Fragmentation to measure methylation and disease

Abstract

Fragmentation of cell-free DNA molecules can be measured and used for various purposes, including, for example, determining methylation at specific sites of DNA molecules, at specific genomic sites in a reference genome of a biological sample of cell-free DNA of a subject (e.g., plasma, serum, urine, saliva), or for specific regions in a reference genome of a biological sample (also simply referred to as a sample). Various types of fragmentation measurements can be used, such as terminal motifs and cleavage profiles. Another purpose is to determine the fractional concentration of DNA (e.g., clinically relevant DNA) of a particular tissue type. Another purpose is to determine a pathology of a subject using a biological sample containing cell-free DNA. The cell-free DNA can be that of the subject or that of a pathogen (e.g., a virus) in the subject's sample. For a particular tissue type, hypermethylated, hypomethylated, 5hmC enriched, and 5hmC depleted sites / regions can be used.

Description

[Technical Field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority from and is a non-provisional application of U.S. Provisional Patent Application No. 63 / 307,622 entitled "Fragmentation For Measuring Methylation And Disease," filed February 7, 2022, U.S. Provisional Patent Application No. 63 / 328,710 entitled "Fragmentation For Measuring Methylation And Disease," filed April 7, 2022, and U.S. Provisional Patent Application No. 63 / 400,244 entitled "Fragmentation For Measuring Methylation And Disease," filed August 23, 2022, the entire contents of which are incorporated herein by reference for all purposes. [Background technology]

[0002] Cell-free DNA (cfDNA) molecules are fragmented nonrandomly. Several studies have demonstrated a correlation between cfDNA fragmentation patterns and nucleosome structure (Sun et al. Proc Natl Acad Sci UAS. 2018;115:E5106, Snyder et al. Cell. 2016;164:57-68). For example, the characteristic size profile of cfDNA shows a modal frequency at approximately 166 bp, with smaller molecules forming a series of peaks with a periodicity of 10 bp (Lo et al. Sci Transl Med. 2010;2:61ra91). Such size patterns of plasma DNA fragments suggest the presence of both internucleosomal and intranucleosomal cleavage during the release of DNA molecules into the blood circulation during cell death (e.g., via apoptosis) or other mechanisms. Furthermore, our group previously reported that a subset of genomic locations was found to be preferentially cleaved during the generation of plasma DNA molecules (Chan et al. Proc Natl Acad Sci USA. 2016;113:E8159-E8168; Jiang et al. Proc Natl Acad Sci USA. 2018;115:E10925-E10933), and such preferred cleavage may reflect the tissue of origin of cfDNA (Jiang et al. Proc Natl Acad Sci USA. 2018;115:E10925-E10933; Sun et al. Proc Natl Acad Sci USA. 2018;115:E5106-E5114). Currently, widely used strategies for methylation analysis are based on bisulfite or enzymatic conversion followed by sequencing. Such conversions, which involve extra steps of DNA purification or DNA degradation under harsh conditions, result in information loss of the DNA molecules. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Sun et al.Proc Natl Acad Sci UAS.2018;115:E5106 [Non-patent document 2] Snyder et al.Cell.2016;164:57-68 [Non-patent document 3] Chan et al.Proc Natl Acad Sci USA.2016;113:E8159-E8168 [Non-patent document 4] Jiang et al.Proc Natl Acad Sci USA.2018;115:E10925-E10933 [Non-Patent Document 5] Sun et al.Proc Natl Acad Sci US A.2018;115:E5106-E5114 Summary of the Invention

[0004] Various embodiments are provided for using fragmentation of cell-free DNA molecules for various purposes. One exemplary purpose is, for example, to determine methylation at a specific site in a DNA molecule, at a specific genomic site in a reference genome of a subject's cell-free DNA biological sample (e.g., plasma, serum, urine, saliva), or for a specific region in a reference genome of a biological sample (also simply referred to as a sample). Various types of fragmentation measurements can be used for this purpose, such as terminal motifs and cleavage profiles, which can include cleavage information at one or more positions surrounding a site, such as a CpG site.

[0005] Another exemplary purpose is to determine the fractional concentration of DNA (e.g., clinically relevant DNA) in a particular tissue type. Various types of fragmentation measurements, such as terminal motifs and cleavage profiles, can be used for this purpose. Certain sites can be selected for use in fragmentation measurements. For example, sites / regions that are hypermethylated and / or hypomethylated for a particular tissue type can be used. As another example, sites that are 5hmC enriched or depleted for a particular tissue type can be used. In addition, techniques can use fragmentation measurements (e.g., terminal motifs and cleavage profiles) to identify 5hmC-enriched and 5hmC-depleted regions for a particular tissue.

[0006] Another exemplary purpose is to determine the pathology of a subject using a biological sample containing cell-free DNA. The cell-free DNA can be that of the subject or that of a pathogen (e.g., a virus) in the subject's sample. Various types of fragmentation measurements, such as terminal motifs and cleavage profiles, can be used for this purpose. Certain sites can be selected for use in the fragmentation measurement. For example, sites / regions that are hypermethylated and / or hypomethylated for a specific tissue type can be used to identify pathology for a specific tissue type. As another example, sites / regions that are similarly methylated but have different cleavage profiles between subjects with and without pathology can be used. As yet another example, sites / regions that are 5hmC enriched or depleted for a specific tissue type can be used to identify pathology for a specific tissue type. Machine learning models can use such tissue-specific methylation patterns (e.g., hypermethylation and hypomethylation, and 5hmC enrichment and depletion) to distinguish between different types of cancer. The cell-free DNA fragments used for the fragmentation measurement can be filtered, for example, based on size. Therefore, pathologies can be detected based on the amount of terminal motifs in cell-free DNA fragments with specific sizes.

[0007] These and other embodiments of the present disclosure are described in detail below. For example, other embodiments are directed to systems, devices, and computer-readable media related to the methods described herein.

[0008] A better understanding of the nature and advantages of embodiments of the present disclosure may be obtained with reference to the following detailed description and accompanying drawings. [Brief explanation of the drawings]

