments, information from both long and short DNA molecules is integrated into one analysis model to enhance the ability of disease classification in nanopore sequencing. Figure 28 shows a box plot 2800 identifying CCCA rates in HCC patients and HBV carriers during nanopore sequencing. CCCA rates were calculated by dividing the CCCA motif frequency of long DNA molecules (>1 kb) by the CCCA motif frequency of short DNA molecules (<200 bp) in HCC patients and HBV carriers. As shown in Figure 28, HCC patients displayed significantly higher CCCA rates than HBV carriers (P value: 0.013, Mann-Whitney U test). These findings are consistent with the data of the present disclosure generated by SMRT sequencing presented in the embodiments of the present disclosure, highlighting the diagnostic potential of long cfDNA fragment features across multiple platforms. G. Machine learning techniques for disease classification based on terminal motif frequencies of long cell-free DNA molecules

[0225] In some embodiments, the terminal motif analysis is carried out using a machine learning model that can extract useful information from the terminal motif signature to classify patients with and without cancer.The machine learning model can include, but is not limited to, convolutional neural network (CNN), linear regression, logistic regression, deep recurrent neural network (e.g., fully connected recurrent neural network (RNN), gated recurrent unit (GRU), long short-term memory, (LSTM)), transformer-based methods (e.g., XLNet, BERT, XLM, RoBERTa), Bayesian classifier, hidden Markov model (HMM), linear discriminant analysis (LDA), K-means clustering, density-based spatial clustering of applications with noise (DBSCAN), random forest algorithm, adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), support vector machine (SVM), or a composite model that includes one or more of the above models. 1. Logistic regression

[0226] In some embodiments, logistic regression analysis is used to evaluate the discrimination ability of using 4-mer terminal motifs to classify HCC and non-HCC subjects. In a logistic regression (LR) model, it is possible to establish a relationship between binary outcome variables. For example, binary outcomes 0 and 1 indicate cancer status of non-cancer and cancer, respectively. LR models the logit-transformed probability with predictor variables as a linear relationship, where the predictor variables include terminal motifs derived from cfDNA molecules of different size ranges. For example, let Y be a binary outcome variable in {0,1} indicating non-cancer and cancer, and p be the probability that y is 1, where p=P(Y=1). The higher p, the more likely the patient has cancer. x 1 , x 2 , …, x k Let x be a set of predictor variables. In one example, the set of predictor variables can be the frequency of 256 5'4-mer end motifs in cfDNA molecules >1 kb. 1 , x 2 , …, x k Logistic regression of Y on β is performed using the maximum likelihood method of the following equation: 0、 β 1 , …, β k The parameter values ​​can be derived.

number

number

[0227] Figure 29 shows a box plot 2900 identifying the results generated by logistic regression analysis of terminal motif features in short DNA molecules with a size of less than 200 bp. Figure 29 shows a logistic regression analysis using short cfDNA molecules of <200 bp. As shown in Figure 29, HCC patients are more likely to be classified as having cancer than control subjects. Figure 30 shows a ROC curve 3000 identifying the performance of logistic regression using terminal motif features in short DNA molecules (<200 bp) in distinguishing subjects with HCC from those without HCC. Figure 30 shows a logistic regression analysis of terminal motifs of short DNA molecules with an AUC of 0.89.

[0228] In addition to short DNA molecules, logistic regression analysis may be extended to long DNA molecules. FIG. 31 shows a box plot 3100 identifying results generated from logistic regression analysis of terminal motif features on long DNA molecules larger than 1 kb in size. FIG. 32 shows a ROC curve 3200 identifying the performance of logistic regression using terminal motif features of long DNA molecules (>1 kb) in distinguishing subjects with HCC from those without HCC. When long cfDNA molecules of >1000 bp are used in logistic regression analysis, HCC patients show a higher probability than healthy and HBV carrier subjects, as compared to the results of FIG. 29. Furthermore, the accuracy of HCC classification reached an AUC of 0.9 as shown in FIG. 32. Using a probability score cutoff to allow for 99% specificity, the use of short DNA molecules in terminal motif analysis could only provide a sensitivity of 42%, whereas the use of long cfDNA molecules in terminal motif analysis improves the sensitivity to 70%.

[0229] The above data suggested that the use of long cfDNA molecules as implemented in the present disclosure can enhance diagnostic performance. Since a study using Illumina short-read sequencing technology revealed that short cfDNA molecules were enriched in the tumor DNA fraction of plasma from cancer patients (Jiang et al. Proc Natl Acad Sci USA. 2015;112:E1317-25), many studies have attempted to focus on the analysis of short cfDNA molecules to improve cancer detection performance (Underhill et al. PloS Genet. 2016;12:e1006162, Mouliere et al. Sci Transl Med. 2018;10:eaat4921, Liu et al. Transl Lung Cancer Res. 2021;10:1501-1511). For example, one study attempted to remove long cfDNA molecules using in vitro size selection with a benchtop microfluidic device prior to sequencing (Mouliere et al. Sci Transl Med. 2018;10:eaat4921). Thus, the inclusion of long cfDNA molecules to enhance diagnostic performance has not been explored before, and the improvement in diagnostic performance with the use of long cfDNA molecules is surprising.

[0230] The terminal motif features of short and long DNA molecules can be used together to further enhance the performance of cancer classification using logistic regression. FIG. 33 shows a box plot 3300 identifying logistic regression analysis using terminal motif features in both long DNA molecules >1 kb and short DNA molecules <200 bp. As shown in FIG. 33, when terminal motif information from both long DNA molecules (>1 kb) and short DNA molecules (<200 bp) are integrated together in logistic regression analysis, HCC subjects can be clearly distinguished from healthy subjects and subjects with HBV. FIG. 34 shows a ROC curve 3400 identifying the performance of logistic regression using a combination of terminal motif features from both long DNA molecules (>1 kb) and short DNA molecules (<200 bp) in distinguishing subjects with HCC from those without HCC. As shown in FIG. 34, the diagnostic ability between HCC and non-HCC subjects was further enhanced to an AUC of 0.92. In some embodiments, patterns of terminal motifs from three or more size ranges may be used. As an example, the frequencies of 256 motifs from molecules in the size range <200 bp, 256 motifs from molecules in the size range 200-600 bp, and 256 motifs from molecules in the size range >600 bp may be combined together in a logistic regression analysis (number of 4-mer features: 256 x 3) with an AUC of 0.93 for distinguishing patients with HCC from those without cancer.

[0231] In some instances, the motif ratio calculated by dividing the motif frequency in long DNA molecules (>1 kb) by the motif frequency in short DNA molecules (<200 bp) is used in logistic regression. Figure 35 shows a box plot 3500 identifying the results generated by logistic regression analysis using the motif ratio. When such motif ratios are used in logistic regression analysis, the generated probabilities for HCC subjects are significantly higher than for healthy and HBV carrier subjects. Figure 36 shows a ROC curve 3600 identifying the performance of logistic regression using the motif ratio in distinguishing subjects with HCC from those without HCC. As shown in Figure 36, the AUC is further improved to 0.97, reflecting that the terminal motif information from both short and long cfDNA molecules can be synergistically utilized to enhance the diagnostic potential for cancer. 2. Support Vector Machine

[0232] In some embodiments, support vector machine (SVM) analysis is used to classify cancer and non-cancer subjects based on 4-mer terminal motifs. Suppose a training dataset for establishing an SVM classifier contains n samples: (M 1 ,Y 1 ), …, (M n ,Y n )(1) In the formula, Y i is 1 (indicating a cancer subject) or −1 (indicating a non-cancer subject) for sample i, and M i is a p-dimensional vector containing the terminal motif pattern for sample i. For example, M i may be a vector containing 256 4-mer terminal motifs. ican be a vector containing values ​​derived from 256 4-mer terminal motifs, such as the ratio of long to short cfDNA molecules. The SVM can be trained using a training dataset to determine a "hyperplane" that separates the non-cancer and cancer groups as accurately as possible. There are various ways to find such a hyperplane. One is to find a set of coefficients (W with p-dimensional vector) that satisfy the following equation:

number

number

number

[0233] Therefore, it is subject to (4). JPEG2024545610000008.jpg59. Based on this principle, the parameters (W and b) of the classifier may be determined. The cancer risk score of a new sample can be calculated using the parameters (W and b) trained in this example.

[0234] FIG. 37 shows a ROC curve 3700 identifying the performance of SVM using terminal motif ratios in distinguishing subjects with HCC from those without HCC. SVM was used to classify a given subject's biological sample using 256 terminal motif ratios, where each terminal motif ratio corresponded to the frequency ratio for each terminal motif (e.g., CCCA) between long and short DNA molecules. As shown in FIG. 37, when terminal motif information from both long DNA molecules (>1 kb) and short DNA molecules (<200 bp) were integrated together in the SVM analysis, the diagnostic ability between HCC and non-HCC subjects reached an AUC of 0.93. In some embodiments, patterns of terminal motifs from three or more size ranges may be used. As an example, the frequencies of 256 motifs from molecules with a size range of <200 bp, 256 motifs from molecules with a size range of 200-600 bp, and 256 motifs from molecules with a size range of >600 bp may be integrated together in a logistic regression analysis. 3. Random Forest

[0235] In some embodiments, random forest tree analysis may be performed using 4-mer terminal motifs to classify HCC subjects and non-HCC subjects. Figure 38 shows a ROC curve 3800 that identifies the performance of random forest analysis using motif ratios in distinguishing subjects with HCC from those without HCC. Random forest tree was used to classify a given subject's biological sample using 256 terminal motif ratios, where each terminal motif ratio corresponded to the frequency ratio for each terminal motif (e.g., CCCA) between long and short DNA molecules. As shown in Figure 38, when terminal motif information from both long DNA molecules (>1 kb) and short DNA molecules (<200 bp) was integrated into random forest tree analysis, the diagnostic ability between HCC subjects and non-HCC subjects reached an AUC of 0.94. 4. Linear Discriminant Analysis

[0236] In some embodiments, linear discriminant analysis (LDA) may be performed using 4-mer terminal motifs to classify HCC subjects and non-HCC subjects. Figure 39 shows a ROC curve 3900 that identifies the performance of LDA analysis using motif ratios in distinguishing subjects with HCC from those without HCC. Linear discriminant analysis was used to classify a given subject's biological sample using 256 terminal motif ratios, where each terminal motif ratio corresponded to the frequency ratio for each terminal motif (e.g., CCCA) between long and short DNA molecules. As shown in Figure 39, when terminal motif information from both long DNA molecules (>1 kb) and short DNA molecules (<200 bp) was integrated into LDA analysis, the diagnostic ability between HCC subjects and non-HCC subjects reached an AUC of 0.97. H. Methods for Analysing Short and Long DNA Molecules

[0237] Figure 40 shows a flow chart 4000 illustrating an example process for analyzing a subject's biological sample based on the relative frequency of sequences with one or more terminal motifs, according to some embodiments. The biological sample may contain DNA originating from normal cells and possibly cells associated with disease (e.g., cancer). In addition, at least a portion of the DNA is acellular in the biological sample.

[0238] In step 4002, sequence reads obtained from sequencing of cell-free DNA molecules may be received. For example, sequence reads may be obtained from biological samples using single molecule real-time sequencing (i.e., SMRT-seq) (e.g., by Pacific Biosciences, PacBio SMRT-seq) and nanopore sequencing (e.g., by Oxford Nanopore Technologies). Other sequencing technologies may also be used, for example, as described herein.

[0239] In step 4004, the sequence reads may be used to determine the size of the cell-free DNA molecules. For example, the number of nucleotides may be counted to determine the size of the cell-free DNA molecules. Other techniques may also be used, for example, paired-end sequencing may be used to align pairs of sequence reads to a reference genome.

[0240] In step 4006, for each of the sequence reads, a sequence motif may be identified for each of the one or more end sequences of the corresponding cell-free DNA molecule. For example, a 4-mer end motif of a sequence read may be identified by analyzing four nucleotides at its end. Continuing with this example, a first sequence read may include CCCA as a sequence motif and a second sequence read may include CCAG as a sequence motif. Similarly, in some embodiments, other types of end motifs may be used, including but not limited to 1-mer, 2-mer, 3-mer, 5-mer, 6-mer, 7-mer, 8-mer, 9-mer, 10-mer, 15-mer, 20-mer, or other combinations.

[0241] In step 4008, for a first set of cell-free DNA molecules having a first size range, a first relative frequency of occurrence of one or more sequence motifs within the first set of cell-free DNA molecules may be determined. The relative frequency may be a rank order of the sequence motifs. As another example, the relative frequency may be a percentage of DNA molecules having a particular sequence motif.

[0242] In some instances, the first relative frequency is a ratio of the first set of cell-free DNA molecules to the number of cell-free DNA molecules from the biological sample. Additionally or alternatively, the first relative frequency is a ratio of the first set of cell-free DNA molecules to the number of cell-free DNA molecules having other sequence motifs. In some instances, the first size range includes an upper limit selected from one of at least 80 bases, at least 100 bases, at least 150 bases, at least 200 bases, or at least 300 bases. For example, the first size range can be 1-200 bp.

[0243] In step 4010, for a second set of cell-free DNA molecules having a second size range, a second relative frequency of occurrence of the one or more sequence motifs in the second set of cell-free DNA molecules may be determined. In some instances, the second relative frequency can be a ratio of the second set of cell-free DNA molecules to the number of cell-free DNA molecules from the biological sample. Additionally or alternatively, the second relative frequency is a ratio of the second set of cell-free DNA molecules to the number of cell-free DNA molecules having other sequence motifs.

[0244] In some instances, the second size range has an upper limit that is greater than the upper limit of the first size range. For example, the first size range may be less than 600 bp and the second size range may be greater than 1000 bases. In some instances, the two size ranges may overlap, for example, the first size range may be less than 800 bp and the second size range may be between 700 bp and 2000 bp. Additionally or alternatively, the second size range includes a lower limit selected from one of at least 300 bases, at least 400 bases, at least 500 bases, at least 600 bases, or at least 800 bases. In some instances, the lower limit of the second size range is greater than the upper limit of the first size range.

[0245] In step 4012, a separation value of the first relative frequency and the second relative frequency may be determined. In some instances, the separation value is a ratio of the first relative frequency and the second relative frequency, or a ratio of each function of the frequencies. For example, the separation value may be a ratio of a second set of cell-free DNA molecules (e.g., long DNA molecules) to a first set of cell-free DNA molecules (e.g., short DNA molecules), where the first and second sets have end sequences corresponding to CCCA. In other instances, the separation value may include not only the subtraction of two frequencies, but also a combination of functions that provide a measure of separation between the frequencies. Identifying the separation value is further described in Sections III.D and III.E of this disclosure.

[0246] In step 4014, the separation value may be used to identify a disease classification of the biological sample. In some instances, the classification is identified by comparing the separation value to one or more cutoff values. The disease may be cancer (e.g., HCC, CRC) and the classification may include multiple stages of the cancer. By way of example, the cancer may be hepatocellular carcinoma, lung cancer, breast cancer, gastric cancer, glioblastoma multiforme, pancreatic cancer, colorectal cancer, nasopharyngeal carcinoma, and head and neck squamous cell carcinoma. In some instances, the disease classification identifies a classification of the severity of the disease. The identification of the disease classification may include the histological state of the cancer, for example, the presence or absence of vascular invasion. The one or more cutoff values ​​may be determined from a reference sample (e.g., a healthy sample, a sample from a subject classified as having a disease) with a known disease classification. In some instances, the cutoff value of the one or more cutoff values ​​may be selected from one of 0.6, 0.65, 0.7, or 0.75.

[0247] As described above, the cutoff value or values ​​may be determined using machine learning with training samples with known disease classifications (e.g., those shown in FIG. 17). In another example, the comparison to the cutoff value or values ​​may be performed using a machine learning model. The machine learning model may be applied to the separation values ​​to generate a disease classification. The machine learning models may include, but are not limited to, convolutional neural networks (CNNs), linear regression, logistic regression, deep recurrent neural networks (e.g., fully connected recurrent neural networks (RNNs), gated recurrent units (GRUs), long short-term memory, (LSTM)), transformer-based methods (e.g., XLNet, BERT, XLM, RoBERTa), Bayesian classifiers, hidden Markov models (HMMs), linear discriminant analysis (LDA), K-means clustering, density-based spatial clustering of applications with noise (DBSCAN), random forest algorithms, adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), support vector machines (SVMs), or hybrid models including one or more of the above models. IV. Methylation pattern analysis of long cell-free DNA molecules for disease classification

[0248] Since tumor formation can occur in any tissue of the human body, identifying the tissue of origin of a single molecule is useful for cancer testing and guiding cancer treatment. One approach that may be based on the hypothesis that the target tissue of origin of plasma DNA corresponds to the fewest mismatches in terms of methylation status across CpG sites between the plasma DNA methylation haplotype and the methylation haplotype of that tissue (i.e., methylation mismatches) is herein named the minimum methylation mismatch approach. The number of methylation mismatches may be identified by pairwise comparison of methylation status across CpG sites between two methylation haplotypes originating from the same genomic location. If two methylation statuses between two methylation haplotypes at the same CpG location are different, it is counted as one mismatch.

[0249] In some embodiments, methylation haplotypes of long DNA molecules obtained from tissue and plasma DNA molecules are used to enhance accuracy, as long methylation haplotypes are more likely to contain informative methylation patterns specific to a particular tissue compared to short methylation haplotypes.

[0250] FIG. 41 shows an exemplary diagram 4100 comparing methylation patterns of long cell-free DNA molecules with methylation patterns of reference tissues, according to some embodiments. In particular, FIG. 41 shows that a short methylation haplotype of plasma DNA with 3 CpG sites does not identify which tissue contributes to such plasma DNA molecules (e.g., liver, brain and lung tissues share the same short methylation haplotype). In contrast, a long methylation haplotype of plasma DNA with 10 CpG sites clearly identifies the liver as the tissue of origin of such plasma DNA molecules, since the methylation haplotype from liver shows the fewest methylation mismatches with 0, compared to the methylation haplotypes of plasma DNA from brain, lung, colon and white blood cells showing 2, 3, 4 and 5 methylation mismatches, respectively.

[0251] In some embodiments, pattern recognition analysis of methylation haplotypes improves the ability to identify the tissue of origin of each long plasma DNA molecule, which can then be used to identify disease classification.

[0252] FIG. 42 illustrates a technique 4200 for analyzing methylation patterns in long cell-free DNA molecules containing at least one methylation mismatch, according to some embodiments. As shown in FIG. 42, identifying the tissue of origin of plasma DNA molecules may be challenging for the minimal methylation mismatch approach. In particular, a given plasma DNA molecule has a methylation mismatch at site "2" when compared to the methylation haplotype in tumor cells, but also has a methylation mismatch at site "5" when compared to the methylation haplotype of non-tumor cells (e.g., buffy coat). Pattern recognition analysis can address this challenge. For example, it can be determined that the presence of three consecutive unmethylated CpG sites at positions 4, 5, and 6 indicates a higher likelihood of HCC. Based on this information, the CpG sites at positions 4, 5, and 6 are given a higher weight indicative of a tumor pattern compared to CpG sites at other positions. In light of this weight, a given plasma DNA molecule can be predicted to be associated with tumor cells based on its unmethylated CpG sites at positions 4, 5, and 6. These types of pattern analyses become more effective when a given plasma DNA molecule exceeds a certain length.