[0009] [Figure 1] 1 illustrates cleavage positions relative to CpG sites (also referred to as CG sites) according to embodiments of the present disclosure. [Figure 2A] 1 shows that the methylation index varies depending on the distance between the 5′ end and the CpG site, according to an embodiment of the present disclosure. [Figure 2B] Same as above. [Figure 3] 1 shows the percentage of methylated CpG sites covered by multiple tetramer end motifs and fragments with a particular tetramer end motif, according to an embodiment of the present disclosure. [Figure 4] 1 shows an example of a method using droplet PCR to determine NCG and CGN motifs in a region of interest, according to an embodiment of the present disclosure. [Figure 5] 1 is an exemplary workflow for sequencing library preparation by selectively amplifying DNA containing NCG or CGN terminal motifs, according to an embodiment of the present disclosure. [Figure 6] FIG. 1 shows a schematic diagram of methylation status inference based on fragmentation patterns, according to an embodiment of the present disclosure. [Figure 7] 1 illustrates the definition of a cut profile using a cut ratio, according to an embodiment of the present disclosure. [Figure 8] 1 shows a comparison of cleavage profiles between methylated and unmethylated CpG sites according to embodiments of the present disclosure. [Figure 9]1 shows cleavage profiles of windows associated with different sequence contexts, including CCGs and methylation status, according to embodiments of the present disclosure. [Figure 10] 1 shows cleavage profiles of windows associated with different sequence contexts, including CCGs and methylation status, according to embodiments of the present disclosure. [Figure 11] A schematic diagram of cfDNA grouping at nearby CpG sites is shown. [Figure 12A] Figure 1 shows the effect of methylation patterns with multiple adjacent CpGs on the CGN / NCG motif ratio. [Figure 12B] Same as above. [Figure 13] 10 illustrates normalization of a cut profile using normalized cut density, according to an embodiment of the present disclosure. [Figure 14] 1 provides a cleavage profile using cleavage densities of various methylation combinations of two CpG sites according to an embodiment of the present disclosure. [Figure 15] 1 provides cleavage profiles using cleavage densities of various methylation combinations of three CpG sites according to embodiments of the present disclosure. [Figure 16] 1 shows the truncation density of a window containing CCGs and methylation status according to an embodiment of the present disclosure. [Figure 17] 1 shows a comparison of cleavage profiles between methylated and unmethylated CpG sites using cleavage density, according to an embodiment of the present disclosure. [Figure 18] 1 illustrates a workflow for using SVM to classify CpG sites as hypermethylated or hypomethylated, according to an embodiment of the present disclosure. [Figure 19] 1 shows the performance of single CpG site methylation status prediction using support vector machine (SVM) based on cleavage profiles, according to an embodiment of the present disclosure. [Figure 20] 1 shows the performance of single CpG site methylation status prediction using a support vector machine (SVM) based on cleavage profiles from both strands, according to an embodiment of the present disclosure. [Figure 21A]1 shows the performance of single-CpG site methylation status prediction using a support vector machine (SVM) based on cleavage profiles conditioned on different sequence contexts, according to an embodiment of the present disclosure. [Figure 21B] Same as above. [Figure 22] FIG. 1 shows a schematic diagram of methylation status prediction using a CNN model based on cleavage profile and sequence context, according to an embodiment of the present disclosure. [Figure 23A] 1 shows the performance of single CpG site methylation status prediction using CNN based on cleavage profile and sequence context, according to an embodiment of the present disclosure. [Figure 23B] A comparison of the sum of the average cleavage ratios at positions −1 and +1 between hypermethylated and hypomethylated CpGs for 39 ucfDNA samples is shown. [Figure 24A] Figure 1 shows the performance of methylation status prediction of a single CpG site using a CNN model based on proximal cleavage profile and sequence context, excluding position 0. [Figure 24B] Figure 1 shows the performance of methylation status prediction of single CpG sites using a CNN model based on proximal cleavage profile and sequence context, excluding positions -1, 0, and 1. [Figure 25A] Box plots illustrating the difference in mean cleavage ratios at various positions in plasma and urine (ucfDNA) between hypermethylated and hypomethylated CpGs. [Figure 25B] Same as above. [Figure 25C] Same as above. [Figure 25D] Same as above. [Figure 25E] Same as above. [Figure 25F] Same as above. [Figure 26A] A comparison of various techniques for determining quantitative values ​​of methylation levels (e.g., density) for CpGs with different methylation levels in eight plasma DNA samples is shown. [Figure 26B] Same as above. [Figure 26C] Same as above. [Figure 27]1 is a flowchart of a method for determining methylation at a site based on fragmentation, according to an embodiment of the present disclosure. [Figure 28] 1 shows the methylation index of CGN and NCG terminal motifs according to an embodiment of the present disclosure. [Figure 29] 1 shows the O / E ratios of CGN and NCG terminal motifs in hypermethylated and hypomethylated regions from eight healthy control samples, according to embodiments of the present disclosure. [Figure 30] 1 illustrates the use of trimer end motifs in TSS regions to determine binary classification or regression, according to embodiments of the present disclosure. [Figure 31A] 1 illustrates an exemplary training procedure, according to an embodiment of the present disclosure. [Figure 31B] Same as above. [Figure 32] 1 shows binary classification results for different k-mer motifs according to an embodiment of the present disclosure. [Figure 33A] 1 shows binary classification results for different k-mer motifs according to an embodiment of the present disclosure. [Figure 33B] Same as above. [Figure 34A] 10 shows the performance of local methylation status prediction using SVM based on local terminal motif frequency or O / E ratio of terminal motifs according to an embodiment of the present disclosure. [Figure 34B] 10 shows the performance of local methylation density estimation using SVR based on local end motif frequency, according to an embodiment of the present disclosure. [Figure 35] 1 illustrates normalization using the expected frequency of regions of each motif (local normalization) according to an embodiment of the present disclosure. [Figure 36] 10 shows binary classification results using an SVM model with context-normalized (locally normalized) terminal motifs, according to an embodiment of the present disclosure. [Figure 37] 1 illustrates per-molecule contextual normalization using expected motif frequencies in molecules, according to an embodiment of the present disclosure. [Figure 38] 1 shows results for a trimer motif with no normalization, local normalization, and molecule-by-molecule normalization, according to embodiments of the present disclosure. [Figure 39] 1 shows the difference in normalized frequency (local normalization) of different dimer end motifs according to an embodiment of the present disclosure. [Figure 40] 1 shows the difference in normalized frequency (local normalization) of different trimer end motifs according to an embodiment of the present disclosure. [Figure 41] 1 is a flowchart illustrating a method for predicting methylation in a region according to an embodiment of the present disclosure. [Figure 42] 1 illustrates an exemplary workflow of a machine learning model for predicting methylation status at the single molecule level, according to an embodiment of the present disclosure. [Figure 43A] 1 shows the performance of single molecule methylation status analysis using CNN based on terminal motifs, sequence context, and fragment size, according to embodiments of the present disclosure. [Figure 43B] 1 shows the performance of single molecule methylation status analysis using molecules with CGN and NCG terminal motifs based on a CNN model, according to an embodiment of the present disclosure. [Figure 44] 1 is a flowchart illustrating a method for predicting the methylation of a DNA fragment according to an embodiment of the present disclosure. [Figure 45] Boxplot showing the O / E ratios of CGN and NCG motifs between 5hmC-enriched and 5hmC-depleted regions. [Figure 46A] Figure 1 shows the performance of 5hmC prediction based on the 5' trimeric CG-containing terminal motif. [Figure 46B] Same as above. [Figure 47] 1 is a flow chart illustrating a method for measuring 5hmC methylation of regions in a subject's genome using cell-free DNA molecules. [Figure 48A] 1 shows motif frequencies of fetal-specific hypermethylated and hypomethylated regions according to embodiments of the present disclosure. [Figure 48B] Same as above. [Figure 49A]49A and 49B show the correlation between fetal DNA fraction and CGN / NCG ratio based on plasma DNA molecules derived from placenta-specific hypermethylated CpG sites (FIG. 49A) and placenta-specific hypomethylated CpG sites (FIG. 49B), according to embodiments of the present disclosure. [Figure 49B] Same as above. [Figure 50A] 5A and 5B show the correlation between fetal DNA fraction and the frequency of NCG and CGN motifs based on plasma DNA molecules derived from placenta-specific hypermethylated CpG sites (50A) and placenta-specific hypomethylated CpG sites (50B), according to embodiments of the present disclosure. [Figure 50B] Same as above. [Figure 51] 1 shows the correlation of fetal fraction inferred by a SNP-based approach (x-axis) and using CGN / NCG ratios (y-axis) from placenta-specific hypermethylated and hypomethylated CpGs, according to an embodiment of the present disclosure. [Figure 52] 1 shows a schematic diagram of CGN / NCG motif ratio analysis based on tissue-specific methylated CpGs in blood and liver. [Figure 53A] Figure 53 shows the CGN / NCG motif ratios from different tissue-specific hypermethylated (Figure 53A) and hypomethylated (Figure 53B) CpGs in plasma DNA of healthy controls, liver transplant cases, and pregnant cases. [Figure 53B] Same as above. [Figure 54] 1 shows that fetal-specific methylation is reflected in the cleavage profile of cfDNA fragmentation, according to embodiments of the present disclosure. [Figure 55] 1 illustrates marker selection, data collection, and training of an SVR model for determining fetal fraction in cell-free samples (e.g., plasma) according to an embodiment of the present disclosure. [Figure 56] 1 illustrates marker selection, data collection, and training of an SVR model for determining fetal fraction in cell-free samples (e.g., plasma) according to an embodiment of the present disclosure. [Figure 57] 1 shows the correlation between proximal cleavage profiles and fetal fraction inferred by a SNP-based approach, according to embodiments of the present disclosure. [Figure 58]1 is a flowchart illustrating a method for determining the fractional concentration of clinically relevant DNA according to an embodiment of the present disclosure. [Figure 59A] 59A and 59B show box plots illustrating the differences in whole genome methylation density (FIG. 59A), whole genome CGN / NCG ratio (FIG. 59B), and CGN / NCG ratio in Alu regions (FIG. 59C) between healthy controls and HCC cases, according to embodiments of the present disclosure. [Figure 59B] Same as above. [Figure 59C] Same as above. [Figure 60A] 1 shows box plots illustrating the difference in CGN and NCG motif frequencies from the whole genome and Alu regions between healthy controls and HCC cases, according to embodiments of the present disclosure. [Figure 60B] Same as above. [Figure 60C] Same as above. [Figure 60D] Same as above. [Figure 61] 1 shows the Alu region CGN / NCG ratios from healthy controls, HCC, HNSCC, CRC, LC, and NPC according to embodiments of the present disclosure. [Figure 62] 1 shows an exemplary technique using droplet PCR to determine NCG and CGN motifs in a region of interest, as well as exemplary structures of probes, according to embodiments of the present disclosure. [Figure 63] 1 shows the top 10 CG probes and the top 10 NCG probes for potential experimental designs according to an embodiment of the present disclosure. [Figure 64] 1 shows an in silico analysis of the CGN / NCG ratio by ddPCR assay using probes containing the Alu-associated sequences "CGCCTGT" and "TCGCTTGA" across healthy controls and subjects with HCC, HNSCC, CRC, LC, or NPC, according to an embodiment of the present disclosure. [Figure 65A] 1 shows ROC analysis in distinguishing patients with HCC, HNSCC, CRC, LC, or NPC from healthy controls using the CGN / NCG ratio from an in silico ddPCR assay, according to an embodiment of the present disclosure. [Figure 65B] Same as above. [Figure 65C] Same as above. [Figure 65D] Same as above. [Figure 65E] Same as above. [Figure 66] 1 shows an exemplary method using qPCR to determine NCG and CGN motifs in a region of interest. [Figure 67A] Examples of primer and probe designs are shown below. [Figure 67B] Same as above. [Figure 67C] Same as above. [Figure 68A] Test results from qPCR (A) and ddPCR (B) are shown. gDNA represents a mechanically fragmented gDNA sample, used as a control in the qPCR assay. [Figure 68B] Same as above. [Figure 69A] Figure 1 shows the CGN / NCG motif ratios in putative hypermethylated and hypomethylated CpGs across different size ranges of pooled single molecule real-time sequencing reads from healthy subjects. [Figure 69B] Same as above. [Figure 70A] Box plots showing the overall CGN / NCG motif ratios of short cfDNA fragments (≦200 bp, 70A) and long cfDNA fragments (≧1000 bp, 70B), and the difference in CGN / NCG motif ratios between short cfDNA fragments (≦200 bp) and long cfDNA fragments (≧1000 bp) across healthy subjects (Healthy), HBV carriers (HBV), and HCC cases (70C). [Figure 70B] Same as above. [Figure 70C] Same as above. [Figure 71A] The performance of HCC detection based on the CGN / NCG motif ratio together with size information is shown. [Figure 71B] Same as above. [Figure 72] 1 is a flowchart illustrating a method for determining a level of pathology according to an embodiment of the present disclosure. [Figure 73A]1 shows the performance of HCC detection based on CGN / NCG motif ratios from liver-specific hypermethylated and hypomethylated regions according to an embodiment of the present disclosure. [Figure 73B] Same as above. [Figure 74A] 1 shows the performance of HCC diagnosis using SVM based on trimer end motif frequency from liver-specific hypomethylated regions, according to an embodiment of the present disclosure. [Figure 74B] Same as above. [Figure 75A] 1 shows the performance of HCC detection based on CG-containing trimer terminal motifs from HCC-specific hypermethylated and hypomethylated CpGs. [Figure 75B] Same as above. [Figure 76] 1 shows ROC analysis in distinguishing patients with and without HCC using a combined data vector from HCC-specific hypermethylated and hypomethylated CpGs. [Figure 77] 1 illustrates exemplary marker selection and data collection according to an embodiment of the present disclosure. [Figure 78] 1 illustrates a workflow that includes generating cutting profiles for various objects to train a machine learning model, according to an embodiment of the present disclosure. [Figure 79] 79A and 79B show PCA analysis of pregnant and non-pregnant subjects by using cleavage profiles from placenta-specific hypermethylated sites (FIG. 79A), placenta-specific hypomethylated sites (FIG. 79B), or a combination of placenta-specific hypermethylated and hypomethylated sites (FIG. 80), according to embodiments of the present disclosure. [Figure 80] 79A and 79B show PCA analysis of pregnant and non-pregnant subjects by using cleavage profiles from placenta-specific hypermethylated sites (FIG. 79A), placenta-specific hypomethylated sites (FIG. 79B), or a combination of placenta-specific hypermethylated and hypomethylated sites (FIG. 80), according to embodiments of the present disclosure. [Figure 81] 1 shows the performance of HCC detection based on cleavage profiles. [Figure 82] FIG. 1 shows a schematic diagram for defining tissue-informative methylation markers. [Figure 83](Figure 83A) Unsupervised clustering analysis of patients with HCC and CRC based on tissue information and 5' trimer CG-containing terminal motifs associated with tissue-shared CpG sites. (Figures 83B-83C) Performance of HCC classifiers based on tissue information and 5' trimer CG-containing terminal motifs associated with tissue-shared CpG sites. [Figure 84] 1 is a flowchart of a method for determining the level of pathology for a particular tissue using tissue-specific differentially methylated regions according to an embodiment of the present disclosure. [Figure 85A] 1 shows the performance of HCC detection based on 5′ trimeric CG-containing terminal motifs from HCC-specific 5hmC enriched and 5hmC-depleted CpGs. [Figure 85B] Same as above. [Figure 86A] 1 shows the performance of EBV-associated lymphoma detection based on the CGN / NCG ratio from lymphoma-specific hypermethylated and hypomethylated CpG sites in plasma EBV DNA according to an embodiment of the present disclosure. [Figure 86B] Same as above. [Figure 87A] 1A-1B show ROC analysis of distinguishing patients with IM from patients with EBV-associated lymphoma using CGN and NCG motif frequencies based on lymphoma-specific hypomethylated (A) and hypermethylated (B) regions, according to embodiments of the present disclosure. [Figure 87B] Same as above. [Figure 88A] 1 shows the performance of NPC detection based on the CGN / NCG ratio from NPC-specific hypermethylated and hypomethylated CpG sites in EBV DNA according to an embodiment of the present disclosure. [Figure 88B] Same as above. [Figure 89A] Figure 1 shows ROC analysis in distinguishing patients with IM from those with NPC using CGN and NCG motif frequencies from NPC-specific hypomethylated (A) and hypermethylated (B) regions. [Figure 89B] Same as above. [Figure 90A]The distribution of cleavage ratios from NPC and non-NPC cases at positions −1 and 1 of each CpG site with a methylation density greater than 80% in both NPC and non-NPC cases is shown. [Figure 90B] Same as above. [Figure 91A] The distributions for both the -1 and 1 positions are shown. [Figure 91B] Cleavage preference when enzyme deficiency is present. [Fig. 92AB] 1 shows box plots illustrating the difference in EBV DNA cleavage ratios between NPC and non-NPC cases in the training and test sets at the −1 and +1 informative positions. [Figure 92CD] Same as above. [Figure 93A] Box plots illustrating the difference in EBV DNA cleavage ratios at the combined −1 and +1 informative positions between NPC and non-NPC cases in the training (A) and test (B) sets are shown. [Figure 93B] Same as above. [Figure 94A] Figure 1 shows the diagnostic power of distinguishing NPC cases from non-NPC cases by using EBV DNA cleavage ratios at combined −1 and +1 informative positions and EBV DNA percentages in the training (A) and test (B) sets. [Figure 94B] Same as above. [Figure 95] By using different measurement references in the training and test datasets, we demonstrate the diagnostic power of distinguishing NPC cases from non-NPC cases. [Figure 96] 1 shows a schematic diagram of the informed CpG selection based on adjusted CGN / NCG motif ratios in the training set and NPC diagnosis in the test set. [Figure 97A] Methylation indices of informative CpGs measured by bisulfite sequencing between non-NPC and NPC cases are shown. [Figure 97B] The adjusted CGN / NCG motif ratios of informative CpGs in the EBV genome between non-NPC and NPC individuals (training set) are shown. [Figure 98A]The adjusted CGN / NCG motif ratios of informative CpGs in the EBV genome between non-NPC and NPC individuals (test set) are shown. [Figure 98B] The positive predictive values ​​achieved by a PCR-based assay, an approach based on EBV DNA fraction and size ratio, and an approach combining EBV DNA fraction, size ratio, and cleavage motif are shown. [Figure 99] 1 is a flow chart illustrating a method for determining the level of pathology in a subject using a biological sample containing subject-derived and virus-derived cell-free DNA molecules. [Figure 100A] 1 shows that CGN and NCG end motif frequencies can be used to predict genomic imprinting, according to embodiments of the present disclosure. [Figure 100B] Same as above. [Figure 101] 1 shows the performance of distinguishing between methylated and unmethylated regions as a function of sequencing depth according to embodiments of the present disclosure. [Figure 102] We show that the relative percentage of sequenced fragments that align to a particular genomic region of interest can be enriched by selectively analyzing a subset of DNA molecules that terminate in a particular terminal motif. [Figure 103] 1 is an exemplary workflow for sequencing library preparation by selectively amplifying DNA containing NCG or CGN terminal motifs from a region of interest, according to an embodiment of the present disclosure. [Figure 104] Figure 1 shows the CGN / NCG motif ratios of putative hypermethylated and hypomethylated CpGs in sequencing libraries of plasma DNA prepared with 4 PCR cycles, 8 PCR cycles, and without PCR, respectively. [Figure 105] Figure 1 shows a comparison of cleavage profiles between hypermethylated and hypomethylated CpG sites in urinary cfDNA samples of healthy controls. [Figure 106A]Boxplots are shown illustrating the differences in methylation density (Figure 106A), CGN / NCG motif ratio (Figure 106B), CGN / CG motif ratio (Figure 106C), and CGN / (NCG+CG) motif ratio (Figure 106D) across the whole genome, Alu regions, and CpG islands in urinary cfDNA from healthy controls. Each dot represents a urinary cfDNA sample. [Figure 106B] Same as above. [Figure 106C] Same as above. [Figure 106D] Same as above. [Figure 107A] Not specified. [Figure 107B] Same as above. [Figure 108] 108A and 108B show the correlation between donor DNA fraction and CGN / NCG motif ratio based on urinary cfDNA molecules derived from kidney-specific hypermethylated CpG sites (FIG. 108A) and kidney-specific hypomethylated CpG sites (FIG. 108B). [Figure 109] Figure 1 shows the correlation between donor DNA fraction and normalized CGN / NCG motif ratio in urinary cfDNA from kidney transplant cases. [Figure 110A] The correlation between donor DNA fraction and CGN / NCG motif ratio based on urinary cfDNA molecules derived from blood cell-specific hypermethylated CpG sites (Figure 110A) and blood cell-specific hypomethylated CpG sites (Figure 110B) is shown. [Figure 110B] Same as above. [Figure 111] Figure 1 shows the correlation between donor DNA fraction and normalized CGN / NCG motif ratio in urinary cfDNA from bone marrow transplant patients. [Figure 112A] Box plot showing CGN / NCG motif ratios from bladder cancer-specific hypermethylated and hypomethylated CpGs across healthy controls (control), low-grade non-muscle-invasive bladder cancer (NMIBC LG), high-grade non-muscle-invasive bladder cancer (NMIBC HG), and patients with muscle-invasive bladder cancer (MIBC). [Figure 112B] Same as above. [Figure 113]Box plot showing normalized CGN / NCG motif ratios from bladder cancer-specific hypermethylated and hypomethylated CpGs across healthy controls (control), low-grade non-muscle-invasive bladder cancer (NMIBC LG), high-grade non-muscle-invasive bladder cancer (NMIBC HG), and patients with muscle-invasive bladder cancer (MIBC). [Figure 114] FIG. 1 is a schematic diagram of SVM model training based on cleavage profiles from bladder cancer-specific hypermethylated and hypomethylated CpG sites to perform bladder cancer detection. [Figure 115A] 1 shows the performance of bladder cancer detection based on truncation profiles. [Figure 115B] Same as above. [Figure 116A] Figure 1 shows the performance of bladder cancer detection based on the 5' trimer CG-containing terminal motif. [Figure 116B] Same as above. [Figure 117] FIG. 1 is a schematic diagram of bladder cancer cases and cancer-free controls. [Figure 118A] Diagnosis of bladder cancer by the number of bins with significantly low normalized CGN / NCG motif ratios (i.e., motif abnormalities) is shown. [Figure 118B] Same as above. [Figure 119] Figure 1 shows residual disease diagnosis by using the amount of bins showing significantly lower normalized CGN / NCG motif ratios (motif aberrations) in urinary cfDNA from bladder cancer patients. [Figure 120] FIG. 1 shows a schematic diagram for defining tissue-informative methylation markers using bladder cancer as an example. [Figure 121] A schematic diagram of multiple cancer clustering with CG-containing terminal motifs is shown. [Figure 122] 1 shows an unsupervised clustering analysis of patients with bladder, kidney, and prostate cancer based on tissue information and 5′ trimeric CG-containing terminal motifs associated with tissue-shared CpG sites. [Figure 123] FIG. 1 shows a schematic diagram of classification of multiple cancer types using multiple binary classification. [Figure 124A]1 shows the performance of a bladder cancer classifier based on tissue information and 5′ trimeric CG terminal motifs associated with tissue-shared CpG sites. [Figure 124B] Same as above. [Figure 125A] 1 shows the performance of kidney cancer and prostate cancer classifiers based on 5′ trimer CG-terminal motifs associated with tissue-informative CpG sites as well as tissue-shared CpG sites. [Figure 125B] Same as above. [Figure 125C] Same as above. [Figure 125D] Same as above. [Figure 126A] 1 shows a schematic diagram of multi-cancer classification using a CNN model based on CG-containing motifs. [Figure 126B] Same as above. [Figure 127] 1 illustrates a measurement system according to an embodiment of the present invention. [Figure 128] FIG. 1 shows a block diagram of an exemplary computer system usable with systems and methods according to embodiments of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0010] term A "tissue" corresponds to a group of cells grouped together as a functional unit. More than one type of cell can be found in a single tissue. Different types of tissue can consist of different types of cells (e.g., liver cells, alveolar cells, or blood cells), but can also correspond to tissues from different organisms (maternal vs. fetal) or healthy vs. tumor cells. A "reference tissue" can correspond to the tissue used to determine tissue-specific methylation levels or tissue-specific fragmentation patterns. Multiple samples of the same tissue type from different individuals can be used to determine tissue-specific methylation levels for that tissue type.

[0011] A "biological sample" refers to any sample obtained from a subject (e.g., a human (or other animal), such as a pregnant woman, an individual with or suspected of having cancer or other disorder, an organ transplant recipient, or a subject suspected of having a disease process involving an organ (e.g., the heart in a myocardial infarction, or the brain in a stroke, or the hematopoietic system in anemia)) and containing one or more nucleic acid molecules of interest. A biological sample can be a bodily fluid, such as blood, plasma, serum, urine, vaginal fluid, fluid from a hydrocele (e.g., of the testes), vaginal washings, pleural effusion, peritoneal fluid, cerebrospinal fluid, saliva, sweat, tears, sputum, bronchoalveolar lavage fluid, nipple discharge fluid, aspirated fluid from different parts of the body (e.g., thyroid, breast), intraocular fluid (e.g., aqueous humor), etc. Fecal samples can 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 via a centrifugation protocol) can be cell-free, e.g., greater than 50%, 60%, 70%, 80%, 90%, 95%, or 99% of the DNA can be cell-free. The centrifugation protocol can include, for example, 3,000 g x 10 minutes, obtaining a fluid portion, and recentrifuging, for example, at 30,000 g for an additional 10 minutes to remove residual cells. As part of the analysis of the biological sample, a statistically significant number of cell-free DNA molecules can be analyzed for the biological sample (e.g., to provide an accurate measurement). 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 1,000,000, or 500,000,000 or more cell-free DNA molecules can be analyzed. At least the same number of sequence reads can be analyzed.

[0012] "Clinically relevant DNA" can refer to DNA of a particular tissue source that is measured, for example, to determine the fractional concentration of such DNA or to phenotype a sample (e.g., plasma). Examples of clinically relevant DNA are fetal DNA in maternal plasma, or tumor DNA in a patient's plasma, or other samples with cell-free DNA. Another example includes measuring the amount of graft-associated DNA in the plasma, serum, or urine of a transplant patient. Further examples include measuring the fractional concentrations of hematopoietic and non-hematopoietic DNA in a subject's plasma, or the fractional concentration of liver DNA fragments (or other tissues) in a sample, or the fractional concentration of brain DNA fragments in cerebrospinal fluid.

[0013] A "sequence read" refers to a chain of nucleotides sequenced from any portion or all of a nucleic acid molecule. For example, a sequence read can be a short chain of nucleotides (e.g., 20-150 nucleotides) sequenced from a nucleic acid fragment, a short chain of nucleotides at one or both ends of a nucleic acid fragment, or the sequencing of an entire nucleic acid fragment present in a biological sample. Sequence reads can be obtained in a variety of ways, for example, using sequencing technologies or probes, such as hybridization arrays or capture probes such as those used in microarrays, or amplification technologies such as polymerase chain reaction (PCR) or linear amplification using a single primer or isothermal amplification. Exemplary sequencing technologies include massively parallel sequencing, targeted sequencing, Sanger sequencing, sequencing by ligation, ion semiconductor sequencing, and single-molecule sequencing (e.g., using nanopores or single-molecule real-time sequencing (e.g., from Pacific Biosciences)). Exemplary PCR technologies include real-time PCR and digital PCR (e.g., droplet digital PCR). As part of the analysis of a biological sample, a statistically significant number of sequence reads can be analyzed, for example, at least 1,000 sequence reads can be analyzed. In other examples, at least 10,000, or 50,000, or 100,000, or 1,000,000, or 500,000,000 or more sequence reads can be analyzed.