[0253] In some embodiments, the length of plasma DNA may include, but is not limited to, ≧500 bp, ≧600 bp, ≧1 kb, ≧2 kb, ≧3 kb, ≧4 kb, ≧5 kb, ≧10 kb, or other combinations. The number of CpG sites may include, but is not limited to, ≧3, ≧4, ≧5, ≧6, ≧7, ≧8, ≧9, ≧10, ≧15, ≧20, ≧25, ≧30, ≧35, ≧40, ≧45, ≧50, ≧60, ≧70, ≧80, ≧90, ≧100, ≧200, ≧300, ≧400, ≧500, ≧1000, or other combinations. In some embodiments, the methylation haplotypes of long DNA molecules from various tissues and tumor tissues are identified by methylation recognition enzyme conversion. One example of such a conversion method is Methyl-seq (EM-seq), which involves a non-destructive enzymatic reaction that utilizes TET2 and APOBEC3A to convert unmethylated (but not methylated) cytosines to uracils that are sequenced as thymines (e.g., NEBNext® Enzymatic Methyl-seq Kit). Traditional bisulfite sequencing has drawbacks when harvesting long DNA molecules, as it degrades the long DNA molecules, resulting in shorter methylation haplotype information and negatively impacting the accuracy of tissue of origin identification of plasma DNA. A. CpG sites in long cell-free DNA molecules

[0254] FIG. 43 shows a comparison 4300 of the prevalence of CpG sites and cancer-derived single nucleotide variants (SNVs) across the genome at 1 kb resolution. In particular, Table A shows the number of 1 kb genomic regions of a given genome (e.g., a reference genome) that have at least a corresponding number of CpG sites (e.g., >1). Table B shows the number of 1 kb genomic regions of a genome that have at least a corresponding number of SNVs (e.g., >2). As shown in FIG. 43, there are 971,880 1 kb regions that contain at least 10 CpG sites, accounting for 33.7% of the human genome, whereas there are only two 1 kb regions that contain at least 10 SNVs when analyzing 38,465 somatic mutations from tumor tissues. Thus, plasma DNA molecules with abundant CpG sites can be obtained in sufficient quantities, and their methylation patterns can be used to predict disease.

[0255] Figure 44 shows a comparison 4400 of the prevalence of CpG sites and cancer-derived SNVs across the genome at 3 kb resolution. Table A shows the number of 3 kb genomic regions of a given genome (e.g., a reference genome) that have at least a corresponding number of CpG sites (e.g., >1). Table B shows the number of 3 kb genomic regions of a genome that have at least a corresponding number of SNVs (e.g., >2). As shown in Figure 44, there are 844,742 3 kb regions that contain at least 10 CpG sites, accounting for 88.0% of the human genome, while there are only two 3 kb regions that contain at least 10 SNVs. These results suggested that methylation haplotype-based analysis of long plasma DNA molecules can significantly improve the information of plasma DNA molecules that can be used to provide information about the presence of cancer.

[0256] In contrast, the amount of CpG sites in short cell-free DNA molecules may not be sufficient for disease classification. Figure 45 shows a comparison of genome-wide CpG sites and cancer-derived SNV prevalence at 200bp resolution 4500. Table A shows the number of 200bp genomic regions in a given genome (e.g., reference genome) that have at least a corresponding number of CpG sites (e.g., >1). Table B shows the number of 200bp genomic regions in a genome that have at least a corresponding number of SNVs (e.g., >2). As shown in Figure 45, the percentage of 200bp regions containing 10 CpG sites quickly drops to as low as 1.9%. This result suggests that the limited number of CpG sites (e.g., <200bp) present in short cell-free DNA molecules adversely affects tissue-of-origin analysis or disease classification based on plasma DNA.

[0257] Thus, the methylation pattern of several CpG sites in a long cell-free DNA molecule can be used to identify one or more biomarkers that can predict the presence of a disease (e.g., cancer). For example, sequence reads corresponding to long cell-free DNA molecules of a plasma sample can be obtained using methylation-aware sequencing (e.g., Enzymatic Methyl-seq). Each sequence read can include a methylation pattern that identifies the methylation status at a set of CpG sites on the sequence read. The methylation pattern of each sequence read can be compared to a reference methylation pattern of a tissue type, thereby identifying a tissue classification of the sequence read. The tissue classification of the sequence read can then be used to identify a disease classification.

[0258] For example, because the pattern specificity of long cell-free DNA molecules is much higher, even if the resolution of the reference tissue methylome is low, methylation analysis can be performed on a molecule-by-molecule basis. For example, even if the reference tissue methylome does not cover all bases of the genome or contains fragmented sections of information, there may be fewer matches with long cell-free DNA molecules in the whole genome. In fact, long cell-free DNA molecules can be aligned to their true matches despite the ambiguity of the reference. As long as a reference methylation pattern exists, the comparison of methylation patterns can be used to predict whether the proportion of long cell-free DNA molecules corresponds to a specific tissue type. A high proportion of long cell-free DNA molecules associated with a specific tissue type in a plasma sample can predict cancer. B. Tissue of origin

[0259] The presence of long plasma DNA molecules in patients with HCC, with a series of CpG sites with distinct methylation patterns (i.e., methylation haplotype information), facilitates tracing their tissue / tumor of origin at the single molecule level.

[0260] FIG. 46 shows a schematic diagram 4600 illustrating an example process of predicting whether a cell-free DNA molecule corresponds to tumor DNA depending on its methylation haplotype information. In FIG. 46, plasma DNA molecules were subjected to SMRT sequencing. The methylation status across CpG sites of each sequenced read was derived using the HK model (Tse et al. Proc Natl Acad Sci USA. 2021;118: e2019768118). To demonstrate the feasibility of tracing the tissue of origin of plasma DNA by its methylation haplotype, the distance in terms of methylation status between plasma DNA molecules (methylation status across CpG sites (0 / 1)) and the aggregate methylation index corresponding to the CpG sites (continuous values ​​ranging from 0 to 1, respectively) in each reference tissue, e.g., buffy coat and HCC tumor, may be calculated. In some instances, dark colors indicate methylation of the corresponding CpG sites ("1") and white indicates non-methylation of the CpG sites ("0"). Each pie chart of the reference tumor tissue can represent the proportion (percentage) of reference DNA molecules that are methylated at the corresponding CpG site, so a large proportion of the pie chart with a darker color means a high proportion of the reference DNA molecules are methylated at the corresponding CpG site.

[0261] Thus, the methylation state of each CpG site of a given long cell-free DNA molecule can be compared to the respective pie charts of the reference tissues, and a tissue match can be identified based on the reference tissue pattern that, taken together, least deviates from the methylation state of all CpG sites present on the long cell-free DNA molecule. For example, for a CpG site, a first distance between the methylation state of the long cell-free DNA molecule and the proportion of reference DNA molecules in the reference tumor tissue can be calculated. As an example, the distance between a methylation site (1) on a DNA molecule and a reference with a methylation index (density) of 60% can be 0.4. If the DNA molecule is unmethylated, the distance can be 0.6. For the same CpG site, a second distance between the methylation state of the long cell-free DNA molecule and the proportion of reference DNA molecules in the reference buffy coat can be calculated. The first and second distances can be compared. In this example, the first distance is less than the second distance, which can indicate that the CpG site has a similar methylation state to the reference tumor tissue.

[0262] In some embodiments, the methylation index for each of the CpG sites in the reference tissue, defined as the percentage or proportion of sequenced CpGs identified as methylated, is obtained from bisulfite sequencing (BS-seq) data. Additionally or alternatively, the aggregate methylation index in the reference tissue may be obtained from Enzymatic Methyl-seq data (i.e., EM-seq).

[0263] Distance between methylation haplotypes of plasma DNA and reference tissue methylome may include, but is not limited to, Euclidean distance, cosine similarity, Hamming distance, edit distance, etc. In some embodiments, distance calculation can be adjusted by weight vector according to different genomic positions. For example, positions with large methylation difference between tumor tissue and non-tumor tissue are assigned higher weight. In contrast, positions with small methylation difference between tumor tissue and non-tumor tissue are assigned lower weight. C. Tissue of origin analysis using methylation scores 1.Hepatocellular carcinoma (HCC)

[0264] In some embodiments, the methylation pattern analysis described herein may include further illustrating the analysis of the tissue of origin of the plasma DNA molecules using additional calculations. As an illustrative example, the methylation pattern of each plasma DNA molecule was identified according to the polymerase kinetic signal around the CpG site using the Holistic Kinetics (HK) model (Tse et al. Proc Natl Acad Sci USA. 2021;118: e2019768118). Such methylation pattern of each plasma DNA molecule was compared with a reference methylation profile, including but not limited to liver tissue, buffy coat, colon tissue, lung tissue, etc. In some embodiments, the reference methylation profile is obtained based on the results of high-depth bisulfite sequencing. For each CpG site in the genome of each reference tissue, a methylation index (MI) was calculated by the following formula:

number

[0265] CpG sites with MI difference between liver tissue and buffy coat greater than 30% were considered informative for downstream analysis. In some embodiments, MI difference includes, but is not limited to, 5%, 10%, 15%, 20%, 25%, 40%, 45%, 50%, 55%, 60%, 65%, 70%, 75%, 80%, 85%, 90%, etc. In some embodiments, a scoring system is used to identify the likelihood that a DNA molecule originates from a particular tissue based on a comparison of the methylation pattern observed in that molecule with a reference methylation profile. For each DNA molecule with n informative CpG sites, a methylation score S(liver) was calculated according to the following formula:

number

[0266] Similarly, to determine the similarity of the methylation patterns between plasma DNA molecules and the buffy coat, another methylation score S(buffy coat) was calculated as follows:

number

[0267] Similarly, to determine the similarity of methylation patterns between plasma DNA molecules and the colon, another methylation score, S(colon), was calculated as follows:

number

[0268] Similarly, to identify the similarity of methylation patterns between plasma DNA molecules and lung, another methylation score S(lung) was calculated as follows:

number

[0269] Among S(liver), S(buffy coat), S(colon) and S(lung), if S(liver) is the highest, the corresponding DNA molecule is classified as being of liver origin, otherwise it is classified as being of hematopoietic, colonic or pulmonary origin, depending on which methylation score is the highest. 2. Prediction of HCC stage using cancer methylation score

[0270] FIG. 47 shows a box plot 4700 identifying the percentage of DNA molecules identified as being of liver origin in HCC patients at different stages based on methylation haplotype analysis according to an embodiment of the present disclosure. In particular, FIG. 47 shows the percentage of plasma DNA molecules in patients with HCC at different stages according to the BCLC staging system. As the disease stage progresses, there is a trend for liver-derived fragments to increase. From the tissue of origin analysis of plasma DNA molecules, the severity of the disease, such as the stage of cancer suffered by the patient, can be identified. Thus, methylation haplotype-based analysis can be effectively used to guide treatment selection and prognosis prediction.

[0271] Based on the methylation haplotype analysis and methylation scores according to the embodiments described in the present disclosure, a metric named cancer methylation score can be used to reflect the presence and / or severity of cancer. The methylation patterns of the plasma DNA molecules were compared with the methylation profiles of reference tissues containing cancer. A first score S(cancer), reflecting the similarity between the DNA molecules and the analyzed tumor in terms of methylation pattern, was calculated according to the following formula:

number

[0272] Similarly, to identify the highest similarity among the comparisons between the methylation pattern of the DNA molecule and tissue reference methylation profiles, including but not limited to buffy coat, liver tissue, colon tissue, lung tissue, etc., a second score S(non-cancer) was calculated by the following formula:

number

[0273] Finally, the cancer methylation score was generated by combining both S(cancer) and S(non-cancer) using the following formula:

number

[0274] FIG. 48 illustrates a box plot 4800 identifying cancer methylation scores of HCC patients across different disease stages, according to some embodiments. FIG. 48 illustrates a cancer methylation score analysis in which the cancer methylation scores (also referred to as "HCC methylation scores") of patients with HCC were identified. The HCC patients had different disease stages according to the BCLC staging system. With the progression of disease stages, the HCC methylation scores gradually increased. Thus, based on the cancer methylation score analysis, the severity of the disease, such as the stage of cancer suffered by the patient, can be identified. Thus, the cancer methylation score can be effectively used to guide treatment selection and prognosis prediction.

[0275] In some embodiments, survival analysis was performed on a cohort of HCC patients based on the HCC methylation score. For example, cases with an HCC methylation score equal to or less than the median HCC methylation score were classified as "group A," and cases with an HCC methylation score greater than the median HCC methylation score were classified as "group B." Kaplan-Meier survival curves can be used to reflect the distribution of survival probabilities between different groups. As described herein, a survival curve corresponds to a graph showing the number or percentage of individuals in a given group who survived to each age. A rapid decline in the survival curve indicates that a given group will die sooner compared to a slow decline in the survival curve.

[0276] FIG. 49 shows a set of survival curves 4900 identifying survival analysis in HCC patients, according to some embodiments. Curve 4902 shows DNA molecules with at least seven CpG sites for HCC methylation score analysis. Curve 4904 shows DNA molecules with less than seven CpG sites for HCC methylation score analysis. As shown in FIG. 49, HCC patients in group B (4906A and 4906B) tend to have worse survival than group A (4908A and 4908B). Using longer cfDNA molecules with at least seven CpG sites (median length: 758 bp, curve 4902) results in a greater difference in Kaplan-Meier survival curves between the two groups than using shorter DNA molecules with less than seven CpG sites (median length: 311 bp, curve 4904), suggesting that using longer cfDNA molecules is more effective than shorter cfDNA molecules in predicting prognosis. For example, after two years, curve 4902 shows that 91% of group A patients will survive and 81% of group B patients will survive, where the corresponding cancer methylation scores are derived from short cfDNA molecules. In contrast, curve 4904 shows that 96% of group A patients will survive and 77% of group B patients will survive, where the corresponding cancer methylation scores are derived from long cfDNA molecules. Thus, cancer methylation score analysis can be used to identify disease survival probability.

[0277] In addition, cancer methylation scores can be effectively used based on long sequence reads obtained using various sequencing platforms. Figure 50 shows a box plot 5000 identifying HCC methylation scores of HBV carriers and HCC patients calculated using data from SMRT-seq (5002) and nanopore sequencing (5004). HCC methylation scores were calculated according to embodiments in the present disclosure. As shown in Figure 50, HCC patients showed significantly higher HCC methylation scores than HBV carriers in both SMRT-seq (P<0.001, Mann-Whitney U test) and nanopore sequencing (P=0.0026, Mann-Whitney U test).

[0278] Therefore, nanopore sequencing by Oxford Nanopore Technologies (ONT) can also be used to analyze nucleic acids. To demonstrate the effectiveness of nanopore sequencing, plasma DNA molecules from 8 HCC patients and 6 HBV carriers were sequenced by both nanopore sequencing and SMRT sequencing. Figure 51 shows a graph 5100 that identifies the percentage of liver-derived cfDNA by single molecule tissue of origin analysis in plasma samples from HBV carriers (5102) and HCC patients (5104) using data from SMRT-seq and nanopore sequencing. As shown in Figure 51, the percentage of cfDNA molecules classified as liver-derived in HBV carriers and HCC patients using nanopore sequencing showed agreement with that by SMRT sequencing (Pearson correlation, r=0.99, P<0.001). 3.Colorectal cancer (CRC)

[0279] In addition to HCC, plasma DNA molecules from CRC patients, HCC patients and healthy subjects were analyzed using SMRT-sequencing and tissue of origin analyzed by methylation score.

[0280] CpG sites with MI difference between colon tissue and buffy coat greater than 30% were considered informative for downstream analysis. In some embodiments, MI difference includes, but is not limited to, 5%, 10%, 15%, 20%, 25%, 40%, 45%, 50%, 55%, 60%, 65%, 70%, 75%, 80%, 85%, 90%, etc. In some embodiments, a scoring system is used to identify the likelihood that a DNA molecule originates from a particular tissue based on a comparison of the methylation pattern observed in that molecule to a reference methylation profile. For a DNA molecule with n informative CpG sites, the methylation score S(colon) was calculated by the following formula:

number

[0281] Similarly, to identify the similarity of the methylation patterns between plasma DNA molecules and the buffy coat, another methylation score S(buffy coat) was calculated as follows:

number

[0282] Similarly, to identify the similarity of methylation patterns between plasma DNA molecules and liver, another methylation score S(liver) was calculated as follows:

number

[0283] Similarly, to identify the similarity of methylation patterns between plasma DNA molecules and lung, another methylation score S(lung) was calculated as follows:

number

[0284] Among S(colon), S(buffy coat), S(liver) and S(lung), if S(colon) is highest, the corresponding DNA molecule is classified as of colon origin, otherwise it is classified as of hematopoietic, liver or lung origin, depending on which methylation score is highest.

[0285] FIG. 52 shows a box plot 5200 identifying the percentage of plasma DNA molecules classified as colon origin in 15 healthy subjects, 45 HCC patients, and 4 CRC patients based on the embodiments presented in the present disclosure. This example analysis included DNA molecules (median length: 896 bp) with at least 7 CpG sites. As shown in FIG. 52, CRC patients had a significantly higher percentage of DNA molecules classified as colon origin than healthy subjects (P value: 0.0005, Mann-Whitney U test), indicating a clear separation between CRC and HCC patients (P value: 0.0018, Mann-Whitney U test). This not only demonstrated the diagnostic ability of the methylation score analysis of the embodiments in the present disclosure in distinguishing subjects with and without colorectal cancer, but also highlighted its specificity in accurately identifying the tissue of origin of the cancer. D. Histological status of the disease

[0286] Further analysis of the long cell-free DNA molecules may be performed to obtain the histological status of the disease. Figure 53 shows a set of bar plots 5300 identifying the percentage of DNA molecules identified as being of HCC tumor origin among HCC patients with and without vascular invasion based on methylation haplotype analysis according to some embodiments. Figure 53 shows that the median percentage of DNA molecules identified as being of HCC tumor origin was higher in HCC patients with vascular invasion (16.68%) than in HCC patients without vascular invasion (14.08%). The data suggested that tumor-derived DNA molecules identified by methylation haplotype-based analysis could be used to provide information about the histological status of the tumor. E. Comparison of Methylation-Based Analysis Using Long vs. Short cfDNA Molecules

[0287] Plasma DNA with a size larger than 1 kb can be used to accurately classify disease for biological samples. Figure 54 shows a set of bar plots 5400 identifying the percentage of DNA molecules identified as HCC tumor origin, according to some embodiments. Figure 54 shows that the percentage of DNA molecules identified as HCC tumor origin was significantly higher in patients with HCC than in patients without HCC (median: 14.78% vs. 10.98%. P value: 0.024, Mann-Whitney U test). The results suggested that the analysis of tissue / tumor origin of each long plasma DNA molecule can serve as a tool for cancer detection.

[0288] Furthermore, to evaluate whether methylation haplotype analysis of long plasma DNA has advantages over the use of short DNA molecules (<600bp), plasma DNA sequence data obtained using PacBio direct methylation HK model analysis (Tse et al. Proc Natl Acad Sci USA. 2021;118: e2019768118) of samples from patients with and without HCC were divided into two groups. The first group of molecules corresponded to a size of >1kb, and the second group of molecules corresponded to a size of <600bp. For the first group, we attempted to detect tumor-derived molecules based on methylation haplotypes according to the embodiments presented in this disclosure. For the second group, we calculated the global methylation level (percentage of methylated CpG sites in the entire human genome by plasma DNA molecules) and identified liver DNA contribution based on the aggregate methylation level instead of methylation haplotype information.

[0289] FIG. 55 shows a set of ROC curves 5500 identifying the cancer detection accuracy of analysis of single molecule methylation sequence data of long cell-free DNA and other analyses using methylation sequence data of short cell-free DNA. Line A (5502) shows methylation haplotype analysis of plasma DNA molecules >1 kb in size according to embodiments presented in this disclosure. Line B (5504) shows the percentage of methylated CpG sites in the entire human genome by plasma DNA molecules <600 bp. Line C (5506) shows the liver contribution derived by aggregate methylation levels of plasma DNA molecules <600 bp instead of methylation haplotype information using a quadratic programming approach.