[0014] A sequence read can include a "terminal sequence" associated with the end of a fragment. The terminal sequence can correspond to the outermost N bases of the fragment, e.g., 1 to 30 bases at the end of the fragment. When a sequence read corresponds to an entire fragment, the sequence read can include two terminal sequences. When paired-end sequencing provides two sequence reads corresponding to the ends of a fragment, each sequence read can include one terminal sequence.

[0015] A "sequence motif" may refer to a short repeating pattern of bases in a DNA fragment (e.g., a cell-free DNA fragment). A sequence motif occurs at the end of a fragment and therefore may be part of or include the terminal sequence. A "terminal motif" (also referred to as a "terminal sequence motif") may refer to a sequence motif of a terminal 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, thereby still corresponding to the terminal sequence. A nuclease may have a specific cleavage preference for a particular terminal motif and a second, most preferred cleavage preference for a second terminal motif. The number of nucleotides (nt) at the end of a fragment used for analysis may be, for example, but is not limited to, 1 nt, 2 nt, 3 nt, 4 nt, 5 nt, 6 nt, 7 nt, 8 nt, 9 nt, and 10 nt or more. In some embodiments, a fragment terminal motif may be defined by one or more nucleotides spanning positions near the end of the fragment. A fragment end motif may be defined by one or more nucleotides in the reference genome that surround the genomic locus to which the ends of the fragment are aligned.

[0016] A "site" (also referred to as 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, a TSS site, a DNase-sensitive site, or a larger group of correlated base positions. A "locus" may correspond to a region containing multiple sites. A locus may contain only one site, making a gene equivalent to the site in that context. Various numbers of regions, sites, or loci can be analyzed, e.g., 50, 100, 200, 500, 1,000, 5,000, 10,000, 50,000, 100,000, 500,000, 1 million, or more. Various techniques, e.g., alignment of sequence reads to a reference genome or use of location-specific probes, can determine that a DNA molecule is located at one or more genomic locations in a reference genome. Location can be relative to part or all of the reference genome, e.g., if only a portion of the genome is analyzed. By way of example, the amount of the genome analyzed can be greater than 0.01%, 0.1%, 1%, 5%, 10%, or 50%. A "cleavage site" can refer to the location where DNA is cleaved by a nuclease, thereby resulting in a DNA fragment.

[0017] A "cleavage profile" may refer to the amount of fragments ending at two or more positions occurring in a window around a site (e.g., a CpG site). The amount of fragments may correspond to different categories according to the end motif (e.g., CGN and NCG for positions 0 and -1, respectively). The amount can be normalized, for example, using the sequencing depth at each position, the depth in the region, or the number of fragments ending in the region. Such a normalized amount at a single position may be referred to as the cleavage ratio, cleavage amount, or cleavage density. In one example, a cleavage profile may be defined as the pattern of the ratio between fragment ends and sequencing depth across genomic coordinates within a window associated with a CpG site, which may be used to infer the methylation pattern of that CpG site. Various types of normalization may be used, as described herein. The window may include, but is not limited to, X nucleotides upstream (i.e., X-nt) and Y nucleotides downstream (i.e., Y-nt) of the CpG site. The values ​​of X and Y can be 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 30, 40, 50, 100, 1000, 5000, etc. The window can cover a nucleosome size range upstream and downstream of the CpG site, for example, from -160 nt to 160 nt.

[0018] The term "allele" refers to alternative DNA sequences at the same physical genomic locus that may or may not result in different phenotypic traits. In any particular diploid organism, which has two copies of each chromosome (except for sex chromosomes in male human subjects), the genotype of each gene contains a pair of alleles present at that locus that are the same in homozygotes and different in heterozygotes. A population or species of organism typically contains multiple alleles at each locus among various individuals. Genomic loci at which more than one allele is found in a population are referred to as 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 variation include, but are not limited to, single nucleotide polymorphisms, simple tandem repeat polymorphisms, insertion / deletion polymorphisms, mutations (which can cause disease), and copy number variation. 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 as few as one pair of loci, or a chromosomal region, or an entire chromosome or chromosomal arm.

[0019] The term "sequence context" may refer to the base composition (A, C, G, or T) and base order in a strand of DNA. Such a strand of DNA may surround a CpG site that is undergoing methylation analysis or is the target of methylation analysis. For example, the sequence context may refer to bases upstream and / or downstream of the CpG site that is undergoing methylation analysis. For example, the sequence context may include a K-mer matrix that specifies the number of instances of each K-mer in the test cell-free DNA molecule, where K is an integer. When K is equal to 2, the K-mer matrix is ​​a dinucleotide matrix that specifies the number of instances of a dinucleotide pair in the test cell-free DNA molecule. As another example, the sequence context may include sequence information from each nucleotide (i.e., the entire sequence) of the test cell-free DNA molecule. One-hot encoding may be used on the sequence, thereby generating a matrix of 0s and 1s with dimensions of 4×N, where N is the number of bases in the sequence.

[0020] Various types of methylation exist. Methylations such as 5mC (5-methylcytosine), 4mC (N4-methylcytosine), 5hmC (5-hydroxymethylcytosine), 5fC (5-formylcytosine), 5caC (5-carboxylcytosine), 1mA (N1-methyladenine), 3mA (N3-methyladenine), 7mA (N7-methyladenine), 3mC (N3-methylcytosine), 2mG (N2-methylguanine), 6mG (O6-methylguanine), 7mG (N7-methylguanine), 3mT (N3-methylthymine), and 4mT (O4-methylthymine) have been found on cytosine, adenine, thymine, and guanine. In vertebrate genomes, 5mC is the most common type of base methylation, followed by that on guanine (i.e., in CpG contexts).

[0021] The "methylation index" or "methylation state" of each genomic site (e.g., a CpG site) may refer to the proportion of DNA fragments (e.g., determined from sequence reads or probes) that exhibit methylation at that site across the total number of reads covering that site. A methylation state may refer to whether a particular site is methylated at a particular site of a DNA fragment. A "read" may correspond to information obtained from a DNA fragment (e.g., the methylation state at a site). A read may be obtained using a reagent (e.g., a primer or probe) that preferentially hybridizes to DNA fragments of a particular methylation state. Typically, such a reagent is applied after treatment with a process that differentially modifies or differentially recognizes DNA molecules depending on 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.

[0022] The "methylation density" of a region or set of sites may refer to the number of reads at a site or set of sites within a region (also referred to as a bin) that exhibit methylation, divided by the total number of reads covering the site or set of sites in the region. A region may contain one or more sites of interest, including at least 1, 2, 3, 4, 5, 10, 20, 50, 100, 200, 500, and 1,000 sites. A site may have the specific characteristic of being a CpG site, for example. Thus, the "CpG methylation density" of a region may refer to the number of reads that exhibit CpG methylation, divided by the total number of reads covering the CpG sites in the region (e.g., a particular CpG site, a CpG site within a CpG island, or a larger region). For example, the methylation density of each 100 kb bin in the human genome can be determined from the total number of cytosines (corresponding to methylated cytosines) at CpG sites that are not converted after bisulfite treatment as a percentage of all CpG sites covered by sequence reads mapped to the 100 kb region. This analysis can also be performed for other bin sizes, such as 500 bp, 5 kb, 10 kb, 50 kb, or 1 Mb. The region can be the entire genome, or a chromosome or part of a chromosome (e.g., a chromosome arm). The methylation index of a CpG site is the same as the methylation density of that region if it contained only that CpG site. The "percentage of methylated cytosines" can refer to the number of cytosine sites "C" that are shown to be methylated (e.g., not converted after bisulfite conversion) across the total number of analyzed cytosine residues in the region, i.e., including cytosines outside of CpG contexts. Methylation index, methylation density, and percentage of methylated cytosines are examples of "methylation levels."Aside 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, single molecule sequencing using methylation-state sensitive enzymes (e.g., methylation-sensitive restriction enzymes), methylation-binding proteins, methylation-state sensitive platforms (e.g., nanopore sequencing (Schreiber et al. Proc Natl Acad Sci USA 2013;110:18910-18915), and Pacific Biosciences single molecule real-time analysis (Tse et al. Proc Natl Acad Sci USA 2021;118:e2019768118).

[0023] The term "hypomethylated" may refer to a site or set of sites (e.g., regions) that are below a specified threshold for methylation level, e.g., have a methylation level of 50%, 45%, 40%, 35%, 30%, 25%, or 20% or less. A site in a genome may be considered unmethylated if its methylation level is below a threshold. The term "hypermethylated" may refer to a site or set of sites (e.g., regions) that are above a specified methylation level, e.g., have a methylation level of 95%, 90%, 80%, 75%, 70%, 65%, or 60% or more. A site in a genome may be considered methylated if its methylation level is above a threshold.

[0024] A differentially methylated region (DMR) is a genomic region that has different DNA methylation states across two or more biological samples. Different DNA methylation states can be defined by certain differences in methylation index or density, such as, but not limited to, 1%, 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 95%, 99%, etc.

[0025] The term "5hmC enriched" can refer to a site or set of sites (e.g., a region) that has a higher relative proportion of 5hmC methylated molecules relative to other types of methylation. The term "5hmC depleted" can refer to a site or set of sites (e.g., a region) that has a lower relative proportion of 5hmC methylated molecules relative to other types of methylation.

[0026] "Relative frequency" (also referred to as "frequency") can refer to a proportion (e.g., a percentage, fraction, or concentration). Specifically, the relative frequency of a particular terminal motif (e.g., CCGA or a single base only) or group of terminal motifs (e.g., CGN or NCG) can provide the proportion of cell-free DNA fragments in a sample that are associated with a set of one or more terminal motifs.

[0027] An "aggregate value" can refer to, for example, a collective property of the relative frequencies of a set of terminal motifs. Examples include the mean, median, sum of the relative frequencies, the variation between the relative frequencies (e.g., entropy, standard deviation (SD), coefficient of variation (CV), interquartile range (IQR), or a particular percentile cutoff (e.g., the 95th or 99th percentile) between different relative frequencies), or the difference (e.g., distance) of the relative frequencies from a reference pattern, as may be implemented in clustering. As another example, the aggregate value can include an array / vector of relative frequencies that can be compared to a reference vector (e.g., representing multidimensional data points).

[0028] The term "sequencing depth" refers to the number of times a locus is covered by sequence reads aligned to the locus. A locus can be as small as a nucleotide, as large as a chromosome arm, or as large as the entire genome. Sequencing depth can be expressed as 50x, 100x, etc., where "x" refers to the number of times a locus is covered by sequence reads. Sequencing depth can also be applied to multiple loci or the entire genome, in which case x can refer to the average number of times a locus, or a haploid genome, or the entire genome, is sequenced, respectively. Ultra-deep sequencing can refer to a sequencing depth of at least 100x.

[0029] A "calibration sample" can correspond to a biological sample with a known characteristic (measurement). Exemplary characteristics include methylation levels or the fractional concentration of clinically relevant DNA. The fractional concentration of clinically relevant DNA (e.g., tissue-specific DNA fraction) can be known or determined through calibration methods, for example, using tissue-specific alleles, such as transplants, where alleles present in the donor's genome but absent from the recipient's genome can be used as markers for the transplanted organ. By way of example, the methylation level can be of a region, genome, or site in or on a genome. The methylation level can be determined using a methylation-recognition assay, such as methylation-recognition sequencing, or PCR. Exemplary methylation-recognition sequencing can include, for example, bisulfite sequencing or single-molecule techniques using nanopores. A calibration sample can have a distinct measurement (e.g., the amount of fragments with a particular terminal motif or cleavage profile) that can be determined to which the desired measurement can be correlated.

[0030] "Calibration data points" include "calibration values" (e.g., the amount of fragments having a particular terminal motif or cleavage profile) and other measured or known values ​​desired to be determined for test samples. Calibration values ​​can be determined from various types of data measured from the DNA molecules of a sample. Calibration values ​​correspond to parameters that correlate with a desired characteristic, e.g., a classification of a genetic disorder or methylation level. For example, calibration values ​​can be determined from measurements determined for calibration samples in which the desired characteristic is known. Calibration data points can be defined in various ways, for example, as discrete points or as a calibration function (also called a calibration curve or calibration surface). A calibration function can be derived from additional mathematical transformations of the calibration data points.

[0031] A "separation value" corresponds to the difference or ratio between two values, for example, two fractional contributions or two methylation levels. A separation value can be a simple difference or ratio. As an example, the direct ratio of x / y is a separation value, as is x / (x+y). A separation value can include other factors, for example, multiplicative factors. As another example, the difference or ratio of a function of the values ​​can be used, for example, the difference or ratio of the natural logarithms (ln) of two values. A separation value can include a difference and a ratio. A separation value can be compared to a threshold to determine whether the separation between two values ​​is statistically significant.

[0032] 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) between a first amount of a first nucleic acid sequence and a second amount of a second nucleic acid sequence is a parameter. A parameter can be used to determine any classification described herein, for example, for fetal, cancer, or transplant analysis. "Separation values" and "aggregate values" (e.g., of relative frequencies) are two examples of parameters (also called measurement references) that provide a measure of sample variation between different classifications (states) and can therefore be used to determine different classifications. An aggregate value can be a separation value when the difference is taken between a set of relative frequencies of a sample and a reference set of relative frequencies, as can be done, for example, in clustering.

[0033] The term "classification" as used herein refers to any number or other character associated with a particular characteristic of a sample. For example, a "+" sign (or the word "positive") may indicate that the sample is classified as having a deletion or amplification. Classification may be binary (e.g., positive or negative) or may have more levels of classification (e.g., a scale of 1 to 10 or 0 to 1).

[0034] The terms "cutoff" and "threshold" refer to predetermined numbers used in operations. 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 is applied. Either of these terms can be used in either 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 reference values ​​can be determined in various ways, as will be understood by those skilled in the art. For example, a measurement reference can be determined for two different cohorts of subjects with different known classifications, and a reference value can be selected to represent a value between one classification (e.g., an average value) or two clusters of measurement references (e.g., selected to obtain a desired sensitivity and specificity). As another example, a reference value can be determined based on statistical simulation of samples. Particular values ​​of cutoffs, thresholds, references, etc. can be determined based on the desired accuracy (e.g., sensitivity and specificity).

[0035] The term "level of cancer" can refer to whether cancer is present (i.e., present or absent), the stage of cancer, tumor size, whether there is metastasis, total tumor burden in the body, the response of cancer to treatment, and / or other measures of the severity of cancer (e.g., cancer recurrence). Cancer level can be a number or other indicator, such as a symbol, alphabetic character, and color. The level can be zero. Cancer level can also include premalignant or precancerous conditions (states). Cancer level can be used in a variety of ways. For example, screening can determine whether cancer is present in a person not previously known to have cancer. Evaluation can investigate a person diagnosed with cancer to monitor the progression of cancer over time, study the effectiveness of treatment, or determine a prognosis. In one embodiment, prognosis can be expressed as the likelihood that a patient will die from cancer, or the likelihood that the cancer will progress after a specific duration or time, or the likelihood or extent that the cancer will metastasize. Detection can mean "screening" or can mean determining whether a person with suggestive characteristics of cancer (e.g., symptoms or other positive test) has cancer.

[0036] "Level of pathology" may refer to the amount, degree, or severity of pathology associated with an organism, and the level may be as described above for cancer. Another example of pathology is rejection of a transplanted organ. Other exemplary pathologies may include autoimmune attacks (e.g., lupus nephritis, which damages the kidney, or multiple sclerosis, which damages the central nervous system), 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 subject's healthy state may be considered a pathology-free category.

[0037] A "machine learning model" may refer to a software module configured to run on one or more processors and provide a classification or numerical value of one or more sample characteristics. An exemplary type of model is supervised learning that can be used with embodiments of the present disclosure. Exemplary supervised learning models may include different 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 handling, 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 methodologies, symbolic machine learning algorithms, less-than-symbolic machine learning algorithms, minimal complexity machines (MCMs), random forests, ensembles of classifiers, ordinal classification, data preprocessing, handling imbalanced datasets, statistical relationship learning, or Proaftn, and multi-reference classification algorithms. Models may include linear regression, logistic regression, deep recurrent neural networks (e.g., long short-term memory, LSTM), hidden Markov models (HMMs), linear discriminant analysis (LDA), k-means clustering, density-based spatial clustering for applications with noise (DBSCAN), random forest algorithms, support vector machines (SVMs), or any of the models described herein. Supervised learning models can be trained in a variety of ways, using various cost / loss functions that define the error from known labels (e.g., least squares and absolute difference from known classifications), and various optimization techniques, e.g., backpropagation, steepest descent, conjugate gradient methods, and Newton and quasi-Newton techniques.

[0038] The term "about" or "approximately" can mean within an acceptable error range for a particular value, as determined by one of ordinary skill in the art, which depends in part on how the value is measured or determined, i.e., the limitations of the measurement system. For example, "about" can mean within 1 or more than 1 standard deviation, according to convention within the art. Alternatively, "about" can mean a range of up to 20%, up to 10%, up to 5%, or up to 1% of a given value. Alternatively, particularly with respect to biological systems or processes, the term "about" or "approximately" can mean within an order of magnitude, within 5-fold, or more preferably within 2-fold of a value. When a particular value is described in this application and claims, unless otherwise stated, the term "about" meaning within an acceptable error range of the particular value should be assumed. The term "about" can have the meaning commonly understood by one of ordinary skill in the art. The term "about" can refer to ±10%. The term "about" can refer to ±5%.

[0039] Where a range of values ​​is provided, unless the context clearly dictates otherwise, it is understood that each intervening value, to the tenth of the unit of the lower limit, between the upper and lower limit of that range is also specifically disclosed. 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 independently be included or excluded within the range, and each range where either, neither, or both limits are included within the smaller range is also encompassed within the disclosure, subject to any specifically excluded limits in the stated range. When a stated range includes one or both of the limits, ranges excluding either or both of those included limits are also included within the disclosure.

[0040] Standard abbreviations may be used, such as bp: base pairs, kb: kilobase, pi: picoliters, s or sec: seconds, min: minutes, h or hr: hours, aa: amino acids, nt: nucleotides, etc.

[0041] 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 potential and exemplary methods and materials may be described here. (Mode for Carrying Out the Invention)

[0042] This paper develops an approach to infer the methylation status of cfDNA molecules based on utilizing cfDNA fragmentation patterns.The fragmentation patterns herein can include, but are not limited to, fragment ends, cleavage profiles, fragment end motifs, genomic coordinates of fragment ends, and fragment sizes.

[0043] We have found cleavage profiles associated with CpG methylation status. The cleavage profiles can be used to predict methylation at single-base resolution for a region or on a fragment-by-fragment basis. Additionally, we have found that cfDNA terminal motifs (e.g., trimers or larger) are useful for predicting local methylation. Thus, cleavage profiles measuring fragments terminating in a window around a site can be used to infer the methylation of a region, a site in a genome, or one or more sites in a fragment. As another example, fragment terminal motifs can be used to infer the methylation level (e.g., local methylation level) of a genomic region of interest. A fragment terminal motif can be defined by one or more nucleotides at the end of a cell-free DNA fragment. As a further example, cfDNA fragmentomics features can be used to infer the methylation status (i.e., the methylation pattern across CpG sites present in a DNA molecule) at single-molecule resolution. For example, terminal motifs from one DNA molecule can be used to predict the methylation status of an entire fragment.

[0044] A further aspect of the present disclosure provides for determining the fractional concentration of clinically relevant DNA using cleavage profiles. For example, but not limited to, methylation-related cfDNA fragmentation features can be used to perform tissue-of-origin analysis of plasma DNA. As described herein, various fragmentation features, such as cleavage profiles and terminal motifs, can be used. Certain sites can be selected using fragmentation measurements to determine such fractional concentrations. For example, sites / regions that are hypermethylated and / or hypomethylated for a particular tissue type can be used. As another example, sites that are 5hmC enriched or depleted for a particular tissue type can be used. In addition, the technique can use fragmentation measurements (e.g., terminal motifs and cleavage profiles) to identify 5hmC-enriched and 5hmC-depleted regions for a particular tissue.

[0045] Yet a further aspect of the present disclosure provides pathology detection. As an example, CG-containing motifs from genome-wide or tissue-specific (e.g., liver-specific) hypomethylated or hypermethylated sites / regions can provide, for example, cancer (e.g., HCC) diagnosis / classification or other diseases that may be associated with a particular tissue. Subject and / or pathogen (e.g., virus) DNA can be used. For either type of DNA, sites with similar methylation levels across different subjects (with or without pathology) but where the subjects have different cleavage profiles can be used. As yet another example, sites / regions that are 5hmC enriched or depleted for a specific tissue type can be used, for example, to identify pathology for a specific tissue type. Machine learning models can use such tissue-specific methylation patterns (e.g., hypermethylated and hypomethylated, and 5hmC enriched and depleted) to distinguish between different types of cancer. Cell-free DNA fragments used for fragmentation measurements can be filtered, for example, based on size. Thus, pathology can be detected based on the amount of terminal motifs in cell-free DNA fragments with a specific size.

[0046] Fragmentation information (fragmentomics features) can be measured in a variety of ways. For example, massively parallel sequencing or whole-genome sequencing can be used. As another example, a targeted sequencing approach can be used to determine cfDNA fragmentomics features from regions of interest used for methylation pattern prediction. As an additional example, cfDNA fragmentomics features can be detected using non-sequencing-based approaches such as quantitative polymerase chain reaction (qPCR), real-time PCR, digital PCR (dPCR), droplet digital PCR (ddPCR), and mass spectrometry.

[0047] Exemplary use of 5' end motifs as an example only. 3' end motifs may be used in addition to or instead of any of the embodiments described herein.

[0048] I. Relationship of Fragmentation to Methylation The present disclosure shows the relationship between the location where DNA is cut (e.g., as part of forming cell-free DNA) and the methylation index (state) of the site in the genome. The location where DNA is cut can be defined as a cleavage profile. In another example, the location where DNA is cut can be represented by a terminal motif.