[0290] Figure 55 shows that methylation haplotype analysis using the first group of molecules (e.g., long cell-free DNA molecules) (AUC: 0.83) outperformed the other two methods tested with the second group of molecules (AUC: <0.7). These results demonstrated that methylation haplotype-based analysis of long plasma DNA molecules is superior to methylation analysis of shorter plasma DNA molecules in cancer detection.

[0291] Another comparative analysis was performed to compare the performance of methylation haplotype-based analysis using long cell-free DNA molecules with plasma DNA tissue mapping analysis of short cell-free DNA molecules.To obtain short cell-free DNA molecules, 34 controls and 38 HCC subjects were sequenced using short-read bisulfite sequencing technology (Illumina) with a median of 211 million 75 bp x 2 paired-end reads (range: 112 to 1,681 million).

[0292] Figure 56 shows a pair of ROC curves 5600 identifying the HCC detection accuracy of a methylation haplotype-based analysis using long DNA 5602 (>1 kb) and a plasma DNA tissue mapping analysis using short-read bisulfite sequencing of short plasma DNA molecules 5604 (<600 bp). As shown in Figure 56, for this cohort of samples, the plasma DNA tissue mapping analysis (Sun et al. Proc Natl Acad Sci USA. 2015;112:E5503-5512) provided an AUC of 0.76 in distinguishing patients with and without HCC. Such an AUC value indicates that the performance of the analysis based on short plasma DNA molecules was below that of the analysis based on methylation haplotypes of long plasma DNA molecules (AUC: 0.83). Plasma DNA tissue mapping, which used aggregate methylation probabilities in genomic regions from a population of short DNA molecules (Sun et al. Proc Natl Acad Sci USA. 2015;112:E5503-5512), did not take into account the information and utility of methylation haplotypes of individual long plasma DNA molecules. F. Methods for Analyzing Methylation Patterns of Long Cell-Free DNA Molecules

[0293] Figure 57 shows a flow chart of process 5700 illustrating an example process for analyzing a subject's biological sample based on the methylation patterns of long cell-free DNA molecules, according to some embodiments. The biological sample may include DNA originating from normal cells and possibly cells associated with one or more of a number of tissue types. In addition, at least a portion of the DNA is cell-free in the biological sample.

[0294] At step 5702, sequence reads obtained from methylation-aware sequencing of the cell-free DNA molecules may be received. The methylation-aware sequencing may include an enzyme treatment. In some instances, the methylation-aware sequencing does not include bisulfite treatment. In other instances, bisulfite treatment is used. Each of the sequence reads may include a methylation pattern of methylation states at a set of sites (e.g., CpG sites) on the sequence read. For example, a sequence read may include six CpG sites that display a methylation pattern of "-MMMUUU-", where "M" represents a methylated state and "U" represents an unmethylated state. In another example, a given sequence read may include at least three CpG sites. The methylation pattern may include the number of bases (e.g., a specified number of bases) between pairs of sites in the set of sites, and the identity of the bases.

[0295] The methylation status at sites of cfDNA molecules can be examined using bisulfite conversion as described herein. Apart from bisulfite conversion, other processes known to those skilled in the art can be used to examine the methylation status of DNA molecules, including, but not limited to, methylation status-sensitive enzymes (e.g., methylation-sensitive restriction enzymes), methylation-binding proteins, single molecule sequencing using methylation status-sensitive platforms (e.g., nanopore sequencing (Schreiber et al. Proc Natl Acad Sci 2013; 110: 18910-18915) and Pacific Biosciences single molecule real-time analysis (Tse et al. Proc Natl Acad Sci USA 2021; 118: e2019768118)).

[0296] In some instances, the sequence reads correspond to long cell-free DNA molecules having a size within a first size range, which may include a lower limit and an upper limit. By way of example, the first size range may include an upper limit of at least 1,000 bases, at least 3,000 bases, or more. In some instances, the lower limit may be selected from one of at least 300 bases, at least 400 bases, at least 500 bases, at least 600 bases, or at least 800 bases.

[0297] The set of sites may be of various numbers. In some instances, the set of sites of each sequence read may include at least N sites. For example, a given sequence read may include at least three CpG sites. Other numbers, including but not limited to at least 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 30, 35, 40, 45, 50, or more than 50 sites, may also be considered. Additionally or alternatively, the sequence read may correspond to a long cell-free DNA molecule having a size within a first size range (e.g., greater than 500 bp) and may include at least N sites (e.g., three CpG sites).

[0298] Steps 5704 and 5706 may be performed for each sequence read among the sequence reads received from step 5702. At step 5704, the methylation pattern of the sequence read may be compared to a first reference methylation pattern. In some instances, the first reference methylation pattern corresponds to a first tissue type. The first tissue type may be a diseased tissue type. In some instances, the first tissue type is associated with a disease. Additionally or alternatively, the methylation pattern of the sequence read may be further compared to each reference methylation pattern among one or more other reference methylation patterns. Each reference methylation pattern may correspond to a tissue type among the plurality of tissue types.

[0299] The values ​​of the reference methylation pattern may be binary values ​​(e.g., 0 and 1 in FIG. 41 or 42) or may have decimal values ​​(e.g., 0.2 indicating a methylation index of 20%). The reference pattern may be general to a tissue type or specific to a particular location. In such cases, the location of the sequence read may be identified. Thus, in some embodiments, comparing the methylation pattern to the reference pattern may include identifying the location of the sequence read (e.g., relative to a reference genome), where the reference methylation pattern corresponds to a reference sequence at the location.

[0300] In some instances, the comparison between the methylation pattern of the sequence read and the first reference methylation pattern may include calculating a similarity metric based on the difference between the methylation state of a site and the methylation index of the first reference methylation pattern at the same site. The similarity metric may be a distance (e.g., Euclidean distance), a cosine similarity, or a methylation score.

[0301] In some instances, the methylation state indicates whether the corresponding site is methylated or not. In this instance, the methylation state comprises a binary value indicating the methylation of the site. A similarity metric may be determined from the set of sites to identify an aggregate value (e.g., sum, average, median) of the sequence reads. The aggregate value may be compared to one or more cutoffs to identify a tissue classification of the sequence reads, where the one or more cutoffs may be identified using a reference sample known to be associated with the first tissue type. Additionally or alternatively, comparing the methylation pattern to the first reference methylation pattern may include identifying a methylation level of the sequence reads based on the methylation state at the corresponding set of sites, identifying a difference between the methylation level and another methylation level identified from the first reference methylation pattern based on the methylation level, and comparing the difference to one or more cutoff values. The methylation level can be a methylation index, a methylation density, a number of molecules that are methylated at one or more sites in a set of sites, or a percentage of molecules (e.g., cytosines) that are methylated at one or more sites in a set of sites.

[0302] In some instances, the similarity metric is a methylation score. To calculate the methylation score, a reference methylation profile representing the first reference methylation pattern may be identified by calculating a methylation index for each CpG in the genome from which the first reference methylation pattern was obtained. Then, for each CpG site in the sequence read, a difference between the methylation state of the CpG site (e.g., a binary value between 0 and 1) and the corresponding methylation index at the same CpG site may be identified. The identified differences across the CpG sites may be aggregated to identify a methylation score for the sequence read. In some instances, the aggregated value is normalized (e.g., by the total number of CpG sites in the sequence read) to identify a methylation score for the sequence read. The step of identifying a methylation score is further described in Sections IV.C and IV.D of the present disclosure.

[0303] At step 5706, a tissue classification of the sequence read may be determined based on the comparison. The tissue classification may be performed as described above with respect to Figures 41, 42, and 46. For example, the comparison may be determined by determining, for each site in the set of sites in the methylation pattern, a similarity metric between the methylation state of the site and the methylation index of the corresponding site in the first reference methylation pattern. The similarity metrics across the set of sites may be aggregated to determine an aggregate value (e.g., a sum of the similarity metrics). If the aggregate value exceeds a cutoff value, the sequence read may be classified as associated with the first tissue type. If the aggregate value does not exceed the cutoff value, the sequence read may be classified as associated with one of the other tissue types. The cutoff value may be determined using one or more reference samples known to be associated with the first tissue type.

[0304] In another example, a closest reference methylation pattern can be identified between a set of reference methylation patterns (e.g., a first reference methylation pattern and one or more other reference methylation patterns) based on their respective tabulation values, and a tissue classification can be determined as the corresponding tissue type of the reference methylation pattern with the highest tabulation value. In particular, a tabulation value as described in the above paragraph can be determined for each reference methylation pattern. The sequence read can then be classified as originating from the tissue type associated with the reference methylation pattern with the highest tabulation value. Thus, the tissue classification can indicate that the sequence read originates from one of a plurality of tissue types (or the degree of origin). The tissue classification can include a probability that the sequence read originates from one of a plurality of tissue types. Probabilities of more than one tissue type may be determined.

[0305] Additionally or alternatively, a closest reference methylation pattern may be identified between a set of reference methylation patterns (e.g., a first reference methylation pattern and one or more other reference methylation patterns) based on a direct comparison of methylation status, and a tissue classification may be identified as the corresponding tissue type of the closest methylation reference pattern. For example, a sequence read has six CpG sites that display a methylation pattern of "-MMMUUU-", where "M" represents a methylation status and "U" represents an unmethylated status. Other molecules contain corresponding CpG sites from other tissues that display reference methylation patterns of "-MUMUMU-", "-MUMMUU-", "-MUUUMM-", "-MMUUMU-", "-MMUUUM-", "-UMMMUU-", "-UUMUMM-", "-UUMMMU-", "-UUUMMM-". However, a reference methylation pattern corresponding to liver tissue (for example) may be "-MMMUUM-". In this example, the sequence read may be identified as being associated with liver tissue. Thus, the combination of methylation patterns across a set of CpG sites in a molecule can act as a "molecular barcode" that indicates cellular properties or disease states.

[0306] When the comparison involves the use of a methylation score, the tissue classification may include identifying another methylation score based on the methylation state of the sequence read and a methylation index from another reference methylation pattern. For example, the first reference methylation pattern may correspond to a first tissue type (e.g., liver), the second methylation pattern may correspond to a second tissue type (e.g., buffy coat), and the third methylation pattern may correspond to a third tissue type (e.g., colon). The tissue type associated with the highest methylation score may then be identified as the tissue classification of the sequence read. In some instances, the comparison includes comparing the methylation scores of two reference methylation patterns (e.g., a first and a second reference methylation pattern), where the first reference methylation pattern corresponds to a first tissue type and the second reference methylation pattern corresponds to one or more other tissue types. Using the methylation score to perform tissue classification is further described in Section IV.C of this disclosure.

[0307] If the first tissue type is a disease type, a first methylation score (e.g., S(cancer) score) can be identified that corresponds to a reference methylation profile of the diseased tissue (e.g., HCC), and a second methylation score (e.g., S(non-cancer) score) can be identified that corresponds to a reference methylation profile of the healthy tissue (e.g., non-HCC). The first and second methylation scores identified for the sequence reads may then be used together to identify a cancer methylation score to classify the disease. Using the cancer methylation score to perform disease classification is further described in Section IV.D of this disclosure.

[0308] In step 5708, a disease classification of the disease of the biological sample may be identified based on a tissue classification of the sequence reads. If the tissue type is a disease tissue type, the tissue classification and the disease classification may be equivalent. For the disease classification, the cancer methylation score identified in step 5706 may be used to identify the disease classification. The disease may be cancer. Identifying the disease classification may include determining the presence or absence of vascular invasion by the cancer. In some instances, identifying the disease classification includes (i) identifying a first amount of sequence reads classified as originating from a first tissue type, and (ii) comparing the first amount to one or more reference values ​​to identify a disease classification of the biological sample.

[0309] The one or more reference values ​​may be determined from a reference sample with known disease classification. If the first amount of sequence reads exceeds the one or more cutoff values, the subject can be classified as having the disease. In contrast, if the first amount of sequence reads does not exceed the one or more cutoff values, the subject can be classified as not having the disease.

[0310] The amount can be a sum of probabilities of the first tissue type. For example, where the tissue classification corresponds to a probability value or a methylation score, the first amount of sequence reads can include a sum of the probability values ​​or methylation scores of sequence reads classified as originating from the first tissue type. In some instances, the sum is determined based on the probability values ​​or methylation scores of sequence reads that exceed a probability threshold.

[0311] In some instances, the disease classification is identified by comparing a first amount of sequence reads corresponding to a first reference methylation pattern with an amount of one or more other reference methylation patterns, where each of the one or more other reference methylation patterns is associated with one or more other tissue types. One or more other amounts of sequence reads classified as originating from one or more other tissue types may be identified. Based on the comparison of the first amount of sequence reads to the one or more other amounts, a disease classification of the biological sample can be identified. For example, if the first amount of sequence reads is the maximum amount, the subject can be identified as having a disease of the first tissue type.

[0312] The classification of the disease may include a classification of the severity of the disease (e.g., no disease, early stage, intermediate stage, advanced stage). For example, the classification of the disease may include a stage of the cancer according to BCLC stages. The classification may then select one of the stages. Thus, the classification may be specified from multiple stages of the disease (e.g., one of the BCLC stages of HCC). In some instances, the disease is cancer. By way of example, the cancer may be hepatocellular carcinoma, lung cancer, breast cancer, gastric cancer, glioblastoma multiforme, pancreatic cancer, colorectal cancer, nasopharyngeal carcinoma, and head and neck squamous cell carcinoma. V. Single-molecule methylation level analysis of long cell-free DNA molecules to predict disease severity

[0313] Assessment of disease severity is important to guide treatment decisions, prognosis, and surveillance. Taking hepatocellular carcinoma (HCC) as an example, patients with early-stage HCC (e.g., Barcelona Clinic Liver Cancer (BCLC) 0 / A) are expected to have a median survival of >5 years and are usually offered curative treatments such as ablation, resection, and transplantation. In contrast, patients with advanced-stage HCC (e.g., BCLC C) are expected to have a median survival of >2 years and are usually offered systemic treatments (Reig. et al. J. Hepatol. 2022;76:681-693).

[0314] According to embodiments of the present disclosure, methylation-based analysis of cfDNA molecules, including but not limited to single molecule methylation pattern analysis and cfDNA tissue of origin analysis, can be used to prognose disease severity, including but not limited to predicting cancer stage. A. Disease classification using single molecule methylation levels

[0315] Single molecule real-time sequencing (SMRT-Seq) was used to sequence plasma DNA molecules from 45 HCC patients, 13 Hepatitis B virus (HBV) carriers, and 15 healthy individuals. FIG. 58 shows a box plot 5800 identifying single molecule methylation levels in different groups of individuals during single molecule real-time sequencing (SMRT-Seq), according to some embodiments. Plasma DNA molecules from HCC patients had significantly lower average single molecule methylation levels than control individuals (P value: 0.005, Mann-Whitney U test) (FIG. 58). Single molecule level herein was defined by the percentage of CpG sites identified as methylated in a single molecule. For example, if a DNA molecule contains 10 CpG sites and 5 of them are identified as methylated, the single molecule methylation level would be 50% (i.e., 5 / 10*100%). Single molecule methylation levels may be determined for each sequence read of a given biological sample, and a statistical value (e.g., average, median) may be determined from the single molecule methylation levels. These data in Figure 58 suggested that the use of single molecule methylation levels may enable detection of HCC patients.

[0316] In some embodiments, criteria can be used to further enhance diagnostic performance since SMRT-seq can obtain more long cfDNA molecules based on criteria not limited to molecular size, number of CpG sites, and methylation level, which was not feasible for Illumina sequencing platforms that cannot sequence long cfDNA molecules (e.g., >600 bp).

[0317] FIG. 59 shows a box plot 5900 identifying single molecule methylation levels in DNA molecules that are >500bp in size, contain at least three CpG sites, and have a methylation level of ≦60% in SMRT-Seq. As shown in FIG. 59, the average single molecule methylation levels in healthy subjects, HBV carriers, and HCC patients, respectively, were analyzed for DNA molecules that are >500bp in size, contain at least three CpG sites, and have a methylation level of ≦60%. We found that HCC patients showed significantly lower methylation levels compared to patients without HCC (P-value: 2.132 x 10 -8 , Mann-Whitney U test).

[0318] FIG. 60 shows a ROC curve 6000 identifying the performance of single molecule methylation levels in distinguishing HCC subjects from non-HCC subjects in SMRT-Seq and short-read sequencing (e.g., Illumina sequencing), according to some embodiments. As shown in FIG. 60, compared to whole DNA molecules not subjected to size selection, such selective analysis based on molecules >500bp in size (6002) improved diagnostic performance, with the area under the receiver operating characteristic (ROC) curve (AUC) increasing from 0.7 to 0.87. In addition, molecules >500bp (6004) used for such methylation analysis based on short-read sequencing only provided an AUC of 0.56 (Jiang et al. Cancer Discov. 2020; 10:664-673), which was much worse than the embodiments disclosed herein.

[0319] FIG. 61 shows a box plot 6100 identifying single molecule methylation levels in HCC patients with different Barcelona Clinic Liver Cancer (BCLC) stages. FIG. 61 shows that the average single molecule methylation levels in patients changed according to different HCC stages according to the BCLC staging system. With the progression of cancer stages, the single molecule methylation levels gradually decreased. Thus, the single molecule methylation levels of plasma DNA molecules can be used to provide information about the severity of disease, such as the stage of cancer suffered by the patient. In fact, the single molecule methylation levels of plasma DNA molecules can guide treatment selection and prognosis prediction. In some instances, the single molecule methylation levels of long plasma DNA molecules (e.g., larger than 600 bp in size) are used to improve the accuracy of identifying the severity of disease. B. Methods for identifying disease classification using single molecule methylation levels in DNA molecules

[0320] 62 shows a flow chart 6200 illustrating an example process for identifying a disease classification using single molecule methylation levels in DNA molecules, according to some embodiments. The biological sample may include DNA originating from normal cells and possibly cells associated with one or more of a plurality of tissue types. In addition, at least a portion of the DNA is acellular in the biological sample.

[0321] In step 6202, sequence reads obtained from methylation-aware sequencing of the cell-free DNA molecules may be received. Each of the sequence reads may include a methylation state corresponding to a set of sites (e.g., CpG sites) on the sequence read. The methylation-aware sequencing may include single molecule sequencing or nanopore sequencing that can be used to identify the methylation state at each CpG site of each cell-free DNA molecule. For example, single molecule real-time sequencing (SMRT-Seq) or nanopore sequencing may be used to sequence the cell-free DNA molecules to obtain the sequence reads. Additionally or alternatively, other processes, including but not limited to bisulfite conversion, enzymes sensitive to methylation states (e.g., methylation-sensitive restriction enzymes), and methylation-binding proteins may be used to identify the methylation state of the CpG sites.

[0322] In some instances, the sequence reads correspond to long cell-free DNA molecules having a size within a first size range, which may include a lower limit and an upper limit. By way of example, the first size range may include an upper limit of at least 1,000 bases, at least 3,000 bases, or more. In some instances, the lower limit may be selected from one of at least 300 bases, at least 400 bases, at least 500 bases, at least 600 bases, or at least 800 bases.