[0049] A. Definition of cutting position 1 illustrates cleavage positions relative to CpG sites (also referred to as CG sites) according to an embodiment of the present disclosure. The horizontal line indicates a portion of a reference sequence, e.g., a reference genome, that contains two CpG sites. After sequencing, the resulting reads can be mapped to this region. The distance between the fragment ends and the positions relative to the CpG sites can be calculated. For example, fragment 110 ends exactly at a CG position and therefore has a distance of 0 (also referred to as position 0).

[0050] Fragment 120 ends one position to the left of CpG site 102. Because the fragment was cut one base before the CpG site, the distance is considered to be 1. In this way, the distance between the cut end and the CpG site can be calculated and fragments with the same distance can be grouped together. As shown below, this distance is related to the methylation level.

[0051] Figure 2A shows how DNA fragments can be grouped according to the distance from the 5' end to the CpG site. If the first two nucleotides at the 5' end of the fragment are CG, the distance is 0. If there is one nucleotide at the 5' end directly preceding the CG, the distance is 1, which corresponds to a fragment with an NCG motif, where N is any one of four bases. Sequenced CpG sites can be grouped into different categories according to their distance from the 5' end. The maximum distance shown in Figure 2A is 8, but additional distances can be used.

[0052] B. Cleavage location correlates with cleavage CpG methylation To explore whether cfDNA fragmentation patterns can be used for methylation prediction, we investigated the relationship between the methylation index and the CpG position relative to the 5' end of the fragment (i.e., the nucleotide distance between CpG positions to the 5' end). Bisulfite sequencing reads from eight plasma DNA samples from healthy subjects were analyzed (median number of paired end sequencing reads: 381 million (interquartile range: 353-404 million)). Sequences were then aligned to the reference genome. The methylation density of each CpG site was determined using the methylation value of the sequence reads at that CpG site.

[0053] The methylation index of each distance can be determined in various ways. For example, the number of methylated CpG sites covered by the DNA fragments in a group divided by the total number of sequenced CpG sites on the DNA fragments of each group can be used. If there are 100 CpG sites on a fragment, and 90 of them are methylated, the index will be 90. As another example, the methylation index can be determined as the number of fragments in a group that occur at hypermethylated sites (for example, the methylation density at sites that exceed a specified percentage, such as 60%, 70%, 80%).

[0054] Figure 2B provides the methylation levels of CpG sites each having a given downstream distance from the cleavage edge of a plasma DNA molecule (e.g., the methylation levels of CpG sites with a distance of 5 nt to the 5' end). Figure 2B shows that the methylation index varies depending on the distance between the 5' end and the CpG site. There is a correlation between the methylation index of CpG sites and their distance to the 5' end. As can be seen, when the distance is 0, the methylation level is highest among CpGs. When the distance is 1, the methylation level is lowest. As the distance further increases, the methylation level becomes fairly stable. When a CpG site is methylated, cleavage occurs more frequently at the methylated C, while when the CpG site is unmethylated, more random cleavage occurs.

[0055] Therefore, CpG sites with a distance of 0 are more likely to be methylated than those with a distance of 1 [mean methylation index: 88.7% (range: 87.3-90.3) vs. 69.1% (range: 58.2-76.9%)]. Because the methylation index varies depending on the distance between the 5' end and the CpG site, the fragmentation pattern can be used to predict the methylation level of those CpG sites.

[0056] The cleavage position is also related to the motif type. For example, if the fragment is cleaved exactly at the CG position, the terminal motif is CGN**** (* represents a base, and N is one of the four bases). If the fragment is cleaved one base before this CG, the terminal motif is NCG*** (* represents a base), etc. Thus, there is a relationship between the motif type and the methylation level of the CpG site around which the fragment is cleaved.

[0057] To verify our observations, we further investigated the methylation levels in different types of tetramer motifs, i.e., 256 possible sequences at the four positions at the ends of the fragments.

[0058] Figure 3 shows the percentage of methylated CpG sites covered by multiple tetramer end motifs and fragments with specific tetramer end motifs according to an embodiment of the present disclosure. In Figure 3, each dot represents a specific tetramer motif. Individual motifs are grouped based on the distance of the cleavage position from the CG. As can be seen, the CGNN motif still revealed the highest methylation level, while the NCGN motif revealed the lowest methylation level, with NNCG in between.

[0059] As illustrated in 301, if a CpG site is methylated, it may be highly preferable to cleave exactly at this CG. In 302, if it is unmethylated, it may be possible to cleave one base before this CpG site. This is why we observed that many fragments starting with CG are methylated, but many fragments starting with NCG are unmethylated.

[0060] Thus, some embodiments can use the amount of CGN and / or NCG motifs to determine the methylation index of a site or a region containing multiple sites. Another example uses a cleavage profile (described in more detail below) and a distance-to-methylation relationship to determine methylation. CGN and NCG terminal motifs can be used to determine a cleavage profile, which can include counts of CGN and NCG terminal motifs (also referred to simply as motifs).

[0061] C. Probe-based counting of a specific end In the above examples, sequencing was used to determine the amount of fragments terminating at a certain distance from the CpG site or to determine the terminal motif of the fragment. Instead of sequencing, probe-based techniques can be used to measure the amount of CGNs (or longer motifs such as CGNNs) relative to NCGs. Therefore, the frequency of CGNs and NCGs from specific regions can be detected through other methods, such as qPCR, digital PCR, and digital droplet PCR.

[0062] 4 shows an example of a method for using droplet PCR to determine NCG and CGN motifs in a region of interest according to an embodiment of the present disclosure. As shown, cfDNA molecules 400 can optionally undergo DNA end pairing, A-tailing (adding A to the end of the fragment), and universal adapter ligation. DNA end repair can blunt the ends so that there are no overhangs between the two strands. Universal adapters 401 can be added to the ends of the fragments.

[0063] The adapter-ligated molecules can be separated into different reactions, such as droplets. A pair of PCR primers can be designed in such a way that one primer (e.g., a common forward primer 410) can bind to the common adapter 401 and the other primer (e.g., a region-specific reverse primer 420) can bind to a specific region of interest.

[0064] DNA molecules are amplified inside a reaction (e.g., droplet) using a pair of PCR primers. Two different fluorescent probes can be used, such as probe 430 for detecting CN-terminated fragments and probe 440 for detecting NCG-terminated fragments. Fluorescent probes specific to a particular terminal motif (such as a CGN or NCG terminal motif) are hydrolyzed to emit a fluorescent signal, thus enabling detection of the presence of the specific motif as well as quantification of the specific motif. For digital PCR, the number of reactions positive for a particular terminal motif can be counted and used to determine the amount of DNA fragments with that terminal motif in the region analyzed. For real-time PCR, the intensity of each signal can be used as a measure of the amount of DNA fragments terminating in the specific motif. The two intensities can be compared with each other.

[0065] D. Targeted sequencing of CGN and NCG Another exemplary technique for analyzing fragmentation is targeted sequencing, for example, by amplifying DNA fragments with sequences at least partially corresponding to one or more regions of interest and / or using capture probes to select such DNA fragments. Specific DNA molecules with specific terminal motifs corresponding to specific distances from the CpG site to the end of the fragment can be amplified. Targeted sequencing approaches can be used to determine cfDNA fragmentomic characteristics from regions of interest in a cost-effective manner. Targeted sequencing approaches can increase the depth of the region of interest and thus improve the accuracy of methylation analysis. DNA fragments containing CGN and NCG motifs can be enriched during sequencing library preparation. This amplification makes sequencing more efficient because the majority of sequence reads contain CGN and / or NCG motifs.

[0066] Figure 5 shows an exemplary workflow for sequencing library preparation by selectively amplifying DNA containing NCG or CGN end motifs according to an embodiment of the present disclosure. As shown, cfDNA molecules 500 can be subjected to optional steps of DNA end pairing, A-tailing, and common adapter ligation, as described above. A pair of PCR primers can be designed such that one primer 510 can bind to the common adapter region 501 (P1), and the other primer 520 or 530 can bind to the junction region between the DNA and the common adapter (P2). DNA fragments containing CG or NCG end motifs can be selectively amplified by controlling the last base in the 3' end of P2 to be CG (primer 520) or NCG (primer 530), i.e., ACG, CCG, TCG, or GCG.

[0067] These libraries are then sequenced. Such sequenced reads are used to perform analysis of CGN and NCG terminal motifs, allowing for the determination of methylation levels, patterns (e.g., levels at different sites or regions), and differences. Furthermore, selective amplification of DNA fragments containing CGN or NCG terminal motifs can enrich reads that map to regions with different methylation levels, for example, from different tissues within the body, from tumors versus non-tumor sources, or from fetuses versus pregnant mothers.

[0068] II. Exemplary Use Cases Using Cleavage Profiles and Terminal Motifs cfDNA is a mixture containing DNA molecules from different sources, which may have different methylation states at one or more CpG sites. Different methylation states affect the plasma DNA fragmentation pattern during the release of DNA molecules into plasma due to cell death (e.g., via apoptosis) or other mechanisms (e.g., active secretion). Different methylation states can also affect the clearance of cell-free DNA molecules. Therefore, the use of cfDNA fragmentation patterns allows the methylation state of CpG sites to be inferred. The influence of the methylation state on CpG on cfDNA fragmentation patterns involves plasma DNA cleavage across a series of genomic locations near or distal to the CpG site, which can be mediated by one or more DNA nucleases, such as DNASE1 and DNASE1L3.

[0069] 6 shows a schematic diagram of methylation status inference based on fragmentation patterns according to an embodiment of the present disclosure. DNA from different tissues has different methylomes. For example, blood cells will have methylation at different sites than liver, as shown for site 610. At these three CpG sites, blood cells will have different methylation levels, and liver tissue will have different methylation levels, i.e., UMU for blood and UMM for liver.

[0070] Both tissues contribute to cell-free DNA in plasma, resulting in different cell-free DNA methylation levels at each site. For example, if a first site is hypomethylated in both tissues, the methylation level will be very low in cell-free DNA. If both are methylated, the cell-free DNA methylation level will be high. If one site is methylated and the other is unmethylated, a partially methylated CpG site will result. Such a scenario is illustrated for methylation levels in Figure 6. Methylation levels represent aggregate levels as can be measured in plasma based on the exemplary methylation patterns depicted for blood cells and liver.

[0071] Meanwhile, both tissues (and other tissues in the body) also contribute their cell-free fragmentomic signatures to the plasma. Thus, if the first site is hypomethylated, there will be more unmethylated motifs (e.g., NCGs) and cleavage will be more random, while at the second site, if both are hypermethylated, there will be more methylated motifs (CGNs). Thus, the terminal motifs and cleavage profiles will be a mixture of tissues, and this information can actually infer the methylation status of each site based on its profiling. This is shown in the section on cfDNA signatures, where site 605 has an unmethylation-associated terminal motif 602 and site 608 has a methylation-associated terminal motif 601.

[0072] Figure 6 also shows two methods for inferring methylation status: method 620 uses cleavage profiles, and method 630 uses terminal motif profiles.