[0323] In some instances, each of the sequence reads includes one or more sites having one or more methylation states that allow for identification of the methylation level of the corresponding cell-free DNA molecule. Each of the one or more sites can be associated with a methylation state. For example, the one or more sites can be CpG sites, and each site can be a CpG site for which a particular methylation state is identified. In some instances, the one or more sites of each of the sequence reads includes at least N sites. For example, a given sequence read can include at least 3 CpG sites. Other numbers, including but not limited to 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 30, 35, 40, 45, 50, or more than 50 sites, can also be considered. Additionally or alternatively, the sequence reads can correspond to long cell-free DNA molecules having a size within a first size range (e.g., greater than 500 bp) and can include at least N sites (e.g., 3 CpG sites). Identifying the methylation status of the set of sites is further described in step 5702 of FIG.

[0324] Steps 6204 and 6206 may be performed for each sequence read among the sequence reads received from step 6202. At step 6204, a methylation state of each of the one or more sites in the sequence read may be determined. In some instances, the methylation state of a given site includes a binary value (e.g., 0 and 1 in FIG. 58) that identifies whether the site is methylated or not.

[0325] In step 6206, a methylation level of the sequence read may be determined based on the methylation status of one or more sites. The methylation level may be a methylation index, a methylation density, a number of molecules methylated at one or more sites in a set of sites, or a percentage of molecules (e.g., cytosines) methylated at one or more sites in a set of sites. For example, the methylation level identifies the methylation percentage of the identified CpG sites based on the number of methylated CpG sites and the total number of CpG sites in the sequence read. For example, if a DNA molecule contains 10 CpG sites and 5 of them are determined to be methylated, the single molecule methylation level is 50% (i.e., 5 / 10*100%).

[0326] In step 6208, a statistical value for the biological sample may be determined based on the determined methylation levels of the sequence reads. For example, the statistical value may be an average, median, or representative value of the methylation levels corresponding to the sequence reads. Additionally or alternatively, the statistical value may be an aggregate value (e.g., a sum) determined from the methylation levels of the sequence reads.

[0327] In step 6210, the statistical value of the cell-free DNA fragments is compared with a reference value to identify the level of the classification of the pathology of the subject. The reference value may include a cutoff value or threshold value, or is for determining a cutoff value or threshold value. The cutoff or threshold value may be derived from a cutoff or threshold value that represents a particular classification or distinguishes between two or more classifications. In some instances, the reference value is determined using a reference sample with a known classification of the pathology. Subjects whose statistical value is above or below the cutoff value (threshold value) may be classified as having a congenital disorder. The cutoff value may be defined by a statistical metric (e.g., significance, P value, Z score) relative to the reference value.

[0328] By way of example, the condition may be cancer. By way of example, the level may be cancer-free, early stage, intermediate stage, or advanced stage. And the classification may select one of the stages. Thus, the classification may be specified from multiple stages of cancer (e.g., one of the BCLC stages). By way of example, the cancer may be hepatocellular carcinoma, lung cancer, breast cancer, gastric cancer, glioblastoma multiforme, pancreatic cancer, colorectal cancer, nasopharyngeal carcinoma, and head and neck squamous cell carcinoma. VI. Machine learning techniques for disease classification based on methylation patterns in long cell-free DNA molecules

[0329] In some embodiments, pattern recognition analysis of methylation haplotypes may be performed using a machine learning model that can extract useful information from methylation haplotypes for classification of patients with and without cancer. Sequence reads may be obtained from methylation-aware sequencing of cell-free DNA molecules without bisulfite treatment, because the chemical reaction of bisulfite treatment may prevent obtaining sequence reads corresponding to long cell-free DNA molecules (e.g., >600bp). In some instances, the methylation pattern of each long cell-free DNA molecule is converted into a matrix of values, where the long cell-free DNA molecule may be associated with a particular tissue type. The matrix can be used to train a machine learning model to identify tissue classification.

[0330] Compared to previously known techniques, the machine learning model can identify certain sites in long cfDNA molecules that are more predictive than other sites in disease classification. Furthermore, the increased number of CpG sites in long cfDNA molecules allows the machine learning model to be trained on a greater variety of methylation patterns. In fact, the machine learning model can more accurately identify disease classification compared to another model trained on short DNA molecules that generally have fewer CpG sites.

[0331] The machine learning models may include, but are not limited to, convolutional neural networks (CNNs), linear regression, logistic regression, deep recurrent neural networks (e.g., fully connected recurrent neural networks (RNNs), gated recurrent units (GRUs), long short-term memory, (LSTM)), transformer-based methods (e.g., XLNet, BERT, XLM, RoBERTa), Bayesian classifiers, hidden Markov models (HMMs), linear discriminant analysis (LDA), K-means clustering, density-based spatial clustering of applications with noise (DBSCAN), random forest algorithms, adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), support vector machines (SVMs), or hybrid models including one or more of the above models. A. Machine Learning Models

[0332] FIG. 63 shows an example diagram 6300 of pattern recognition of methylation haplotypes using a machine learning model, according to some embodiments. In some embodiments, the machine learning model is a hybrid model including a CNN followed by an LSTM. In particular, FIG. 63 shows an example of the use of pattern recognition of methylation haplotypes to classify tumor and non-tumor DNA in the plasma of patients with cancer. 1. Training Data

[0333] EM-seq was used to obtain long methylation haplotypes (e.g., >5 kb) from tumor cells (green) and non-tumor cells (blue). A horizontal line (DNA) with a series of filled and unfilled dots (i.e., methylated and unmethylated CpG sites) represents a methylation haplotype. In some instances, to obtain DNA fragments from tumor cells, tissue DNA is sonicated to obtain molecules with specific sizes (e.g., 5 kb, 10 kb). Methylation-aware sequencing, including bisulfite sequencing, may be used to obtain sequence reads for data training. In some instances, non-tumor DNA fragments are obtained from T cells, B cells, neutrophils, lung tissue, liver, etc.

[0334] Each long methylation haplotype is programmed into a data matrix that includes both sequence composition and methylation pattern. The data matrix can include one-hot encoding of bases to identify the methylation state of each CpG site of the corresponding cell-free DNA molecule. The term "one-hot encoding" refers to a technique for quantifying categorical data such that the categorical data is converted into a numerical representation. In particular, the technique can include generating a vector (e.g., bases) of length equal to the number of categories in the data set. For example, if a base belongs to the T category, the components of this vector are assigned a value of 0, except for the T component, which is assigned a value of 1. One-hot encoding allows categories to be tracked in a numerically meaningful manner.

[0335] The first row of the matrix shows the sequence information "...ACGTACGTCT...", where "..." indicates that the bases are omitted for brevity. For illustrative purposes, the first CpG site is unmethylated and the second CpG site is methylated. In column "i", a "1" was entered at the intersection of column "i" and row "A" (herein referred to as cell) corresponding to an "A" nucleotide. The other cells in the same column were entered with a "0". In column "ii", a "1" was entered in the cell corresponding to row "uCG" corresponding to an unmethylated "C" nucleotide immediately followed by a "G" nucleotide. The other cells in column "ii" were entered with a "0". In column "vi", a "1" was entered in the cell corresponding to row "mCG" corresponding to a methylated "C" nucleotide immediately followed by a "G" nucleotide. The other cells in column "vi" were entered with a "0". Based on these rules, a data matrix was constructed containing sequence configurations and methylation patterns associated with methylation haplotypes. 2. Training

[0336] In some embodiments, several data matrices obtained from tumor cells and non-tumor cells, respectively, can be used to train a machine learning model to distinguish between tumor-associated and non-tumor-associated methylation haplotypes. The trained machine learning model can be used to determine the likelihood that a methylation haplotype present in the plasma DNA originates from a tumor cell or a non-tumor cell. Additionally or alternatively, several data matrices obtained from plasma DNA associated with patients with and without cancer, respectively, can be used to train a machine learning model to distinguish between tumor-associated and non-tumor-associated methylation haplotypes.

[0337] In some embodiments, a two-dimensional (2D) matrix of the form [molecule length x 6] was input to a convolutional neural network (CNN). The matrix was passed to a CNN 1D convolutional layer consisting of 128 filters with a kernel size of 10. A rectified linear unit (ReLU) activation function was employed. Then, a max pooling layer with a pool size of 2 and a stride of 2 was applied. This was followed by a bidirectional long short-term memory (LSTM) layer with 32 units. The LSTM may resolve each time point to correspond to the location of a CpG site. Thus, the methylation status of the sequence of CpG sites may be analyzed by the LSTM so that the LSTM is trained to associate the methylation pattern with the presence of a disease. In some instances, a bidirectional LSTM is used. A hyperbolic tangent (tanh) activation function is employed in this layer. The output was then flattened and passed to two dense layers containing 128 and 64 neurons, respectively. A ReLU activation function was employed for the two dense layers. The final layer uses a single neuron with a sigmoid activation function to output probability values ​​indicating the likelihood of a tumor or non-tumor DNA molecule. A higher probability value corresponding to a plasma DNA molecule suggests a higher likelihood that the plasma DNA molecule originates from a tumor.

[0338] In some embodiments, the probability cutoff may be greater than a certain value for detecting tumor-derived plasma DNA molecules, including but not limited to 0.5, 0.6, 0.7, 0.8, and 0.9, etc. The probability cutoff may be less than a certain value for detecting non-tumor-derived plasma DNA molecules, including but not limited to 0.5, 0.4, 0.3, 0.2, and 0.1, etc. In some embodiments, the activation function may include, but is not limited to, rectified linear unit (ReLU), exponential linear unit (ELU), leaky rectified linear unit (Leaky ReLU), parametric rectified linear unit (Parametric ReLU), scaled exponential linear unit (SELU), Gaussian error linear unit (GELU), hyperbolic tangent (tanh) function, sigmoid function, softmax function, swish function, etc.

[0339] In some embodiments, when a model is trained with a data matrix derived from methylation haplotypes from different tissues, including but not limited to neutrophils, T cells, B cells, megakaryocytes, erythrocytes, monocytes, NK cells, liver, lung, esophagus, heart, pancreas, colon, small intestine, adipose tissue, adrenal gland, brain, breast, kidney, bladder, thyroid, prostate, uterus, etc., such a model can be used to identify the tissue / tumor of origin of each plasma DNA molecule based on the methylation haplotype. B. Working Example

[0340] To evaluate the feasibility and potential performance of the above proposed pattern recognition of methylation haplotypes, we simulated various numbers of methylation haplotypes of plasma DNA molecules derived from HCC tumors and buffy coat (i.e., white blood cell) samples, respectively. A probabilistic model was used to simulate the methylation haplotypes of plasma DNA molecules with a certain size (e.g., 2 kb). The methylation status of k CpG sites (k≧1) on a plasma DNA molecule was defined as M=(m 1 , m 2 , …, m k ), where m iis 0 (unmethylated) or 1 (methylated) at CpG site i on a plasma DNA molecule. The probability of M associated with a plasma DNA molecule derived from an HCC tumor may depend on the previous methylation distribution in the HCC tissue. The probability of M associated with a plasma DNA molecule derived from a buffy coat may depend on the previous methylation distribution in the buffy coat.

[0341] The prior methylation distribution of the corresponding CpG sites at 1, 2, …, k in HCC tissues and buffy coat samples follows a beta distribution. The beta distribution is parameterized by two positive parameters α and β, denoted as Beta(α,β). Values ​​derived from the beta distribution range from 0 to 1. Based on the high-depth bisulfite sequencing data of the tissues of interest, the parameters α and β were determined by the number of sequenced cytosines (methylated) and thymines (unmethylated), respectively, at each CpG site in that particular tissue. For HCC tumor tissues, such a beta distribution is defined as Beta(α T ,β T For the buffy coat sample, this beta distribution was defined as Beta(α N ,β N ) respectively. T ,β T ) and Beta(α N ,β N ) were sampled for the methylation status of k CpG sites (k≧1) for tumor-derived and non-tumor-derived plasma DNA molecules. Prior probability distributions for co-methylation and co-unmethylation within a certain nucleotide distance can be integrated into the simulation. For example, 79.6%, 75.6%, 71.6%, 68.6%, 66.4%, 65.1%, 62.5%, 61.1%, and 60.7% of two consecutive CpG sites were found to be co-methylated or co-unmethylated within a nucleotide distance of 5 bp, 10 bp, 20 bp, 30 bp, 40 bp, 50 bp, 100 bp, 200 bp, and 500 bp, respectively.

[0342] Tumor-derived and non-tumor-derived DNA molecules contributing to the plasma DNA pool were simulated at several different depths, 1x, 5x, 10x, 20x, 25x, 30x, 35x, 40x, 50x, 60x, 70x, 80x, 90x, and 100x, across 5,000 randomly selected genomic regions. At a particular depth, data matrices were constructed for tumor-derived and non-tumor-derived DNA molecules, respectively, according to the embodiments in the present disclosure. During the training process, output values ​​corresponding to the data matrix of tumor-derived DNA molecules were labeled with "1". Output values ​​corresponding to the data matrix of non-tumor-derived DNA molecules were labeled with "0". Deep learning models, including CNN and LSTM, were trained using the data matrices related to tumor-derived and non-tumor-derived DNA molecules. The deep learning model parameters were determined by minimizing the prediction error between the predicted output values ​​and the expected output values. The trained models were applied to classify newly simulated plasma DNA that was not used during the training process. The area under the receiver operating characteristic curve (AUC) was used to evaluate model performance at different depths.

[0343] FIG. 64 shows a set of bar graphs 6400 identifying the performance of machine learning models for distinguishing tumor and non-tumor DNA in plasma across different sequencing depths used in the training process. For each sequencing depth (X-axis), blue bars 6402 identify the AUC of the machine learning models based on classifying the training data, and orange bars 6404 identify the AUC of the machine learning models based on classifying the test data. FIG. 64 shows that the performance of distinguishing tumor and non-tumor DNA in plasma improved with increasing sequencing depth used in training. The performance reached a plateau at 70x sequencing depth used in training, with an AUC of 0.90. The performance using the deep learning model based on methylation haplotypes (AUC: 0.90) significantly outperformed the minimal methylation mismatch approach (AUC: 0.8). These data suggested that the proposed pattern recognition of methylation haplotypes is a general and informative approach for detecting cancer-derived cell-free DNA from any genomic region.

[0344] In some embodiments, plasma DNA molecules from differentially methylated regions (DMRs) between tumor and non-tumor genomes are selectively analyzed to further strengthen the model performance. 5,000 DMRs that are highly methylated in tumor genomes compared to buffy coat genomes were obtained (e.g., methylation levels differ by at least 20%). Tumor-derived and non-tumor-derived DNA molecules were simulated at depths of 1x, 5x, 10x, 20x, 25x, 30x, 35x, 40x, 50x, 60x, 70x, 80x, 90x, and 100x for those 5,000 DMRs.

[0345] FIG. 65 shows a set of bar graphs 6500 identifying the performance of machine learning models for distinguishing tumor DNA from non-tumor DNA in plasma, where the machine learning was trained using differentially methylated regions across different sequencing depths. For each sequencing depth (X-axis), blue bars 6502 identify the AUC of the machine learning model based on classifying the training data, and orange bars 6504 identify the AUC of the machine learning model based on classifying the test data. FIG. 65 shows that with increasing sequencing depth used in training, the performance of distinguishing tumor DNA from non-tumor DNA in plasma improved. At a sequencing depth of 30x used in training, the performance reached a plateau with an AUC of 0.91 (FIG. 65), which exceeded non-DMR genomic regions at the same sequencing depth (FIG. 64). In some embodiments, a specificity of 86% and a sensitivity of 81% can be achieved at a sequencing depth of 30x, when 0.5 is used as a cutoff for the probability of classifying a molecule as tumor-derived. In contrast, a specificity of 91% and a sensitivity of 87% can be achieved at a sequencing depth of 100x. The performance using a deep learning model based on methylation haplotypes (AUC: 0.91) significantly outperformed the minimal methylation mismatch approach (AUC: 0.87). These data suggested that in some embodiments, selective analysis of a subset of the genome enhances cancer detection.

[0346] The longer the cell-free DNA molecule, the more CpG sites it contains. More CpG sites in plasma DNA molecules can generally improve the accuracy of identifying the tissue of origin of plasma DNA molecules. Figure 66 shows a table 6600 identifying the performance of machine learning models for differentiating tumor and non-tumor DNA in the plasma of cancer patients with plasma DNA molecules of different lengths. Figure 66 shows that analysis of methylation haplotypes of 200 bp DNA molecules only yielded an AUC of 0.62, whereas analysis of methylation haplotypes of 1 kb plasma DNA molecules improved the AUC to 0.84. Analysis of methylation haplotypes of 5 kb plasma DNA molecules further improved the AUC to 0.98. These results suggested that analysis of long cell-free DNA in cancer patients provides an approach to more accurately identify the tissue of origin / tumor of origin of plasma DNA molecules, and significantly improves the performance of detection and monitoring of cancer or other diseases, including but not limited to autoimmune diseases, organ transplants, trauma, etc. C. Methods for using machine learning models to identify disease classification based on methylation patterns of long cell-free DNA molecules

[0347] 67 shows a flowchart 6700 illustrating an example process of subject biological sample analysis using a machine learning model to identify tissue type characteristics based on methylation patterns of long cell-free DNA molecules, according to some embodiments. The biological sample may include DNA originating from normal cells and possibly cells associated with one or more of a number of tissue types. In addition, at least a portion of the DNA is cell-free in the biological sample.

[0348] In step 6702, sequence reads obtained from methylation-aware sequencing of the cell-free DNA molecules may be received. The methylation-aware sequencing may include an enzyme treatment. In some instances, the methylation-aware sequencing does not include a bisulfite treatment to generate sequence reads for disease classification. In other instances, a bisulfite treatment is used. The bisulfite treatment may be used for the methylation-aware sequencing to generate training data to train the machine learning model. Each of the sequence reads may include a methylation pattern of methylation states at a set of sites (e.g., CpG sites) on the sequence read. For example, a given sequence read may include at least three CpG sites. The methylation pattern may include the number of bases between pairs of sites in the set of sites, and the identity of the bases.

[0349] In some instances, the sequence reads correspond to long cell-free DNA molecules having a size within a first size range, which may include a lower limit and an upper limit. By way of example, the first size range may include an upper limit of at least 1,000 bases, at least 3,000 bases, or more. In some instances, the lower limit may be selected from one of at least 300 bases, at least 400 bases, at least 500 bases, at least 600 bases, or at least 800 bases.

[0350] The set of sites may be of various numbers. In some instances, the set of sites of each sequence read may include at least N sites. For example, a given sequence read may include at least three CpG sites. Other numbers, including but not limited to 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 25, 30, 35, 40, 45, 50, or more than 50 sites, may also be considered. Additionally or alternatively, the sequence read may correspond to a long cell-free DNA molecule having a size within a first size range (e.g., greater than 500 bp) and may include at least N sites (e.g., three CpG sites). The steps of obtaining sequence reads and identifying the methylation state of the sequence reads are further described in step 5702 of FIG. 57.

[0351] Steps 6704 and 6706 may be performed for each sequence read among the sequence reads received from step 6702. At step 6704, a methylation pattern of the sequence read may be input to a machine learning model. In some instances, inputting the methylation pattern into the machine learning model includes inputting a sequence of the sequence read into the machine learning model. Inputting the methylation pattern of the sequence read into the machine learning model may include forming a matrix of the sequence reads, for example, where the matrix includes one-hot encoding of bases and methylation states of a set of sites of the sequence reads. In some instances, the sequence reads may be located (e.g., by aligning the sequence reads to corresponding positions of a reference sequence), where the positions are also input to the machine learning model.

[0352] The machine learning model can be trained using a first training set of sequence reads labeled as being from a first tissue type and a second training set of sequence reads labeled as being from one or more other tissue types. In some instances, the machine learning model includes a convolutional neural network (CNN) and a recurrent neural network (RNN), for example, as described with respect to FIG. 63. In some instances, the first or second training set of sequence reads are obtained from one or more differentially methylated regions (DMRs). The one or more other tissue types can include 1, 2, 3, 4, 5, 10, 15, 20, or more than 20 tissue types. The one or more other tissue types can include, but are not limited to, T cells, B cells, neutrophils, lung tissue, or liver. The one or more other tissue types can include buffy coat.