[0073] The cleavage profile can correspond to the amount of DNA fragments terminating at positions surrounding the CpG site. Using the methylation pattern in this example, sites that are methylated in both tissues have the highest peak. For unmethylated sites, no peak is present. For highly methylated sites, a high peak is present. For partially methylated sites (e.g., methylated in only some tissues), a lower peak is present. As shown, the cleavage profile associated with a CpG site can be used to predict the aggregate methylation index of plasma DNA molecules from different tissues, such as the liver, lung, colon, small intestine, lymphocytes, and neutrophils. The cleavage profile can be defined as the amount of fragment ends at each position over the number of genomic positions surrounding the CpG site being analyzed (also referred to as the cleavage measurement window). The number of genomic positions (referred to as the window size) can be, but is not limited to, 1 nt, 2 nt, 3 nt, 4 nt, 5 nt, 6 nt, 7 nt, 8 nt, 9 nt, and 10 nt or more. The amount of fragment ends can be a normalized value, such as, but not limited to, the number of fragment ends divided by the sequencing depth at each position.

[0074] Method 630 can use the terminal motif profile to predict methylation levels in plasma. When the methylation level is lower (e.g., the first site shown), more unmethylation-associated terminal motifs are present. When the methylation level is higher, more methylation-associated terminal motifs are present. These two signals can be used to predict whether a CpG site or region is unmethylated, highly methylated, or partially methylated. The terminal motif profile (i.e., the local terminal motif profile) of a genomic region of interest can be used to infer its corresponding methylation density (e.g., the percentage of sequenced CpGs determined to be methylated).

[0075] Because methylation is tissue-specific and fragmentomics features reflect methylation, methylation can further reflect the tissue of origin. Fragmentomics features can provide the fractional concentration of DNA from a specific tissue (e.g., of clinically relevant DNA), and can be used directly (or, in turn, used to determine methylation) to determine the fractional concentration of a specific tissue type (e.g., using tissue deconvolution when determining multi-tissue concentrations). Cleavage profiles and terminal motif profiles can be used to estimate the fraction of a specific tissue. Therefore, methylation-related cfDNA fragmentomics features, such as cleavage profiles and motif profiles, can be used to infer the contribution (fractional concentration) of cfDNA from different tissues of origin.

[0076] Some embodiments may also use fragmentomics features to infer whether a pathology (e.g., a disease such as cancer) is present.

[0077] III. Fragmentation-based site methylation prediction This section describes a technique for predicting methylation index at single-CpG resolution in a genome using cleavage profiles. Cleavage profiles can be determined in a variety of ways, including using specific terminal motifs, such as CGN terminal motifs. Exemplary definitions and results are provided below. Results are consistent regardless of how the cleavage profile is defined, e.g., regardless of the type of normalization used.

[0078] A. Normalization by total reads covering a position Cleavage profiles were constructed according to the cleavage ratio across genomic coordinates within the measurement window relative to the CpG site. The cleavage ratio at a position within the measurement window of interest can be calculated by the following formula:

number

[0079] FIG. 7 illustrates the definition of a cleavage profile using a cleavage ratio, according to an embodiment of the present disclosure. A cleavage measurement window 710 of width 11 is shown, although other widths can be used. A set of cell-free DNA fragments 700 is shown above a reference sequence 720. Relative positions are with respect to the C of the CpG site. The row marked "End" indicates the number of cell-free DNA fragments that terminate at that particular base position. The row marked "Depth" indicates the number of cell-free DNA fragments that overlap that base position. The cleavage ratio is provided using the equation shown.

[0080] The C in the CpG site is at position 0, which corresponds to the CGN end motif. Positions -5 to -1 are upstream. Position -1 corresponds to the NCG end motif. The G is at position 1, and the other four are downstream up to position 5. These positions are the cleavage measurement window 710. Within the window, all cell-free fragments are mapped inward. For each position, the number of fragments terminating at that position is calculated. This is shown in the row labeled "End."

[0081] The depth of each position is then calculated. For A at position -5, there are two fragments mapped to that position, so the depth is 2. Position -5A has one fragment terminating at that position and a depth of 2, so the truncation ratio is 0.5 (or 50%).

[0082] 1.CG cutting profile Cleavage profiles were analyzed for two sets of CG sites: putatively unmethylated and putatively methylated. Because the majority of cfDNA molecules in healthy participants originate from white blood cells (Sun et al. Proc Natl Acad Sci USA. 2015;112:E5503-5512), 1,000 putatively methylated CpG sites (beta value > 0.8) and 1,000 putatively unmethylated CpG sites (beta value < 0.1) were identified based on publicly available Illumina HumanMethylation450 BeadChip data (GSE40279) from white blood cells. If a site was methylated in white blood cells, it was assumed to be putatively unmethylated. To investigate how methylation affected cfDNA fragmentation, cleavage profiles of 1,000 putatively unmethylated CpG sites and 1,000 putatively methylated CpG sites were analyzed for eight healthy control cfDNA samples with a median of 381 million paired-end sequencing reads (interquartile range: 353-404 million). Probe-based techniques, including targeted sequencing, as described herein can also be used. "Presumed" means that methylation is derived from locus array data derived from publicly available databases.

[0083] 8 shows a comparison of cleavage profiles between a methylated CpG site 810 and an unmethylated CpG site 820 according to an embodiment of the present disclosure. The x-axis represents the nucleotide position relative to the CpG site within the measurement window. As an example, the window herein was defined as 5 nucleotides (i.e., 5 nt) upstream and downstream of the C base in the CpG site (i.e., the window size was 11 nt).

[0084] The y-axis represents the average cleavage ratio. For each sample, cleavage profiles from putative methylated or unmethylated CpG sites were merged by calculating the average cleavage ratio for each position within an 11-nt window. As shown, the merged cleavage profiles of putative methylated CpG sites showed a higher cleavage preference at position "0" (i.e., the cytosine nucleotide of the CpG site in question) (median cleavage ratio: 1.24, range: 0.91-1.37) than putative unmethylated CpG sites (median cleavage ratio: 0.54, range: 0.39-0.62) (P-value = 0.00016, Mann-Whitney U test). Thus, when C is methylated, cleavage is highly favored at C, whereas when C is unmethylated, cleavage is unfavorable at C.

[0085] 2. Cleavage Profiles in Other Sequence Contexts In addition to methylation status, sequence context can affect cleavage profiles. To study the effect of sequence context, merged cleavage profiles associated with methylated and unmethylated CpG sites were determined from a set of cleavage measurement windows (6,928,652 measurement windows in total) in which position "-1" corresponds to a cytosine nucleotide, i.e., a "CCG" subsequence.

[0086] 9-10 show cleavage profiles of windows associated with different sequence contexts, including CCGs and methylation status, according to embodiments of the present disclosure. The data are generated using bisulfite sequencing results (391 million paired-end sequencing reads) from one healthy control sample as an example.

[0087] Figure 9 shows that the merged cleavage profile of a window contains C, C, and G nucleotides at positions -1, 0, and 1. The black line 910 represents the merged cleavage profile of a window containing a methylated C at position 0, and the gray line 920 represents the profile of a window containing an unmethylated C at position 0. In this example, the methylation state (represented by M) was defined by CpG sites with a methylation index greater than 70%, while the unmethylated state was defined by CpG sites with a methylation index less than 30% (represented by U).

[0088] The cleavage ratio at position "0" appeared to be much higher at methylated CpG sites (0.90%) than at unmethylated sites (0.51%). When the C is methylated, cleavage is favored at the C, and while the C is unmethylated, cleavage is unfavorable at that C.

[0089] Figure 10 shows merged cleavage profiles of 615,465 windows containing C, G, C, and G nucleotides at positions 0, 1, 2, and 3 (i.e., CGCG subsequences). The cleavage profiles were classified into four groups, namely, "MM," "MU," "UM," and "UU," according to the methylation patterns of the two CpG sites for the CGCG subsequences. The solid black line 1010 represents the merged cleavage profile of a window containing a methylated C at positions 0 and 2, while the dashed gray line 1020 represents the profile of a window containing an unmethylated C at positions 0 and 2. The dashed black line 1030 represents the merged cleavage profile of a window containing a methylated C at position 0 and an unmethylated C at position 2, while the solid gray line 1040 represents the profile of a window containing an unmethylated C at position 0 and a methylated C at position 2.

[0090] Relatively higher cleavage ratios were observed at both cytosine positions in the CGCG subsequence for the "MM" group (0.86 and 0.89 cleavage ratios, respectively) compared with the "UU" group (0.53 and 0.42 cleavage ratios, respectively). When the methylation status of these two CpG sites differed (i.e., the "MU" and "UM" groups), relatively higher cleavage tended to occur at the methylated cytosine. These data suggested that the cleavage profile was related to the methylation pattern across a range of different CpG sites. Therefore, the cleavage profile can be used to infer the methylation pattern across a range of different CpG sites.

[0091] 3. CGN / NCG ratios in different sequence contexts For example, Figures 2-3 show that CGN and NCG motifs can be used to reflect the methylation status of CpGs immediately adjacent to the cleavage site of interest, and Figure 10 shows the cleavage pattern of two CpGs next to each other (i.e., in tandem).

[0092] To understand the relationship between cfDNA cleavage patterns and the methylation status across several adjacent CpGs, the CGN / NCG motif ratio was analyzed across different combinations of methylation status for multiple adjacent CpG sites in a molecule. The CGN / NCG motif ratio corresponds to the ratio of the amount of DNA fragments with CGN motifs to the amount of DNA fragments with NCG motifs. Here, adjacent CpGs were defined as CpG sites that are located within a 75 bp range but are not tandem. In other embodiments, adjacent CpGs can be defined as CpG sites that are located within a range of, but not limited to, 5 bp, 10 bp, 20 bp, 30 bp, 40 bp, 50 bp, 100 bp, 200 bp, 500 bp, 600 bp, 1000 bp, etc. In yet another embodiment, tandem CpGs can be used. Because the majority of cfDNA molecules (approximately 93.5%) contain three or fewer CpG sites in each molecule within a 75-bp range, we analyzed the CGN / NCG motif ratio across different combinations of methylation states for molecules with two and three CpG sites, respectively.

[0093] Figure 11 shows a schematic diagram of cfDNA grouping at adjacent CpG sites. The cleavage site 1105 is on the left at the 5' end. In this example, the range 1110 is 75 nt. The first two examples show DNA fragments with two CpG sites. The last two examples show DNA fragments with three CpG sites. The right side shows different combinations of methylation states.

[0094] For cfDNA molecules with two CpGs, there were a total of four combinations of methylation states. One combination could be a methylated CpG at the 5' end followed by a methylated CpG (represented by "MM", where "M" represents a methylated CpG and "-" represents any one or more nucleotides). Other combinations could be "MU", "UM", or "UU", where "U" represents an unmethylated CpG. For cfDNA molecules with three CpGs, there were a total of eight combinations of methylation states, namely, "MMM", "MMU", "MUU", "UMM", "UMU", "UUM", and "UUU".

[0095] Figures 12A-12B show the effect of methylation patterns with multiple adjacent CpGs on the CGN / NCG motif ratio. As shown in Figures 12A-12B, the CGN / NCG motif ratio was significantly higher in cfDNA molecules beginning with a methylated CpG at the 5' end (i.e., "MM," "MU," "MMM," ​​"MMU," "MUM," and "MUU") compared to molecules beginning with an unmethylated CpG at the 5' end (i.e., "UM," "UU," "UMM," "UMU," and "UUU"). This is similar to the behavior of a single CpG site, indicating that the dominant effect is the terminal motif at the intended cleavage site.

[0096] Interestingly, cleavage of molecules beginning with an unmethylated CpG at the 5' end was apparently enhanced by the presence of methylation of the adjacent CpG. For example, for molecules with two CpG sites, the "UM" group showed an approximately 27.9% increase in the CGN / NCG motif ratio compared with the "UU" group (P value = 0.018) (Figure 12A).