[0353] At step 6706, a classification of the sequence read may be determined based on the output of the machine learning model. The classification may indicate that the sequence read is derived from (or the degree of derivation from) the first tissue type. The tissue classification may include a probability that the sequence read is derived from the first tissue type. The first tissue type may be a disease tissue type. In some instances, the first tissue type is associated with a disease. A probability of more than one tissue type may be determined.

[0354] In step 6708, the classification of the sequence reads may be used to identify a characteristic of the first tissue type. The characteristic of the first tissue type may identify the amount of sequence reads classified as originating from the first tissue type. In some instances, the characteristic of the first tissue type may identify a disease state of a disease associated with the first tissue type. The disease may be cancer. The characteristic of the first tissue type may further identify a predicted prognosis of the disease associated with the first tissue type. For example, the predicted prognosis may be the presence of vascular invasion associated with cancer.

[0355] In some instances, identifying the characteristic includes (i) identifying a first quantity of sequence reads classified as derived from the first tissue type, and (ii) identifying a disease classification of the biological sample of the first tissue type based on the first quantity. Identifying a disease classification using the first quantity is further described in step 5708 of FIG. VII. Combined analysis of variants and methylation patterns

[0356] To enhance the accuracy of disease classification using long cell-free DNA molecules, methylation pattern analysis of long cell-free DNA molecules may be combined with SNV-based analysis. For example, in a plasma sample, sequence reads may be identified as having mutations (e.g., SNVs) based on comparing the sequence reads with reference sequences, such as reference sequences determined from white blood cells represented in the innate genome. The methylation patterns of the sequence reads leading to such genetic mutations may then be analyzed.

[0357] The method of using combined analysis of SNVs and methylation patterns of long cell-free DNA molecules includes obtaining sequence reads from methylation-aware sequencing of cell-free DNA molecules of a biological sample. Each sequence read includes a methylation pattern, where the methylation pattern identifies the methylation status at a set of CpG sites on the sequence read. The sequence reads of the sequence reads may be aligned to corresponding locations of a reference genome. The sequence reads are then compared to the sequence of the corresponding locations to identify the presence of one or more variants (e.g., single nucleotide variants, single nucleotide polymorphisms, amplifications). The variants may be microsatellite expansions, insertions, deletions, structural mutations, sequence duplications, amplifications, rearrangements, translocations, inversions, and / or microdeletions. In some instances, an SNV of a sequence read is identified if the SNV is detected above a threshold number (e.g., 5 times) of other reads. If one or more variants are identified, the methylation pattern of the sequence read may be further analyzed to identify a disease classification. A. SNVs and methylation patterns

[0358] In some embodiments, the buffy coat DNA and plasma DNA of the patient are sequenced. The buffy coat DNA may be sequenced using, but not limited to, Illumina sequencing. The plasma DNA may be sequenced using, but not limited to, SMRT-seq, so that sequence reads corresponding to long cell-free DNA molecules can be obtained. FIG. 68 shows a schematic diagram 6800 illustrating an example of a combined analysis using SNV and CpG methylation haplotype information, according to some embodiments. As shown in FIG. 68, plasma DNA with alleles (e.g., G nucleotides) that are not present in the buffy coat DNA sequencing results were referred to as somatic mutations. Analysis of the methylation haplotypes of plasma DNA molecules with such somatic mutations can identify the anatomical location of potential cancers.

[0359] In some embodiments, methylation haplotypes associated with cancer signals may be linked to so-called somatic mutations for disease classification. Combination analysis may be used to reduce false positives when using only somatic mutations for disease classification. For example, somatic mutations supported by sequenced plasma DNA molecules identified as tumor of origin (e.g., based on methylation patterns) are more likely to be true mutations compared to somatic mutations supported by sequenced plasma DNA molecules not identified as tumor of origin. Thus, selecting somatic mutations supported by sequenced plasma DNA molecules identified as tumor of origin can increase the positive predictive value of detecting tumor-derived mutations.

[0360] FIG. 69 shows the characteristics 6900 of the first group of plasma DNA molecules with wild-type alleles and the second group of plasma DNA molecules with mutations. For the first group, four long plasma DNA molecules with sizes of 8.9 kb, 3.7 kb, 4.3 kb, and 3.9 kb, respectively, were identified with the wild-type alleles. These four plasma DNA molecules were identified as being of leukocyte origin based on the abundance of their respective methylation states (red bars) across CpG sites. For the second group, three long plasma DNA molecules with sizes of 9 kb, 2.3 kb, and 5.5 kb, respectively, were identified with mutations not present in leukocytes. These three plasma DNA molecules were identified as being of HCC tumor origin based on the abundance of their respective unmethylation states (green bars) across CpG sites. Thus, patients with the second group of plasma DNA molecules could be diagnosed as having HCC based on the synergistic analysis combining SNVs and CpG methylation haplotypes, which was consistent with the clinical diagnosis.

[0361] In some embodiments, the longer the DNA molecule, the more accurate the classification of tumor DNA and non-tumor DNA molecules, since the longer the DNA molecule, the more CpG sites it contains, which facilitates the analysis of the tissue of origin.Thus, the longer the DNA molecule with SNV, the more accurate the tumor localization analysis.For example, for illustrative purposes, the number of CpG sites in the region surrounding the identified somatic mutation from tumor tissue with a size of 200bp and 1kb, etc., was analyzed.In total, 38,465 somatic mutations were analyzed.

[0362] FIG. 70 shows a table 7000 identifying the distribution of the number of CpG sites in 200bp or 1kb regions around somatic mutations. The reference genome is divided into equal-sized regions (e.g., 200bp, 1kb). The corresponding number of CpG sites (e.g., 0, ≧1, ≧10) and the number of these regions with at least one SNV were identified. As shown in FIG. 70, 29.7% of the regions with a size of 200bp had no CpG sites, while only 4.4% of the regions with a size of 1kb had no CpG sites. Furthermore, 5.1% of the regions with a size of 200bp had at least 10 CpG sites, while this percentage increased to 35.7% for the regions with a size of 1kb. These results suggested that the use of long plasma DNA carrying mutations based on methylation haplotypes with more abundant CpG sites is beneficial for identifying the tissue of origin of plasma DNA. Similar conclusions can be drawn for other lengths. Although 90% of the regions 3 kb in size had at least 10 CpG sites, there are only two regions across the genome that are actually SNVs. B. Allelic imbalance between haplotypes and methylation patterns in disease classification

[0363] Cancer cells frequently exhibit copy number aberrations (Chan et al. Proc Natl Acad Sci USA. 2013;110:18761-8; Chan et al. Clin Chem. 2013;59:1,211-224; Zeira and Raphael. Bioinformatics. 2020;36: i344-i352). Such copy number aberrations are generally absent in non-tumor cells. Copy number aberrations include copy number gains and copy number losses.

[0364] For patients with cancer, plasma DNA is a mixture that includes tumor-derived and non-tumor-derived DNA molecules. Variants can cause copy number differences between tumor cells and non-tumor cells. Such differences can lead to distinct differences in the concentration of tumor-derived DNA throughout the human genome. For example, copy number gain regions cause relatively high tumor DNA concentrations, while copy number loss regions cause relatively low tumor DNA concentrations. Copy number gains and losses often occur monoallelic in cancer cells, leading to allelic imbalances (e.g., loss of heterozygosity (LOH)) (Vattathil et al. Genome Res. 2013:23:152-158).

[0365] In other words, variants such as copy number gains and losses generally involve one haplotype block. In some situations, both haplotypes are affected by copy number gains, but the number of amplified blocks between the two haplotype blocks may differ. Thus, the observed amount of plasma DNA molecules between two innate haplotype blocks affected by copy number gains or losses is different. Since genome-wide hypomethylation is frequently observed in cancer (Chan et al. Proc Natl Acad Sci USA. 2013;110:18761-8, Ehrlich. Oncogene. 2002;21: 5400-13), haplotypes with increased contribution from tumor DNA are expected to have lower methylation levels than other haplotypes with decreased contribution from tumor DNA. Thus, relative haplotype methylation imbalance is a new metric that provides information about the presence of cancer. In some embodiments, unequal haplotype methylation levels between haplotypes in cancer cells contribute to such haplotype methylation imbalance in the plasma of cancer patients when cancer-derived DNA molecules are shed into the blood circulation.

[0366] Other types of variants may also be considered in this analysis, including but not limited to microsatellite expansions, insertions, deletions, structural variations, sequence duplications, amplifications, rearrangements, translocations, inversions, and / or microdeletions.

[0367] Figure 71 shows a schematic 7100 of DNA molecules with relative haplotype imbalance with skewed allele ratios and skewed methylation levels that provide information about the presence or absence of cancer. As shown in Figure 71, the resulting skewed allele ratios and methylation levels between two haplotypes may be used to identify the presence of cancer in a patient. In Figure 71, non-tumor cells contain haplotypes I and II (referred to as Hap I and Hap II, respectively). Tumor cells with copy number abnormalities, e.g., copy number gains, contain one haplotype I and three haplotypes II. Plasma DNA molecules are sequenced and assigned to haplotype I and haplotype II, respectively.

[0368] For simplicity, two allelic sites are chosen for illustrative purposes. A higher number of molecules are assigned to Hap II compared to Hap I, resulting in a higher allelic ratio for the C and A alleles on Hap II compared to the T and G alleles on Hap I. CpG sites upstream and downstream of the alleles are analyzed. The CpG sites associated with the C and A alleles are hypomethylated in this case with a methylation level of 20%. Such a methylation level differs from the CpG sites associated with the T and G alleles, which in this case have a methylation level of 75%. The increase in allelic ratio, coupled with the decrease in the methylation level of the CpG sites associated with the alleles in Hap II, reflects copy number gain and hypomethylation, thus providing information on the contribution of plasma DNA from tumor cells.

[0369] In some embodiments, the number of plasma DNA molecules assigned to alleles in the same haplotype block may be counted together to enhance classification power since the increased number of plasma DNA molecules reduces sampling variation. In some instances, the methylation pattern of each of the plasma DNA molecules in the same haplotype block is used to identify disease classification. Statistical approaches for identifying whether haplotype methylation imbalance exists in plasma may include, but are not limited to, sequential probability ratio test, binomial proportion test, Pearson's chi-square test, z-test of two proportions, etc. The number of CpG sites analyzed may include, but is not limited to, ≧3, ≧4, ≧5, ≧6, ≧7, ≧8, ≧9, ≧10, ≧15, ≧20, ≧25, ≧30, ≧35, ≧40, ≧45, ≧50, ≧60, ≧70, ≧80, ≧90, ≧100, ≧200, ≧300, ≧400, ≧500, ≧1000, or other combinations. C. Combined variant and methylation pattern analysis methods for identifying tissue of origin

[0370] FIG. 72 shows a flowchart 7200 illustrating an example process for analyzing a biological sample using variants and methylation patterns to identify tissue of origin based on the methylation pattern of long cell-free DNA molecules, according to some embodiments. The biological sample may include DNA originating from normal cells and possibly cells associated with one or more of multiple tissue types. In addition, at least a portion of the DNA is cell-free in the biological sample.

[0371] At step 7202, sequence reads obtained from methylation-aware sequencing of the cell-free DNA molecules may be received. In some instances, the methylation-aware sequencing does not include bisulfite treatment to generate sequence reads for disease classification. In other instances, bisulfite treatment is used. Each of the sequence reads may include a methylation pattern of methylation states at a set of sites (e.g., CpG sites) on the sequence read. For example, a given sequence read may include at least 15 CpG sites. The methylation pattern may include the number of bases between pairs of sites in the set of sites, and the identity of the bases.

[0372] In some instances, the sequence reads correspond to long cell-free DNA molecules having a size within a first size range, which may include a lower limit and an upper limit. By way of example, the first size range may include an upper limit of at least 1,000 bases, at least 3,000 bases, or more. In some instances, the lower limit may be selected from one of at least 500 bp, 600 bp, 1 kbp, 2 kbp, 3 kbp, 4 kbp, 5 kbp, 6 kbp, 7 kbp, 8 kbp, 9 kbp, 10 kbp.

[0373] The set of sites may be of various numbers. In some instances, the set of sites of each sequence read may include at least N sites. For example, a given sequence read may include at least three CpG sites. Other numbers, including but not limited to at least 3, 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100 sites, may also be considered. Additionally or alternatively, the sequence read may correspond to a long cell-free DNA molecule having a size within a first size range (e.g., greater than 1 kbp) and may include at least N sites (e.g., 10 CpG sites). The steps of obtaining sequence reads and identifying the methylation state of the sequence reads are further described in step 5702 of FIG. 57.

[0374] In step 7204, a location of the first sequence read of the sequence reads may be identified. The location of the first sequence read may be identified by aligning the first sequence read to a reference genome. In some instances, the location of the first sequence read is identified by aligning the first sequence read to a native genome of the subject.

[0375] In step 7206, a variant in the first sequence read corresponding to the position may be detected. The variant in the first sequence read may be a variant relative to the reference sequence at the position. By way of example, the variant may be a polymorphism, a microsatellite expansion, an insertion, a deletion, a structural variation, a sequence duplication, an amplification, a rearrangement, a translocation, an inversion, and / or a microdeletion. An example of a variant that is a single nucleotide polymorphism is shown in FIG. 68.

[0376] In step 7208, the methylation pattern of the first sequence read may be used to identify the tissue of origin of the variant. Identification of the tissue of origin (tissue classification) may use any technique described herein, including those described in Sections IV, V, and VI of this disclosure. For example, the tissue of origin associated with the variant may be identified by comparing the methylation pattern of the first sequence read to one or more reference methylation patterns, as described in steps 5706 and 5708 of FIG. 57. Such descriptions of other methods apply equally to this method. For example, identifying the tissue of origin includes comparing the methylation pattern to a first reference methylation pattern at the location. The first reference methylation pattern may correspond to a disease tissue type of the disease. In some instances, the first reference methylation pattern corresponds to a particular tissue type (e.g., liver). Based on the comparison, the sequence read can be classified as originating from one of a plurality of tissue types.

[0377] The value of the reference pattern may be binary (e.g., 0 and 1 in FIG. 41 or 42) or may have a decimal value (e.g., 0.2 indicating a methylation index of 20%). A closest reference pattern may be identified from a set of reference patterns, and a tissue classification may be determined as the corresponding tissue type of the closest reference pattern. The closest reference pattern may be determined by taking the difference in the methylation state or index at each site relative to the reference pattern. The tissue classification may indicate that the sequence read originates from one of multiple tissue types (or the degree of origin). The tissue classification may include a probability that the sequence read originates from one of multiple tissue types. A probability of more than one tissue type may be determined.

[0378] In some instances, identifying the tissue of origin may include inputting the location and methylation pattern into a machine learning model. The machine learning model may be trained using a first training set of sequence reads labeled as originating from the first tissue type and a second training set of sequence reads labeled as originating from one or more other tissue types. In some instances, the machine learning model includes a convolutional neural network (CNN) and a recurrent neural network (RNN). In some instances, the first or second training set of sequence reads are obtained from one or more differentially methylated regions (DMRs). Based on the output of the machine learning model, identifying whether the sequence reads originate from the first tissue type. D. Methods for combined analysis of variants and methylation patterns to identify cancer classification

[0379] FIG. 73 shows a flowchart 7300 illustrating an example process of analyzing a biological sample using variants and methylation patterns to identify a cancer classification based on the methylation patterns of long cell-free DNA molecules, according to some embodiments. The biological sample may contain DNA originating from normal cells and possibly cells associated with cancer. In addition, at least a portion of the DNA is cell-free in the biological sample.

[0380] At step 7302, sequence reads obtained from methylation-aware sequencing of the cell-free DNA molecules may be received. The methylation-aware sequencing may include an enzyme treatment. In some instances, the methylation-aware sequencing does not include bisulfite treatment. In other instances, bisulfite treatment is used. Each of the sequence reads may include a methylation pattern of methylation states at a set of sites (e.g., CpG sites) on the sequence read. For example, a given sequence read may include at least 15 CpG sites. The methylation pattern may include the number of bases between pairs of sites in the set of sites, and the identity of the bases.

[0381] A single molecule methylation level in a short cell-free DNA molecule may not be statistically sufficient to identify a cancer classification. To address this issue, a long cell-free DNA molecule may be used. In some instances, the sequence reads correspond to long cell-free DNA molecules having a size within a first size range, which may include a lower limit and an upper limit. By way of example, the first size range may include an upper limit of at least 1,000 bases, at least 3,000 bases, or more. In some instances, the lower limit may be selected from one of at least 500bp, 600bp, 1kbp, 2kbp, 3kbp, 4kbp, 5kbp, 6kbp, 7kbp, 8kbp, 9kbp, 10kbp. By sequencing a long DNA molecule, the number of CpG sites in a set of a given long DNA molecule may be large (e.g., at least 5, 10, 20, 50, 100, 200, 500, or 1,000 CpG sites). In this way, the total percentage of methylated sites can be an accurate statistical specification rather than just fragments with only one or two sites.

[0382] The set of sites may be of various numbers. In some instances, the set of sites of each sequence read may include at least N sites. For example, a given sequence read may include at least three CpG sites. Other numbers, including but not limited to at least 3, 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100 sites, may also be considered. Additionally or alternatively, the sequence read may correspond to a long cell-free DNA molecule having a size within a first size range (e.g., greater than 1000 kbp) and may include at least N sites (e.g., 10 CpG sites). The steps of obtaining sequence reads and identifying the methylation state of the sequence reads are further described in step 5702 of FIG. 57.

[0383] In step 7304, a position of the first sequence read of the sequence reads may be identified. The position of the first sequence read may be identified by aligning the first sequence read to a reference genome. In some instances, the position of the first sequence read is identified by aligning the first sequence read to a native genome of the subject. Additionally or alternatively, positions of each of the other sequence reads may also be identified such that a cancer classification can be identified based on sequence reads from the same position of the first sequence read.

[0384] In step 7306, a variant in the first sequence read corresponding to the location may be detected. The variant in the first sequence read may be a variant relative to a reference sequence at the location. The variant may be a microsatellite expansion, insertion, deletion, structural mutation, sequence duplication, amplification, rearrangement, translocation, inversion, and / or microdeletion. In some instances, the variant may be a known tumor marker, such as microsatellite instability (e.g., copy number abnormalities) or a particular sequence variant that is a marker for cancer (e.g., a single nucleotide variant).

[0385] In step 7308, the methylation pattern and variants of the first sequence read may be used to identify a classification of cancer (or other disease or condition). For example, the classification of cancer may be identified based on the methylation level of the methylation pattern, where the methylation level is identified from the methylation state of a set of sites in the first sequence read. The methylation level may be a methylation index, a methylation density, a number of methylated molecules at one or more sites in the set of sites, or a percentage of methylated molecules (e.g., cytosines) at one or more sites in the set of sites. In some instances, the methylation level identifies a methylation percentage of CpG sites in the first sequence read, determined based on the number of methylated CpG sites and the total number of CpG sites in the first sequence read. For example, if a DNA molecule contains 10 CpG sites, of which 5 are identified as methylated, the single molecule methylation level is 50% (i.e., 5 / 10*100%). The variants of the first sequence read can be used as a first indication of whether the corresponding DNA molecule is from a tumor, and the methylation pattern of the sequence read can be used as a second indication of the DNA molecule being from a tumor. If the sequence read has a tumor marker and has a single molecule methylation level below a threshold (e.g., below 20%, 30%, 40%, 50%, or 60%), the classification of the cancer can be cancer present. For example, the decrease occurs as a result of widespread hypomethylation by the cancer.

[0386] Additionally or alternatively, single molecule methylation may be determined to be greater than a threshold, for example, if a particular location in the genome (CpG island) is known to be hypermethylated (e.g., greater than 40%, 50%, 60%, 70%, 80%, 90%, or 95%). In this example, if a sequence read has a tumor marker and has a single molecule methylation level above a threshold (e.g., greater than 40%, 50%, 60%, 70%, 80%, or 90%), the cancer classification may be cancer present.

[0387] In some instances, identifying the cancer classification includes comparing the methylation pattern to a first reference methylation pattern at the location. Thus, instead of single molecule methylation levels, a pattern of methylated or unmethylated sites may be used. For example, a cfDNA molecule with six CpG sites originating from the liver may have a methylation pattern of "-MMMUUU-", where "M" represents a methylated state and "U" represents an unmethylated state. However, other molecules containing the corresponding CpG sites from other tissues may have a methylation pattern of "-MUMUMU-", "-MUMMUU-", "-MUUUMM-", "-MMUUMU-", "-MMUUUM-", "-UMMMUU-", "-UUMUMM-", "-UUMMMU-", "-UUUMMM-". For this example, single molecule methylation levels cannot be used to distinguish liver-derived molecules from other tissues, since all molecules exhibit the same single molecule methylation level with a value of 0.5. In contrast, the methylation pattern across these six CpG sites makes the liver-derived molecule unique compared to the molecules that are derived from other tissues.In this situation, the methylation pattern across a set of CpG sites in the molecule can serve as a "molecular barcode" that indicates the cell nature or disease state, for example, the disease state in the specific tissue type that corresponds to the methylation pattern.The first reference methylation pattern can correspond to a specific tissue type that is associated with cancer.Based on the comparison, the subject can be identified as having cancer.

[0388] The value of the reference methylation pattern may be binary (e.g., 0 and 1 in FIG. 41 or 42) or may have a decimal value (e.g., 0.2 indicating a methylation index of 20%). For example, a closest reference methylation pattern may be identified from a set of reference methylation patterns, and a disease classification may be specified as the disease of the closest reference methylation pattern. The closest reference methylation pattern may be identified by taking the difference in the methylation state or index at each site relative to the reference methylation pattern. Additional details of identifying the closest reference methylation pattern are described at least in process 5700 of FIG. 57 and in Section IV of this disclosure.

[0389] In some instances, identifying the cancer classification includes inputting the location and the methylation pattern into a machine learning model. The machine learning model may be trained using a first training set of sequence reads labeled as being from cancer cells and a second training set of sequence reads labeled as being from normal cells. In some instances, the machine learning model includes a convolutional neural network (CNN) and a recurrent neural network (RNN). In some instances, the first or second training set of sequence reads are obtained from one or more differentially methylated regions (DMRs). Based on the output of the machine learning model, identifying whether the sequence reads are from cancer cells or not.

[0390] Additionally or alternatively, the plurality of DNA molecules can be used to identify a cancer classification. Each DNA molecule of the plurality of DNA molecules can include a variant. The variant can also be a known tumor marker, such as microsatellite instability (copy number abnormality) or a specific sequence variant that is a marker for cancer. Based on their respective methylation patterns, the methylation levels of all of the sequence reads having the tumor variant can be identified based on the methylation status of their respective sets of sites. Such methylation levels can be of only one site or across multiple sites and can occur across multiple regions (e.g., across CpG islands). In some instances, the methylation level includes a ratio of methylated sites to the total number of sites of the sequence read. The methylation level can be compared to a threshold value to identify hypomethylation or hypermethylation of the sequence read. If the methylation level exceeds the threshold value, cancer can be identified as present in the subject. Examples of threshold values ​​are provided above. The threshold value can be identified by examining the methylation levels of a reference sample obtained from a subject with a known cancer classification (e.g., healthy, cancer present). Indeed, the methylation levels determined from variant-bearing cell-free DNA molecules can be used as another indication of whether cancer is present in a subject.

[0391] In yet another embodiment, in another example of the multiple DNA molecules used, the variants may be copy number abnormalities (CNAs), such as deletions or amplifications. Copy number abnormalities can be identified in various ways, for example, by comparing the number or reads in a region to the number in another region (e.g., a region, an average read density across multiple regions, a region on another chromosome(s), or the total number of reads across the genome). The ratio of the number may be compared to a cutoff value to classify whether a CNA is present or not. For a region with a CNA, the aggregate methylation level (e.g., the sum, average, or median methylation level identified for the sequence reads) of one or more sites of sequence reads aligned to the region may be compared to a threshold value. The threshold value may be determined based on the methylation levels of a reference sample obtained from a subject with a known cancer classification (e.g., healthy, cancer present). For example, a genomic region with an amplification generally has a lower methylation level (due to extensive hypomethylation) since there are more fragments from that genomic region compared to another genomic region without a CNA. Thus, if the CNA corresponds to an amplification of sequence reads, then if the aggregate methylation level is below a threshold, then the subject can be classified as having cancer. In another example, if a particular location is known to be hypermethylated in subjects with cancer, then it can be determined whether the methylation level is greater than a threshold. If the CNA corresponds to a deletion of sequence reads in a particular region, then if the aggregate methylation level is greater than a threshold, then the subject can be classified as having cancer. Thus, the methylation pattern of multiple DNA molecules can be used as an additional indication of whether a subject has cancer.

[0392] In yet another embodiment, the haplotype technique may be used on multiple DNA molecules. For example, the allele ratio at one or more SNPs (heterozygous loci) may be determined. When multiple SNPs are used, the allele ratio may be determined based on a first number of sequence reads at one haplotype and a second number of sequence reads at the other haplotype. As will be appreciated by those skilled in the art, the sizes of the DNA fragments of different regions or different haplotypes may also be used. The methylation levels of the regions or haplotypes having an abnormality (e.g., an aggregate of single methylation levels or levels determined across DNA molecules) may then be determined and compared to a threshold. The threshold may be determined based on the methylation levels of reference samples obtained from subjects with known cancer classifications (e.g., healthy, cancer present). If the aggregate methylation level of sequence reads corresponding to a particular haplotype is below the threshold, the subject may be classified as having cancer present if the allele ratio at that position indicates an amplification of the DNA molecule at the particular haplotype. In another example, a subject can be classified as having cancer if the aggregate methylation level of sequence reads corresponding to a particular haplotype is greater than a threshold value if the allele ratio at that position indicates a deletion of the DNA molecule at that particular haplotype. Thus, for deletions, the methylation level is increased due to global methylation or is increased due to regions within the tumor known to have hypermethylation. VIII. Machine learning for disease classification based on multiple features of long CFDNA molecules

[0393] Multiple features of long cfDNA molecules (e.g., methylation patterns, sequence motifs, sequence composition) can be used to identify a disease classification of interest. In particular, machine learning models can be trained using long cfDNA molecules (e.g., sequences greater than 600 bp in size) obtained from training samples with known disease classifications. A. Training

[0394] FIG. 74 shows a schematic diagram 7400 illustrating an example process of training a machine learning model to distinguish patients with and without cancer based on fragment information and a posteriori information present in plasma DNA molecules. As shown in FIG. 74, several plasma DNA samples may be obtained from several patients with and without cancer. Sequence reads of long cell-free DNA molecules of a biological sample may be obtained by single molecule sequencing or cluster-based sequencing. The sequence reads of a given long cell-free DNA molecule may be analyzed to identify a corresponding set of features. Features of each plasma DNA molecule, including but not limited to ends, size, sequence composition, end motifs, methylation haplotypes, ragged ends, genomic coordinates, etc., may be programmed into a data matrix. In some instances, the sequence composition identifies the nucleotide sequence (e.g., 4-mer) of at least a portion of the plasma DNA molecule. The sequence composition may span the entire plasma DNA molecule.

[0395] The data matrix from patients with and without cancer can be used to train a statistical model to classify patients as having or not having cancer, which may include, but is not limited to, linear regression, logistic regression, deep recurrent neural networks (e.g., fully connected recurrent neural networks (RNN), gated recurrent units (GRU), long short-term memory (LSTM)), transformer-based methods (e.g., XLNet, BERT, XLM, RoBERTa), Bayesian classifiers, hidden Markov models (HMM), linear discriminant analysis (LDA), K-means clustering, density-based spatial clustering of applications with noise (DBSCAN), random forest algorithms, adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), and support vector machines (SVM). B.Classification

[0396] The trained model can be used to identify whether a new sample has cancer. FIG. 75 shows a schematic diagram 7500 illustrating an example process of applying the trained model to cancer detection using fragment information and a posteriori information present in plasma DNA molecules. For example, sequence reads may be obtained from a plasma DNA sample, where at least some sequence reads have a length greater than a threshold size (e.g., 600 bp). For each sequence read, one or more features are identified. The one or more features may include, for the sequence read, the location of the ends in a reference genome, the sequence composition, the size, the sequence motif at one or more ends, or the DNA methylation pattern. The features can be input into a trained machine learning model. The machine learning model can generate an output that can be used to identify a classification of the sequence read. The classification can identify whether the sequence read is from a first tissue type or from another tissue type.

[0397] The classification of sequence reads may be analyzed. For example, an amount of sequence reads corresponding to a first tissue type may be identified. If the amount exceeds a cutoff value, the subject may be identified with a disease classification corresponding to the first tissue type. C. Methods for using machine learning models for disease classification based on multiple features of long cell-free DNA molecules

[0398] 76 shows a flowchart 7600 illustrating an example process of analyzing a subject's biological sample to identify a disease classification based on multiple features of long cell-free DNA molecules using a machine learning model, according to some embodiments. The biological sample may include DNA originating from normal cells and possibly cells associated with a disease of a first tissue type. In addition, at least a portion of the DNA is cell-free in the biological sample.

[0399] In step 7602, sequence reads obtained from methylation-aware sequencing of the cell-free DNA molecules may be received. The methylation-aware sequencing may include an enzyme treatment. In some instances, the methylation-aware sequencing does not include a bisulfite treatment to generate sequence reads for disease classification. In contrast, bisulfite treatment may be used for methylation-aware sequencing to generate training data to train a machine learning model. Each of the sequence reads may include a methylation pattern of methylation states at a set of sites (e.g., CpG sites) on the sequence read. The methylation pattern may include the number of bases between pairs of sites in the set of sites, and the identity of the bases.

[0400] In some instances, the sequence reads correspond to long cell-free DNA molecules having a size within a first size range, which may include a lower limit and an upper limit. By way of example, the first size range may include an upper limit of at least 1,000 bases, at least 3,000 bases, or more. In some instances, the lower limit may be selected from one of at least 300 bases, at least 400 bases, at least 500 bases, at least 600 bases, or at least 800 bases.

[0401] The set of sites may be of various numbers. In some instances, the set of sites of each sequence read may include at least N sites. For example, a given sequence read may include at least three CpG sites. Other numbers, including but not limited to at least 3, 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100 sites, may also be considered. Additionally or alternatively, the sequence read may correspond to a long cell-free DNA molecule having a size within a first size range (e.g., greater than 500 bp) and may include at least N sites (e.g., three CpG sites). The steps of obtaining sequence reads and identifying the methylation state of the sequence reads are further described in step 5702 of FIG. 57.

[0402] Steps 7604 and 7606 may be performed for each sequence read among the sequence reads received from step 7602. At step 7604, one or more features of the sequence read may be input to the machine learning model. In some instances, the one or more features include at least one selected from a position of an end in a reference genome, a sequence composition, a size, a sequence motif at one or more ends, and a DNA methylation pattern. For example, the feature may be a sequence composition of the sequence read, where the sequence composition includes a nucleotide base composition and / or a nucleotide base order of the sequence read (as described in Section IB of this disclosure). Another feature may be a position of an end of the sequence read, where identifying the position of the end may include aligning the sequence read to a reference genome. In another example, the feature may be a DNA methylation pattern of the sequence read, where the DNA methylation pattern includes a methylation state at a set of sites on the sequence read (as described in Sections IV, V, and VI of this disclosure).

[0403] The machine learning model was trained using a first training set of sequence reads labeled as being from a first tissue type and a second training set of sequence reads labeled as being from one or more other tissue types. In some instances, the machine learning model includes a convolutional neural network (CNN) and a recurrent neural network (RNN). In some instances, the first or second training set of sequence reads are obtained from one or more differentially methylated regions (DMRs). The machine learning model may be selected from one of a convolutional neural network (CNN), linear regression, logistic regression, deep recurrent neural network (e.g., fully connected recurrent neural network (RNN), gated recurrent unit (GRU), long short-term memory, (LSTM)), transformer-based methods (e.g., XLNet, BERT, XLM, RoBERTa), Bayesian classifier, hidden Markov model (HMM), linear discriminant analysis (LDA), K-means clustering, density-based spatial clustering of applications with noise (DBSCAN), random forest algorithm, adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), support vector machine (SVM), or a composite model including one or more of the above machine learning models.

[0404] The one or more other tissue types may include 1, 2, 3, 4, 5, 10, 15, 20, or more than 20 tissue types. The one or more other tissue types may include, but are not limited to, T cells, B cells, neutrophils, lung tissue, or liver. The one or more other tissue types may include buffy coat.

[0405] At step 7606, a classification of the sequence read may be determined based on the output of the machine learning model. The classification indicates that the sequence read originates from the first tissue type. The tissue classification may include a probability that the sequence read originates from the first tissue type. The first tissue type may be a disease tissue type. In some instances, the first tissue type is associated with a disease.

[0406] In step 7608, an amount of sequence reads classified as originating from the first tissue type may be identified. In some instances, a parameter representative of the amount of sequence reads is determined. The parameter may include a ratio of the amount of sequence reads to an amount of other sequence reads not classified as originating from the first tissue type.

[0407] In step 7610, the amount of sequence reads may be used to identify a disease classification of the biological sample. For example, identifying a disease classification of the biological sample includes comparing the amount to one or more cutoff values, where the one or more cutoffs are determined using a reference sample with a known disease classification. The disease may be cancer. In some instances, identifying a disease classification includes determining the presence or absence of vascular invasion. Using the amount to identify a disease classification is further described in step 5708 of FIG. 57.

[0408] The classification of the disease may include a classification of the severity of the disease (e.g., no disease, early stage, intermediate stage, advanced stage). For example, the classification of the disease may include a stage of cancer according to BCLC stages. The classification may then select one of the stages. Thus, the classification may be specified from multiple stages of the disease (e.g., one of the BCLC stages of HCC). In some instances, the disease is cancer. By way of example, the cancer may be hepatocellular carcinoma, lung cancer, breast cancer, gastric cancer, glioblastoma multiforme, pancreatic cancer, colorectal cancer, nasopharyngeal carcinoma, and head and neck squamous cell carcinoma. IX. Microsatellite Instability

[0409] Microsatellite instability is associated with various cancers, including colon cancer, gastric cancer, ovarian cancer, etc. Microsatellites are tandem repeats of DNA in which a sequence motif of 1-6 nucleotides is repeated multiple times. Figure 77 shows an example of a set of microsatellite sequences 7700 in a DNA molecule.

[0410] These repeat sequences can occur in thousands of regions and have a higher mutation rate than other regions of the human genome (Brinkmann et al. Am J Hum Genet. 1998;62:1408-15). Microsatellite instability (MSI) is thought to occur when DNA mismatch repair (MMR) does not function properly, leading to gain or loss of repeat units in targeted microsatellites and resulting changes in somatic cell size. The widespread instability associated with MMR deficiency indicates a rapid accumulation of somatic mutations that can inactivate genes in key regulatory processes and cause tumor formation. Current research has identified MMR deficiency in many forms of cancer, more often in early stage disease (Le et al. Science. 2017;357:409-413). MSI detection was first performed in colorectal cancer by either PCR against specific markers followed by PAGE and autoradiography (Thibodeau et al. Science. 1993;260:816-819). However, these methods were laborious, time consuming, invasive and had poor dimensional accuracy.

[0411] Subsequently, MSI detection was performed in plasma of small cell lung cancer patients using PCR against specific markers selected for the most frequent microsatellite alterations, followed by gel electrophoresis and autoradiography (Chen et al. Nat Med. 1996; 2(9):1033-5). However, these PCR-based methods limited the application of MSI detection to a limited number of markers and could not be applied to cancer patients carrying other MSI that were not targeted by the PCR primers. They were also laborious, time-consuming, and had low dimensional accuracy.

[0412] Massively parallel next-generation sequencing (NGS) (i.e., short-read sequencing) has also been proposed for the detection of MSI in cancer (Cortes-Ciriano et al. Nat Commun. 2017;8:15180). However, the detectable size range of microsatellites is limited by the read length of the NGS technology, typically 50-150 bp. In addition, the short read nature of NGS and the highly repetitive nature of microsatellites make alignment results prone to inaccuracies, resulting in false positives during MSI analysis.

[0413] Analysis of long plasma DNA molecules in cancer patients provides a more accurate tool to determine the presence or absence of MSI. Using long plasma DNA molecules sequenced by single molecule sequencing, it is possible to obtain unique sequence information of the entire length of the repeat and its vicinity, thereby allowing accurate examination of the genomic location of such repeats and the size of the microsatellite of interest. Since the polymorphic region may be longer than conventional cfDNA (e.g., 160 bp), some of these markers cannot be utilized without analyzing long cfDNA in plasma. In some instances, other tandem repeat polymorphisms (e.g., minisatellites) can also be detected using long cell-free DNA molecules, which can then be used to detect tumor-derived DNA. In addition, the use of long-read sequencing for MSI detection can be applied to cancer patients carrying any MSI (e.g., at a genome-wide level), in contrast to using PCR-based methods described in the literature that are limited to a small number of known MSI markers.

[0414] Figure 78 illustrates an example overview 7800 of detecting tumor-derived DNA based on cancer-specific microsatellite markers. As shown in Figure 78, in some embodiments, microsatellite alterations (CAGs) that are uniquely present in cancer cells but absent in normal cells are detected. 30 In some embodiments, methylation haplotypes associated with microsatellite alterations can be used to localize tumors according to embodiments presented in the present disclosure. X. Treatment A. Treatment options

[0415] In embodiments of the present disclosure, disease recurrence can be accurately predicted, facilitating early intervention and selection of appropriate treatment to improve disease outcome and overall survival of subjects. For example, intensive chemotherapy can be selected for subjects in the event that corresponding samples can predict disease recurrence. In another example, biological samples of subjects who have completed initial treatment can be sequenced to identify viral DNA that can predict disease recurrence. In such an example, an alternative treatment regimen (e.g., higher dose) and / or a different treatment can be selected for the subject, since the subject's cancer may be resistant to the initial treatment.

[0416] The embodiment may also include treating the subject in response to the identification of a classification of recurrence of the condition. For example, if the prediction corresponds to locoregional failure, surgery may be selected as a possible treatment. In another example, if the prediction corresponds to distant metastasis, chemotherapy may additionally be selected as a possible treatment. In some embodiments, the treatment includes surgery, radiation therapy, chemotherapy, immunotherapy, targeted therapy, hormonal therapy, stem cell transplantation, or precision medicine. Based on the identified classification of recurrence, a treatment plan may be developed that reduces the risk of harm to the subject and increases overall survival. The embodiment may further include treating the subject according to the treatment plan. B. Types of Treatment

[0417] The embodiment may further include treating the patient's condition after identifying the classification for the subject. The treatment may be provided according to the identified level of the condition, the fractional concentration of clinically relevant DNA, or the tissue of origin. For example, the identified mutation may be targeted with a specific drug or chemotherapy. The tissue of origin may be used to guide surgery or any other form of treatment. The level of the condition may then be used to determine the aggressiveness of any treatment, which may be determined based on the level of the condition. The condition (e.g., cancer) may be treated by chemotherapy, drugs, diet, therapy, and / or surgery. In some embodiments, the more the value of the parameter (e.g., amount or size) exceeds the reference value, the more aggressive the treatment may be.

[0418] Treatment may include resection. As an example, for bladder cancer, treatment may include transurethral resection of bladder tumor (TURBT). This procedure is used for diagnosis, staging, and treatment. During TURBT, the surgeon inserts a cystoscope through the urethra into the bladder. The tumor is then removed using tools that include small wire loops, lasers, or high-energy electricity. For patients with non-muscle invasive bladder cancer (NMIBC), TURBT may be used to treat or remove the cancer. Another treatment may include radical cystectomy and lymph node dissection. Radical cystectomy is the removal of the entire bladder and possible surrounding tissues and organs. Treatment may also include urinary diversion. Urinary diversion is when the doctor creates a new path for urine to leave the body when removing the bladder as part of treatment.

[0419] Treatment may include chemotherapy, which is the use of drugs that destroy cancer cells, usually by preventing them from growing and dividing. Drugs may include, but are not limited to, for example, mitomycin-C (available as a generic drug), gemcitabine (Gemzar), and thiotepa (Tepadina) for intravesical chemotherapy. Systemic chemotherapy may include, but are not limited to, for example, cisplatin gemcitabine, methotrexate (Rheumatrex, Trexall), vinblastine (Velban), doxorubicin, and cisplatin.

[0420] In some embodiments, the treatment may include immunotherapy. Immunotherapy may include immune checkpoint inhibitors that block a protein called PD-1. Inhibitors may include, but are not limited to, atezolizumab (Tecentriq), nivolumab (Opdivo), avelumab (Bavencio), durvalumab (Imfinzi), and pembrolizumab (Keytruda).

[0421] Therapeutic embodiments may also include targeted therapy. Targeted therapy is a treatment that targets specific genes and / or proteins of cancer that contribute to cancer growth and survival. For example, erdafitinib is an orally administered drug approved for the treatment of people with locally advanced or metastatic urothelial carcinoma with FGFR3 or FGFR2 gene mutations, which causes cancer cells to continue to grow or invade.

[0422] Some treatments may include radiation therapy. Radiation therapy may include using high-energy photons (e.g., x-rays) or other particles to destroy cancer cells. In addition to each individual treatment, a combination of these treatments described herein may be used. In some embodiments, a combination of treatments may be used when the value of a parameter exceeds a threshold value, which itself exceeds a reference value. Information regarding treatments in references is incorporated herein by reference. XI. System Examples

[0423] FIG. 79 illustrates a measurement system 7900 according to an embodiment of the present disclosure. As shown, the system includes a sample 7905, such as cell-free DNA molecules, in an assay device 7910, where an assay 7908 can be performed on the sample 7905. For example, the sample 7905 can be contacted with reagents of the assay 7908 to provide a signal of a physical characteristic 7915. An example of an assay device can be a flow cell including probes and / or primers of an assay or a tube through which droplets (along with droplets containing an assay) travel. The physical characteristic 7915 (e.g., fluorescence intensity, voltage, or current) is detected from the sample by a detector 7920. The detector 7920 can take measurements at regular intervals (e.g., periodic intervals) to obtain data points that constitute a data signal. In one embodiment, an analog-to-digital converter converts the analog signal from the detector to digital form multiple times. The assay device 7910 and the detector 7920 can form an assay system, such as a sequencing system, that performs sequencing according to embodiments described herein. The data signal 7925 is transmitted from the detector 7920 to a logic system 7930. By way of example, the data signal 7925 can be used to identify the sequence and / or location in a reference genome of a DNA molecule. The data signal 7925 can correspond to multiple signals, as the data signal 7925 can include various measurements made simultaneously, e.g., different colors of fluorescent dyes or different electrical signals for different molecules of the sample 7905. The data signal 7925 can be stored in a local memory 7935, an external memory 7940, or a storage device 7945.

[0424] The logic system 7930 may be or include a computer system, ASIC, microprocessor, graphics processing unit (GPU), etc. It may include or be coupled to a display (e.g., monitor, LED display, etc.) and user input devices (e.g., mouse, keyboard, buttons, etc.). The logic system 7930 and other components may be part of a stand-alone or networked computer system, or may be directly attached to or integrated into an apparatus (e.g., a sequencing apparatus) including the detector 7920 and / or assay device 7910. The logic system 7930 may also include software that executes in the processor 7950. The logic system 7930 may include a computer readable medium that stores instructions for controlling the measurement system 7900 to perform any of the methods described herein. For example, the logic system 7930 may provide commands to a system including the assay device 7910 such that sequencing or other physical operations are performed. Such physical operations may be performed in a particular order, such as, for example, adding and removing reagents in a particular order. Such physical manipulations can be performed by a robotic system, including, for example, a robotic arm, so that samples can be obtained and used to perform assays.

[0425] The system 7900 may also include a treatment device 7960 capable of providing a treatment to the subject. The treatment device 7960 may determine a treatment and / or may be used to administer a treatment. Examples of such treatments may include surgery, radiation therapy, chemotherapy, immunotherapy, targeted therapy, hormone therapy, and stem cell transplantation. The logic system 7930 may be connected to the treatment device 7960, for example, to provide results of the methods described herein. The treatment device may receive input from other devices, for example imaging devices and user inputs (e.g., to control a treatment, such as to control a robotic system).

[0426] Any of the computer systems referred to herein may utilize any suitable number of subsystems. An example of such a subsystem is shown in FIG. 80 as computer system 8000. In some embodiments, a computer system includes a single computing device, where the subsystems may be parts of the computing device. In other embodiments, a computer system may include multiple computing devices, each with its own subsystem and internal parts. Computer systems may include desktop and laptop computers, tablets, mobile phones, and other mobile devices.

[0427] The subsystems shown in FIG. 80 are interconnected via a system bus 75. Additional subsystems are shown, such as a printer 74, a keyboard 78, a storage device(s) 79, a monitor 76 (e.g., a display screen such as LEDs, etc.) coupled to a display adapter 82. Peripherals and input / output (I / O) devices coupled to an I / O controller 71 can be connected to the computer system by any number of means known in the art, such as an input / output (I / O) port 77 (e.g., USB, FireWire, etc.). For example, the I / O port 77 or an external interface 81 (e.g., Ethernet, Wi-Fi, etc.) can be used to connect the computer system 8000 to a wide area network such as the Internet, a mouse input device, or a scanner. The interconnection via the system bus 75 allows the central processor 73 to communicate with each subsystem and control the execution of instructions from the system memory 72 or storage device(s) 79 (e.g., fixed disks such as hard drives, or optical disks), and the exchange of information between the subsystems. The system memory 72 and / or the storage device(s) 79 may embody a computer readable medium. Another subsystem is a data collection device 85, such as a camera, microphone, accelerometer, etc. Any of the data mentioned herein can be output from one component to another, or to a user.

[0428] A computer system may include multiple identical components or subsystems connected together, for example, by an external interface 81, an internal interface, or through a removable storage device that can be connected and disconnected from one component to another. In some embodiments, computer systems, subsystems, or devices may communicate over a network. In such an example, one computer may be considered a client and another computer a server, where each may be part of the same computer system. The client and server may each include multiple systems, subsystems, or components.

[0429] Aspects of the embodiments may be implemented in the form of control logic using computer software stored in memory with hardware circuitry (e.g., application specific integrated circuits or field programmable gate arrays) and / or processors that are generally programmable in a modular or integrated fashion, such that a processor may include an FPGA having software instructions and configuration instructions that configure the hardware circuitry or an ASIC. As used herein, a processor may include a single-core processor on the same integrated chip, a multi-core processor, or multiple processing units on a single circuit board or networked, as well as dedicated hardware. Based on the disclosure and teachings provided herein, one of ordinary skill in the art will recognize and understand other ways and / or methods for implementing embodiments of the present disclosure using hardware and combinations of hardware and software.

[0430] Any of the software components or functions described in this application may be implemented as software code executed by a processor using any suitable computer language, such as, for example, Java, C, C++, C#, Objective-C, Swift, or scripting languages, such as Perl or Python, using, for example, conventional or object-oriented techniques. The software code may be stored as a series of instructions or commands on a computer-readable medium for storage and / or transmission. Suitable non-transitory computer-readable media may include random access memory (RAM), read-only memory (ROM), magnetic media, such as hard drives or floppy disks, or optical media, such as compact disks (CDs) or DVDs (digital versatile disks) or Blu-ray disks, flash memory, and the like. The computer-readable medium may be any combination of such devices. In addition, the order of operations may be rearranged. A process may terminate when its operations are completed, but may have additional steps not included in the figures. A process may correspond to a method, a function, a procedure, a subroutine, a subprogram, and the like. When a process corresponds to a function, its termination may correspond to a return of the function to the calling function or to the main function.

[0431] Such a program may be encoded and transmitted using a carrier signal adapted for transmission over wired, optical, and / or wireless networks following various protocols, including the Internet. Thus, a computer-readable medium may be created using a data signal encoded with such a program. A computer-readable medium encoded with program code may be packaged with a compatible device or provided separately from other devices (e.g., via Internet download). Any such computer-readable medium may reside on or within a single computer product (e.g., a hard drive, CD, or an entire computer system), or may reside on or within different computer products in a system or network. A computer system may include a monitor, printer, or other suitable display for providing any of the results mentioned herein to a user.

[0432] Any of the methods described herein may be performed in whole or in part by a computer system including one or more processors that may be configured to perform the steps. Any operation (e.g., alignment, determination, comparison, operation, calculation) performed by the processor may be performed in real time. The term "real time" may refer to an operation or process that is completed within a certain time constraint. The time constraint may be one minute, one hour, one day, or seven days. Thus, the embodiments may relate to a computer system configured to perform any of the steps of the methods described herein, which may include different parts performing individual steps or groups of individual steps. Although shown as numbered steps, the steps of the methods herein may be performed at the same time or different times or in different orders. In addition, some of these steps may be used with some of the other steps from other methods. Also, all or some of the steps may be optional. In addition, any of the steps of any of the methods may be performed using modules, units, circuits, or other means of a system for performing these steps.

[0433] The specific details of the particular embodiments may be combined in any suitable manner without departing from the spirit and scope of the embodiments of the present disclosure, however, other embodiments of the present disclosure may be directed to specific embodiments relating to each individual aspect, or specific combinations of these individual aspects.

[0434] The above description of exemplary embodiments of the present disclosure has been presented for purposes of illustration and description. It is not intended to be exhaustive or to limit the disclosure to the precise form described, and many modifications and variations are possible in light of the above teachings.

[0435] The terms "a," "an," or "the" are intended to mean "one or more," unless expressly stated to the contrary. The use of "or" is intended to mean "inclusive or" rather than "exclusive or," unless expressly stated to the contrary. A reference to a "first" element does not necessarily require that a second element is provided. Additionally, a reference to a "first" or "second" element does not limit the referenced element to a particular location unless expressly stated. The term "based on" is intended to mean "based at least in part on."

[0436] When a group of substitutions is disclosed herein, it is understood that all individual members of the group and all subgroups and classes that can be formed using the substitutions are disclosed separately. When a Markush group or other grouping is used herein, all individual members of the group and all possible combinations and subcombinations of the group are intended to be included separately in the disclosure. As used herein, "and / or" means that one, all, or any combination of the items in the list separated by "and / or" is included in the list, for examp...

Claims

1. 1. A method for analyzing a biological sample from a subject, comprising: receiving sequence reads obtained from methylation-aware sequencing of a plurality of cell-free DNA molecules from the biological sample of the subject, each of the sequence reads comprising a methylation pattern of methylation states at a set of sites on the sequence read; For each of the sequence reads, comparing the methylation pattern to a first reference methylation pattern, the first reference methylation pattern corresponding to a first tissue type of a plurality of tissue types; and identifying a tissue classification of the sequence reads as originating from one of the plurality of tissue types based on a comparison of the methylation pattern to the first reference methylation pattern; and identifying a disease classification of the biological sample based on the tissue classification of the sequence reads.

2. The step of identifying a disease classification comprises: identifying a first quantity of sequence reads classified as originating from the first tissue type; and determining a disease classification of the biological sample based on the first amount.

3. The method of claim 2 , wherein identifying a disease classification of the biological sample based on the first amount comprises comparing the first amount to a cutoff value.

4. The method of claim 3 , wherein the cutoff value is determined based on a reference sample for which the disease classification is known.

5. The method of claim 3 , wherein identifying a classification of the disease comprises applying a machine learning model to the first quantity to generate an output indicative of the classification of the disease.

6. The method of claim 5 , wherein the machine learning model is trained using training samples for which the disease classification is known.

7. For each of the sequence reads, comparing said methylation pattern to one or more other reference methylation patterns corresponding to one or more other tissue types; identifying one or more other quantities of sequence reads classified as originating from one or more other tissue types; The method of claim 1 , further comprising identifying a disease classification of the biological sample based on the one or more other quantities.

8. The method of claim 1 , wherein the first tissue type is a diseased tissue type.

9. The method of claim 1 , wherein the first tissue type is associated with the disease.

10. 10. The method of any one of claims 1 to 9, wherein the disease is cancer.

11. The method of claim 10 , wherein the step of identifying a disease classification comprises determining the presence or absence of vascular invasion of the cancer.

12. 12. The method of claim 1, further comprising identifying a position of the sequence read, wherein the first reference methylation pattern corresponds to the position.

13. 13. The method of claim 1, wherein the methylation pattern comprises the number of bases between site pairs of the set of sites.

14. 14. The method of any one of claims 1 to 13, wherein the tissue classification comprises a probability that the sequence read originates from one of the plurality of tissue types.

15. 15. The method of any one of claims 1 to 14, wherein the classification of the disease identifies the severity of the disease.

16. 16. The method of claim 15, wherein the severity of the disease comprises a stage selected from a plurality of stages of the disease.

17. 17. The method of any one of claims 1 to 16, wherein the first reference methylation pattern comprises a plurality of sites of a reference tissue of a first tissue type, each of the plurality of sites identifying a methylation index at the site.

18. The step of comparing the methylation pattern to the first reference methylation pattern comprises: For each site of the set of sites of the methylation pattern, determining a similarity metric between the methylation state of said site and said methylation index of the corresponding site of said first reference methylation pattern; and determining an aggregate value for the sequence reads based on the determined similarity metric.

19. 20. The method of claim 18, wherein the similarity metric is determined based on the difference between a binary value representing the methylation state of the site and the methylation index of the corresponding site.

20. 20. The method of claim 18 or 19, wherein the aggregate value is a sum, a representative value, or a median value of the determined similarity metrics.

21. one of the plurality of tissue types comprises the first tissue type, and identifying the tissue classification of the sequence reads as originating from one of the plurality of tissue types comprises: determining that the aggregate value determined for the first reference methylation pattern is greater than another aggregate value determined for a second reference methylation pattern, the second reference methylation pattern corresponding to another tissue type of the plurality of tissue types; and identifying the tissue classification of the sequence reads as originating from the first tissue type.

22. wherein one of the plurality of tissue types includes one of the other tissue types of the plurality of tissue types, and identifying the tissue classification of the sequence read as originating from one of the plurality of tissue types comprises: determining that the aggregate value determined for the first reference methylation pattern is less than another aggregate value determined for a second reference methylation pattern, the second reference methylation pattern corresponding to one of the other tissue types; and identifying the tissue classification of the sequence reads as originating from one of the other tissue types.

23. a plurality of reference methylation patterns including the first reference methylation pattern, each of the plurality of reference methylation patterns corresponding to a particular tissue type of the plurality of tissue types, the method comprising: For each of the sequence reads, for each reference methylation pattern of the plurality of reference methylation patterns, comparing said methylation pattern to said reference methylation pattern; and identifying the tissue classification of the sequence reads as originating from one of the plurality of tissue types based on a comparison of the methylation pattern to the reference methylation pattern; 23. The method of any one of claims 1 to 22, further comprising identifying a disease classification of the biological sample based on identifying a maximum amount of sequence reads associated with the tissue classification of the first tissue type.

24. 24. The method of any one of claims 1 to 23, wherein the plurality of tissue types comprises two tissue types, the first tissue type being a diseased tissue type and the second tissue type being a non-diseased tissue type.

25. 25. The method of any one of claims 1 to 24, wherein the disease is cancer.

26. 26. The method of claim 25, wherein the cancer is one of hepatocellular carcinoma, lung cancer, breast cancer, gastric cancer, glioblastoma multiforme, pancreatic cancer, colorectal cancer, nasopharyngeal carcinoma, or head and neck squamous cell carcinoma.

27. 1. A method for analyzing a biological sample from a subject, comprising: receiving sequence reads obtained from methylation-aware sequencing of cell-free DNA molecules from the biological sample of the subject, each of the sequence reads comprising a methylation pattern of methylation states at a set of sites on the sequence read; For each of the sequence reads, inputting the methylation patterns of the sequence reads into a machine learning model trained using a first training set of sequence reads labeled as being from a first tissue type and a second training set of sequence reads labeled as being from one or more other tissue types; and determining a classification of the sequence reads as being from or not from the first tissue type based on an output of the machine learning model; and identifying a characteristic of the first tissue type using the classification.

28. Identifying a characteristic of the first tissue type using the classification includes: identifying a first quantity of sequence reads classified as originating from the first tissue type; and determining a disease classification of the biological sample of the first tissue type based on the first amount.

29. 29. The method of claim 27 or 28, further comprising inputting the sequence reads into the machine learning model.

30. 30. The method of any one of claims 27 to 29, further comprising forming a matrix of one-hot encoding of bases and methylation states.

31. 31. The method of any one of claims 27 to 30, wherein the machine learning models include convolutional neural networks (CNN) and recurrent neural networks (RNN).

32. 32. The method of any one of claims 27 to 31, wherein the first or second training set of sequence reads are obtained from one or more differentially methylated regions (DMRs).

33. 33. The method of any one of claims 27 to 32, further comprising identifying positions of said sequence reads, said positions also being input into said machine learning model.

34. 34. The method of any one of claims 27 to 33, wherein the methylation pattern comprises the number of bases between site pairs of the set of sites.

35. 35. The method of any one of claims 27 to 34, wherein the characteristic of the first tissue type identifies a quantity of sequence reads classified as originating from the first tissue type.

36. 35. The method of any one of claims 27 to 34, wherein the characteristic of the first tissue type identifies a disease state of a disease associated with the first tissue type.

37. 37. The method of claim 36, wherein the disease is cancer.

38. 37. The method of claim 36, wherein said characteristic of said first tissue type further identifies a predictive prognosis of said disease associated with said first tissue type.

39. the disease is cancer, 39. The method of claim 38, wherein the predicted prognosis comprises the presence of vascular invasion associated with the cancer.

40. 40. The method of any one of claims 27 to 39, wherein the one or more other tissue types comprise T cells, B cells, neutrophils, lung tissue, or liver.

41. 41. The method of any one of claims 1-40, wherein the set of sites for each sequence read comprises at least three sites.

42. 1. A method for analyzing a biological sample from a subject, comprising: receiving sequence reads obtained from methylation-aware sequencing of cell-free DNA molecules from the biological sample of the subject, each of the sequence reads comprising a methylation pattern of methylation states at a set of sites on the sequence read; identifying a position of the first sequence read; detecting a variant in the first sequence read that corresponds to the location; and identifying a tissue of origin of the variant using the methylation pattern of the first sequence read.

43. The step of identifying the tissue of origin comprises: comparing said methylation pattern to a first reference methylation pattern at said location, said first reference methylation pattern corresponding to a disease tissue type of a disease; and classifying the first sequence read as originating from one of a plurality of tissue types based on a comparison of the methylation pattern to the first reference methylation pattern.

44. The step of identifying the tissue of origin comprises: inputting the locations and the methylation patterns into a machine learning model trained using a first training set of sequence reads labeled as being from a first tissue type and a second training set of sequence reads labeled as being from one or more other tissue types; and determining whether the first sequence read is from the first tissue type based on an output of the machine learning model.

45. 45. The method of any one of claims 42 to 44, wherein the variant is a microsatellite expansion, insertion, deletion, structural variation, sequence duplication, amplification, rearrangement, translocation, inversion, and / or microdeletion.

46. 1. A method for analyzing a biological sample from a subject, comprising: receiving sequence reads obtained from methylation-aware sequencing of cell-free DNA molecules from the biological sample of the subject, each of the sequence reads comprising a methylation pattern of methylation states at a set of sites on the sequence read; identifying a position of the first sequence read; detecting a variant in the first sequence read that corresponds to the location; and identifying a cancer classification using the methylation pattern and the variants of the first sequence reads.

47. wherein the variants are associated with a cancer of known classification and the methylation pattern is a methylation level, and identifying the cancer classification comprises: determining the methylation level of the first sequence read based on a methylation state of the set of sites in the first sequence read; comparing the methylation level of the first sequence read to a threshold, the threshold being determined based on the methylation level of a reference sample having a known cancer classification; and identifying a classification of the cancer based on the variant and a determination that the methylation level exceeds the threshold.

48. identifying a position of each of the sequence reads; determining a plurality of sequence reads from the sequence reads, wherein each sequence read of the plurality of sequence reads is from a same position of the first sequence read and comprises the variant, and the plurality of sequence reads comprises the first sequence read; The step of identifying a cancer classification comprises: determining an aggregate methylation level of the plurality of sequence reads; comparing the aggregate methylation level of the plurality of sequence reads to a threshold, the threshold being determined based on methylation levels of a reference sample with known cancer classification; and identifying a classification of the cancer based on the variants and a determination that the aggregate methylation level exceeds the threshold.

49. Determining an aggregate methylation level of the plurality of sequence reads includes: determining, for each sequence read of the plurality of sequence reads, a methylation level based on the methylation state of the set of sites in the sequence read; 49. The method of Claim 48, comprising determining an aggregate value based on methylation levels of said plurality of sequence reads, said aggregate value being said aggregate methylation level.

50. 50. The method of claim 49, wherein the aggregate value is a representative, sum, or median of the identified methylation levels of the plurality of sequence reads.

51. 50. The method of Claim 49, wherein said locations of said plurality of sequence reads are locations of a particular haplotype.

52. 52. The method of claim 51 , wherein the variant is an amplification or deletion of a DNA molecule at the particular haplotype.

53. 49. The method of claim 47 or 48, wherein determining that the methylation level exceeds the threshold comprises identifying the methylation level as being below the threshold.

54. 49. The method of claim 47 or 48, wherein hypermethylation at said location is associated with said known classification of cancer, and determining that said methylation level exceeds said threshold comprises identifying said methylation level as being higher than said threshold.

55. 47. The method of claim 46, wherein the cancer is identified as being of a particular tissue type.

56. The step of identifying a cancer classification comprises: inputting the locations and the methylation patterns of the first sequence reads into a machine learning model trained using a first training set of sequence reads labeled as being from cancer cells and a second training set of sequence reads labeled as being from normal cells; and determining whether the first sequence read is derived from the cancer cell based on an output of the machine learning model.

57. 57. The method of any one of claims 46 to 56, wherein the methylation pattern is a global methylation level of the biological sample.

58. 58. The method of any one of claims 46 to 57, wherein identifying a classification of the cancer comprises comparing the methylation pattern to a reference methylation pattern associated with the cancer.

59. 59. The method of any one of claims 46 to 58, wherein the variant is a microsatellite expansion, insertion, deletion, structural mutation, sequence duplication, amplification, rearrangement, translocation, and / or inversion.

60. 60. The method of any one of claims 42 to 59, wherein the first sequence read has a size within a size range, the lower limit of the size range being at least one of 500 bp, 600 bp, 1 kbp, 2 kbp, 3 kbp, 4 kbp, 5 kbp, 6 kbp, 7 kbp, 8 kbp, 9 kbp, or 10 kbp.

61. 61. The method of any one of claims 42-60, wherein the set of sites of the first sequence reads comprises at least three sites.

62. 62. The method of any one of claims 42 to 61, wherein identifying the location of the first sequence read comprises aligning the first sequence read to a reference sequence, and wherein the variant in the first sequence read is relative to the reference sequence at the location.

63. 62. The method of claims 42-61, wherein identifying the location of the first sequence read comprises aligning the first sequence read to a reference sequence, and wherein the variants in the first sequence read are relative to the subject's native genome.

64. 64. The method of any one of claims 1 to 63, wherein the methylation-aware sequencing does not include bisulfite treatment.

65. 64. The method of any one of claims 1 to 63, wherein the methylation-aware sequencing comprises bisulfite treatment.

66. 1. A method of analyzing a biological sample from a subject, the biological sample comprising DNA originating from normal cells and, optionally, DNA originating from cells associated with cancer, at least a portion of the DNA being acellular in the biological sample; measuring the size of a plurality of cell-free DNA molecules from the biological sample; measuring a first amount of cell-free DNA molecules having a size within a first size range, the upper limit of the first size range being at least 1,000 bases; generating a normalized parameter value using the first quantity; and identifying a classification of a level of cancer using the normalized parameters.

67. 67. The method of claim 66, wherein the first size range includes a lower limit that is greater than zero.

68. receiving sequence reads obtained from sequencing the plurality of cell-free DNA molecules from the biological sample; identifying, for each of the sequence reads, a sequence motif for each of one or more end sequences of a corresponding cell-free DNA molecule; and identifying a group of the plurality of cell-free DNA molecules that have at least one of the set of one or more sequence motifs in a termination sequence, 68. The method of claim 66 or 67, wherein the first amount is identified from a subgroup of the group of the plurality of cell-free DNA molecules, the subgroup of cell-free DNA molecules having a size within the first size range.

69. 69. The method of claim 68, wherein the set of one or more sequence motifs comprises CCCA.

70. 68. The method of claim 67, wherein the lower limit is selected from one of at least 300 bases, at least 400 bases, at least 500 bases, at least 600 bases, or at least 800 bases.

71. 71. The method of any one of claims 66 to 70, wherein the upper end of the first size range is at least 3,000 bases.

72. 72. The method of any one of claims 66 to 71, wherein the classification of the level of cancer comprises a classification of the histological state of the cancer.

73. 73. The method of claim 72, wherein the histological state of the cancer comprises the presence of vascular invasion of the cancer.

74. Measuring the size of a plurality of cell-free DNA molecules from the biological sample includes: performing single molecule real-time sequencing or nanopore sequencing of the plurality of cell-free DNA molecules; and counting the number of nucleotides in each cell-free DNA molecule of the plurality of cell-free DNA molecules.

75. 76. The method of any one of claims 66-75, wherein the normalized parameter comprises the relative frequency of the cell free DNA molecules having a size within the first size range.

76. 76. The method of any one of claims 66-75, wherein the normalized parameter is determined using a second amount of cell free DNA molecules having a size within a second size range, wherein the first size range and the second size range are different.

77. 77. The method of any one of claims 66 to 76, wherein the step of identifying a classification of the level of cancer comprises comparing the normalized parameter with a cutoff value.

78. 78. The method of claim 77, wherein the cutoff value is determined using a reference sample with known cancer classification.

79. 78. The method of claim 77, wherein comparing the normalized parameter to a cutoff value comprises comparing the normalized parameter to the cutoff value using a machine learning model, wherein the cutoff value is determined using one or more training datasets comprising reference samples with known classification of levels of cancer.

80. 1. A method for analyzing a biological sample from a subject, comprising: receiving sequence reads obtained from sequencing cell-free DNA molecules from the biological sample of the subject; identifying, for each of the sequence reads, a sequence motif for each of one or more end sequences of a corresponding cell-free DNA molecule; For each of a set of N sequence motifs, determining the relative frequency of N sequence motifs; using said N relative frequencies to generate a vector of N frequencies, each normalized to one another or to other frequencies of sequence motifs in a group of reference samples; comparing said N frequency vectors to a plurality of reference vectors determined from a group of reference samples of known disease classification; and identifying a disease classification of the biological sample based on a comparison of the N frequency vectors to the plurality of reference vectors.

81. 81. The method of claim 80, wherein the frequency of the vector of N frequencies is a z-score.

82. 82. The method of claim 80 or 81, wherein comparing the N frequency vectors to a plurality of reference vectors comprises applying a machine learning model to the plurality of reference vectors to generate a set of clusters, each cluster of the set of clusters comprising one or more reference vectors of the plurality of reference vectors.

83. The step of identifying the classification includes: determining a first distance between the N frequency vectors and a closest reference vector of a first cluster of reference vectors of the set of clusters, the first cluster of reference vectors representing a first subgroup of the group of reference samples classified as having the disease; determining a second distance between the N frequency vectors and a closest reference vector of a second cluster of reference vectors of the set of clusters, the second cluster of reference vectors representing a second subgroup of the group of reference samples classified as not having the disease; determining that the first distance is less than the second distance; and identifying the subject as not having the disease based on a determination that the first distance is greater than the second distance.

84. The step of identifying the classification includes: determining a first distance between the N frequency vectors and a closest reference vector of a first cluster of reference vectors of the set of clusters, the first cluster of reference vectors representing a first subgroup of the group of reference samples classified as having the disease; determining a second distance between the N frequency vectors and a closest reference vector of a second cluster of reference vectors of the set of clusters, the second cluster of reference vectors representing a second subgroup of the group of reference samples classified as not having the disease; determining that the first distance is less than the second distance; and identifying the subject as having the disease based on a determination that the first distance is less than the second distance.

85. 85. The method of any one of claims 80-84, wherein the relative frequency is determined based on a proportion of cell-free DNA molecules having an end sequence corresponding to the sequence motif relative to the cell-free DNA molecules from the biological sample.

86. 85. The method of any one of claims 80-84, wherein the relative frequency is determined based on the proportion of cell-free DNA molecules having an end sequence that corresponds to the sequence motif relative to the number of cell-free DNA molecules having end sequences that correspond to other sequence motifs of the set of N sequence motifs.

87. 87. The method of any one of claims 80 to 86, wherein the relative frequency is a ranking of the sequence motifs.

88. 88. The method of claim 87, wherein the rank of the sequence motif is determined based on the number of the cell-free DNA molecules that have an end sequence that corresponds to the sequence motif.

89. 89. The method of any one of claims 80 to 88, wherein generating the N frequency vector comprises normalizing the relative frequency of the sequence motif using other frequencies of the sequence motif within the group of reference samples.

90. 90. The method of claim 89, wherein each frequency in the vector of N frequencies is determined by comparing the relative frequency to the average frequency of the sequence motif within the group of reference samples.

91. 1. A method for analyzing a biological sample from a subject, comprising: receiving sequence reads obtained from sequencing cell-free DNA molecules from the biological sample of the subject; using the sequence reads to size the cell-free DNA molecules; determining, for each of the sequence reads, a sequence motif for each of one or more end sequences of a corresponding cell-free DNA molecule; for a first set of cell-free DNA molecules having a first size range, determining a first relative frequency of occurrence of one or more sequence motifs within the first set of cell-free DNA molecules; determining a second relative frequency of occurrence of one or more sequence motifs within the second set of cell-free DNA molecules for a second set of the cell-free DNA molecules having a second size range, the second size range having an upper limit that is greater than the upper limit of the first size range; determining a separation value between the first relative frequency and the second relative frequency; and using the separation value to identify a disease classification.

92. 92. The method of claim 91, wherein identifying the classification comprises comparing the separation value to one or more cutoff values.

93. 93. The method of claim 92, wherein a cut-off value of the one or more cut-off values ​​is determined using a reference sample with a known classification of the disease.

94. 93. The method of claim 92, wherein the cutoff value of the one or more cutoff values ​​is one of 0.6, 0.65, 0.7, or 0.

75.

95. 94. The method of any one of claims 91 to 93, wherein the classification of the disease comprises a classification of the severity of the disease.

96. 96. The method of claim 95, wherein the severity of the disease comprises a stage selected from a plurality of stages of the disease.

97. 97. The method of any one of claims 91 to 96, wherein the first size range includes an upper limit.

98. 98. The method of claim 97, wherein the upper limit is selected from one of at least 80 bases, at least 100 bases, at least 150 bases, at least 200 bases, or at least 300 bases.

99. 99. The method of any one of claims 91 to 98, wherein the second size range includes a lower limit.

100. 100. The method of claim 99, wherein the lower limit is selected from one of at least 300 bases, at least 400 bases, at least 500 bases, at least 600 bases, or at least 800 bases.

101. 101. The method of any one of claims 91 to 100, wherein the separation value is a ratio value of the first relative frequency and the second relative frequency.

102. 102. The method of any one of claims 91 to 101, wherein using the separation values ​​to identify a classification of a disease comprises applying a machine learning model to the separation values ​​to generate an output indicative of the classification of the disease.

103. The method of claim 102, wherein the machine learning model is trained using training samples for which the disease classification is known.

104. 103. The method of claim 102, wherein the machine learning model is selected from one of the following: a convolutional neural network (CNN), a linear regression, a logistic regression, a deep recurrent neural network (e.g., a fully connected recurrent neural network (RNN), a gated recurrent unit (GRU), a long short-term memory (LSTM)), a transformer-based method (e.g., XLNet, BERT, XLM, RoBERTa), a Bayesian classifier, a hidden Markov model (HMM), a linear discriminant analysis (LDA), K-means clustering, density-based spatial clustering of applications with noise (DBSCAN), a random forest algorithm, adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), a support vector machine (SVM), or a composite model comprising one or more of the above machine learning models.

105. 105. The method of any one of claims 91-104, wherein the first relative frequency is a ratio of the first set of cell free DNA molecules to the cell free DNA molecules from the biological sample.

106. 105. The method of any one of claims 91-104, wherein the first relative frequency is a ratio of the first set of cell-free DNA molecules to a number of cell-free DNA molecules having other sequence motifs.

107. 1. A method for analyzing a biological sample from a subject, comprising: receiving sequence reads obtained from sequencing the cell-free DNA molecules; For each of the sequence reads, inputting one or more features of the sequence reads into a machine learning model trained using a first training set of sequence reads labeled as being from a first tissue type and a second training set of sequence reads labeled as being from one or more other tissue types; and determining a tissue classification of whether the sequence reads originate from the first tissue type based on an output of the machine learning model; identifying a first quantity of sequence reads classified as originating from the first tissue type; and identifying a first disease classification of the biological sample of the first tissue type based on the first amount.

108. 108. The method of claim 107, wherein the one or more features comprise at least one selected from the position of a terminus in a reference genome, a sequence composition, a size, a sequence motif at one or more termini, and a DNA methylation pattern.

109. 109. The method of claim 108, wherein the one or more features comprise a sequence configuration of the sequence read, the sequence configuration comprising a nucleotide base composition and / or a nucleotide base order of the sequence read.

110. 109. The method of Claim 108, wherein the one or more features comprise positions of termini of the sequence reads, and wherein identifying the positions of the termini comprises aligning the sequence reads to the reference genome.

111. 109. The method of Claim 108, wherein said one or more features comprise a DNA methylation pattern of said sequence reads, said DNA methylation pattern comprising a methylation state at a set of sites on said sequence reads.

112. 112. The method of any one of claims 107 to 111, wherein the first tissue type is a diseased tissue type.

113. 113. The method of any one of claims 107 to 112, wherein the first tissue type is associated with the disease.

114. 114. The method of any one of claims 107 to 113, wherein the disease is cancer.

115. 115. The method of claim 114, wherein identifying the disease classification comprises determining the presence or absence of vascular invasion of the cancer.

116. 116. The method of any one of claims 107 to 115, wherein the cancer is one of hepatocellular carcinoma, lung cancer, breast cancer, gastric cancer, glioblastoma multiforme, pancreatic cancer, colorectal cancer, nasopharyngeal carcinoma, or head and neck squamous cell carcinoma.

117. 117. The method of any one of claims 107 to 116, wherein the machine learning model is selected from one of the following: Convolutional Neural Networks (CNN), Linear Regression, Logistic Regression, Deep Recurrent Neural Networks (e.g., Fully Connected Recurrent Neural Networks (RNN), Gated Recurrent Units (GRU), Long Short-Term Memory, (LSTM)), Transformer-based methods (e.g., XLNet, BERT, XLM, RoBERTa), Bayesian classifiers, Hidden Markov Models (HMM), Linear Discriminant Analysis (LDA), K-Means Clustering, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), Random Forest Algorithm, Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), Support Vector Machines (SVM), or a composite model comprising one or more of the above machine learning models.

118. 118. The method of any one of claims 107 to 117, wherein identifying a disease classification of the biological sample comprises comparing the first amount to one or more cutoff values.

119. 119. The method of claim 118, wherein a cut-off value of the one or more cut-off values ​​is determined using a reference sample with a known classification of the disease.

120. 1. A method for analyzing a biological sample from a subject, comprising: receiving sequence reads obtained from methylation-aware sequencing of cell-free DNA molecules from the biological sample of the subject, each of the sequence reads comprising a methylation state at a set of sites in the sequence read; For each of the sequence reads, identifying the methylation state of each of the set of sites in the sequence reads; and determining a methylation level of the sequence reads based on the methylation state; determining a statistical value based on the determined methylation levels of the sequence reads; comparing the statistical value to a reference value to identify a level of classification of the subject's condition.

121. 121. The method of claim 120, wherein the methylation-aware sequencing comprises single molecule sequencing.

122. 121. The method of claim 120, wherein the methylation recognition sequencing comprises nanopore sequencing.

123. 123. The method of any one of Claims 120 to 122, wherein determining a methylation level of the sequence reads comprises determining a number of one or more sites that are methylated in the set of sites.

124. 124. The method of any one of claims 120 to 123, wherein the statistical value is a representative or median value of the identified methylation levels of the sequence reads.

125. 125. The method of any one of claims 120 to 124, wherein the condition comprises cancer.

126. 126. The method of claim 125, wherein the level of classification of the pathology comprises a stage of cancer.

127. 127. The method of any one of claims 120 to 126, wherein each of the sequence reads has a size within a first size range.

128. 128. The method of claim 127, wherein the first size range includes a lower limit.

129. 129. The method of claim 128, wherein the lower limit is selected from one of at least 300 bases, at least 400 bases, at least 500 bases, at least 600 bases, or at least 800 bases.

130. 130. The method of any one of claims 127 to 129, wherein the first size range includes an upper limit.

131. 131. The method of claim 130, wherein the upper end of the first size range is at least 3,000 bases.

132. 132. The method of any one of claims 120 to 131, wherein the set of sites comprises at least three sites.

133. 133. The method of any one of claims 120 to 132, wherein the reference value is determined using a reference sample for which the classification of the pathology is known.

134. A computer product comprising a non-transitory computer readable medium storing a plurality of instructions which, when executed, cause a computer system to perform the method of any one of the preceding claims.

135. A computer product according to claim 134; and one or more processors for executing instructions stored on the computer-readable medium.

136. A system comprising means for performing any of the above methods.

137. A system comprising one or more processors configured to perform any of the above methods.

138. A system including modules each performing the steps of any of the above methods.

Images