[0097] Additionally, for cfDNA molecules with three CpGs starting from an unmethylated CpG at the 5' end, there was a tendency for the CGN / NCG motif ratio to gradually decrease as more adjacent CpGs became unmethylated. The "UMM" group showed an increase of approximately 20.6%, approximately 25.0%, and approximately 52.2% in the CGN / NCG motif ratio compared with the "UMU", "UUM", and "UUU" groups, respectively (Figure 12B).

[0098] Therefore, the methylation status of CpGs directly adjacent to the intended cleavage site had a more significant effect on cfDNA cleavage than CpGs distant from the cleavage site. Additionally, the data also suggested that 5'-end CpG cleavage may be, at least in part, influenced by the methylation status of adjacent CpGs. This finding may have several potential applications.

[0099] In one embodiment, the 5'-terminal CGN and NCG motifs can be used to infer the methylation index of several adjacent CpGs in a reference genome. While it is optimal to use the terminal motif at the cleavage site of interest, for example, the CpG site may be present within a nucleosome, and therefore DNase may have fewer opportunities to cleave at such sites, resulting in an insufficient number of DNA fragments terminating at that site. In such cases, the methylation status is inferred based on the CGN / NCG ratio at the upstream CpG site. For example, as shown, a motif ratio of approximately 2 indicates that the next CpG site is methylated, while a motif ratio of approximately 1.7 indicates that the next CpG site is unmethylated. Therefore, the motif ratio at the cleavage site can determine the methylation index at the cleavage site, but it can also determine the methylation index at one or two downstream CpG sites.

[0100] In another embodiment, the cfDNA cleavage pattern associated with one or more adjacent CpGs can be used to enhance the diagnostic power of pathological conditions.For example, the cfDNA cleavage pattern associated with the number of CpGs within a certain nucleotide distance that show the same methylation pattern can be used to facilitate the selection of diagnostic markers.As another example, the fragmentomics features derived from cancer-specific hypomethylation markers can be used for cancer diagnosis.In some cases, the use of cleavage patterns across CpG sites with adjacent hypomethylated CpGs can be superior to the use of CpGs with adjacent hypermethylated CpGs.

[0101] B. Normalization by total ends in the region Another example of normalization uses the number of fragments that terminate in a region surrounding the site, which may or may not be the same as the window used to determine the cleavage profile.

[0102] Figure 13 illustrates normalization of cleavage profiles using cleavage density according to an embodiment of the present disclosure. The black bar 1305 represents a reference genome containing CG1301. Similar to the example using cleavage ratios, the number of fragments terminating at a CpG site can be calculated to infer the methylation status of that CpG site. For normalization, the number of ends at each position was divided by the total ends in the entire region. This is another exemplary technique that can be used to normalize the amount of DNA fragments with a particular end motif or end at a particular position.

[0103] Similar analysis is performed to determine the profile of various CpG sites for cleavage density, as was performed for cleavage ratio. For example, one CG, two CG, and three CG analyses are performed. Sequencing analysis uses merged data from multiple samples to obtain higher sequencing depth. If the methylation index exceeds the first threshold (80% in this example), the site is methylated, and if the methylation index is below the second threshold (30% in this example), the site is not methylated. If the methylation index of a site is between the thresholds, it is not used in this example. In other embodiments, intermediate classifications can be used.

[0104] Figure 14 provides a cleavage profile using cleavage densities of various methylation combinations of two CpG sites according to an embodiment of the present disclosure. The CpG sites are tandem in this example, as in Figure 10. There are four groups according to the methylation status of the first and second CGs. The two dashed lines 1403 and 1404 represent the positions of these two cytosines.

[0105] As can be seen, in the top two profiles, the first C1401 is unmethylated and there is no peak at C1201. However, in the bottom two profiles, the first C is methylated, and there is a dramatic increase from -1 to 0. The second C is methylated for the second and fourth profiles. In these profiles, there is a peak for the second C1402. The second C is unmethylated for the first and third profiles, and the peak is low or absent. The cleavage density 1405 has a high value, but there is no relative increase. The relative increase is more important for inferring the methylation state rather than the exact number at a position. These results are consistent with the results in Figure 10, which uses the cleavage ratio.

[0106] Figure 15 provides cleavage profiles using cleavage densities of various methylation combinations of three CpG sites according to an embodiment of the present disclosure. Different profiles correspond to different combinations of methylation states of different sites. As can be seen, the top two profiles have the first CpG methylated, while the bottom two profiles are unmethylated. Also, for the first C1501, prominent peaks 1503 and 1505 were observed for the top two profiles. However, the peaks disappear in the bottom two profiles for the first C1501.

[0107] The second CG is methylated for the first and third profiles. Peak 1504 is observed for the second C1502. The second CG is unmethylated for the second and fourth profiles, and the peak is reduced or absent.

[0108] The third CG is methylated and has a visible peak in the first three profiles, whereas the third CG is unmethylated in the fourth profile and has a very weak peak.

[0109] Figure 16 shows the cleavage density of a window containing a CCG and methylation status according to an embodiment of the present disclosure. Because DNASE1L3 prefers to cleave at CC or CCC ends, a region containing a CCG was analyzed. However, if the CG is methylated, it is hypothesized that cleavage exactly before the CG (between the C) is preferred. And if it is methylated, peak 1601 is observed. Therefore, if the CG is methylated, cleavage between two Cs is more preferred. If the CG is unmethylated, the peak disappears. These results are consistent with the results in Figure 9 using the cleavage ratio.

[0110] Figure 17 shows a comparison of cleavage profiles between methylated and unmethylated CpG sites using cleavage density, according to an embodiment of the present disclosure. Figure 17 corresponds to Figure 8 and shows similar results, but uses a different definition of the cleavage profile, i.e., using cleavage density as opposed to cleavage ratio. As in Figure 8, putative hypermethylated or hypomethylated CpG states were selected from published leukocyte methylation array data. A set of 1000 CpGs that are methylated in leukocytes (beta value > 0.8) and 1000 CpG sites that are unmethylated in leukocytes (beta value > 0.1) were randomly selected.

[0111] Each red line 1720 represents a sample with 1000 methylated CpG regions, and each blue line 1730 represents a sample with 1000 unmethylated CpG sites. A dramatic peak is observed in the red line, indicating that in highly methylated regions, the methylation status can be inferred by this peak. For the blue lines, this peak is now reduced or even disappeared. Thus, when a CpG is methylated, there is a spike of fragments terminating at position 1701. The spike can be seen in the schematic graph 1707 representing the cleavage profile. When the C is unmethylated, this CG-terminal preference disappears, as shown in the schematic graph 1709. The results demonstrate that the cleavage profile can predict methylation status at single-CpG resolution.

[0112] C. Results using a support vector machine In some embodiments, machine learning (ML) techniques can be used. Such techniques can use two or more quantities within a cleavage measurement window and can include the full cleavage profile in a window around the CpG site. Machine learning techniques can also use sequence context, such as the entire sequence or an instance of a k-mer, which can be tracked in various ways. The use of sequence context can also be used outside of machine learning; for example, the relative increase in positions from -1 to 0 can vary depending on the base preceding the C. Machine learning can identify such patterns in any of the positions used in the feature vector input into the ML model. Various ML models can be used. This section presents results using a support vector machine (SVM) to classify sites as hypermethylated or hypomethylated.

[0113] 1. Use cutting profiles only In some embodiments, cleavage profiles were constructed using cleavage ratios from an 11-nt cleavage measurement window centered on the CpG site. The cleavage profiles from the Watson and Crick strands were used to train a machine learning model used to infer the methylation status of the CpG site at the center of the window. In one embodiment, a support vector machine (SVM) was used to predict whether the aggregate methylation index at a CpG site was greater than 95% or less than 20% based on the cleavage profile associated with that CpG site.

[0114] Figure 18 shows a workflow for using SVM to classify CpG sites as hypermethylated or hypomethylated according to an embodiment of the present disclosure. For the training set, 33,147 hypomethylated and 33,147 hypermethylated CpG sites in the reference genome are used. The actual classification of each site was determined using bisulfite sequencing. In this example, hypomethylation is a methylation index greater than 20 percent, and hypermethylation is a methylation index greater than 95 percent. Half of the sites were used for training, and the other half were used for testing. The trained model can predict whether a CpG site is hypomethylated or hypermethylated based on the cleavage profile. The input feature vector is the cleavage ratio at 11 positions, as shown for each site. Feature vector 1810 is an example of such a feature vector.

[0115] Figure 19 shows the performance of single-CpG site methylation status prediction using a support vector machine (SVM) based on cleavage profiles, according to an embodiment of the present disclosure. The two plots show data from the Watson and Crick strands. Using only data from the Watson or Crick strands, an AUC of approximately 0.8 can be achieved. In this example, the sequencing depth at each position was used to normalize the cleavage ratio within the cleavage profile.

[0116] Typically, the methylation patterns at CpG sites in the Watson strand and the corresponding sites in the complementary Crick strand will be symmetric. Cleavage profiles from both the Watson and Crick strands can be merged to train a machine learning model to infer the methylation status of the CpG site in the center of the window. Thus, in some embodiments, cleavage profiles from each strand are used. Thus, the feature vector may have 22 values ​​in this example of 11 positions in the measurement window. In most cases, the two strands share the same methylation status. Therefore, both strands can be used to estimate the methylation status of both CpG sites.

[0117] Figure 20 shows the performance of single-CpG site methylation status prediction using a support vector machine (SVM) based on cleavage profiles from both strands, according to an embodiment of the present disclosure. An AUC of approximately 0.89 was achieved in the test dataset, which appeared to be superior to either model based on data from the Watson or Crick strands alone (P-value < 0.0001, DeLong test). When the cutoff of the SVM-based output probability was set as 0.5 to determine whether a CpG site was methylated or not, a sensitivity of 83.8% could be achieved, along with a specificity of 77.2%. When the cutoff of the SVM-based output probability was set as 0.55 to determine whether a CpG site was methylated or not, a sensitivity of 81.4% could be achieved, along with a specificity of 80.2%. These data suggested that it is feasible to use cleavage profiles to predict methylation status using machine learning algorithms.

[0118] The sequencing depth used for the training samples (e.g., 10x) can be higher than that used for the test samples.

[0119] 2. Using Array Context The use of sequence contexts around CpG sites can be analyzed to improve model performance. For example, a training dataset containing cleavage profiles associated with methylated and unmethylated CpG sites can be divided into four categories according to the nucleotide type at the "-1" position of the Watson strand. Sequence contexts can be used by training different models, one for each of the four sequence contexts.

[0120] 21A-21B show the performance of single-CpG site methylation status prediction using a support vector machine (SVM) based on cleavage profiles conditioned on different sequence contexts, according to an embodiment of the present disclosure. Using sequence context can improve accuracy.

[0121] Figure 21A shows the combined data from the Watson (W) and Crick (C) strands. The different lines represent the ROC curves when using different subsequences at positions -1, 0, and 1 within the measurement window of the Watson strand. Having a C at the -1 position provides the highest accuracy. For this embodiment, there are essentially four different models: A, C, G, and T on the Watson strand.

[0122] As shown in Figure 21A, focusing on cleavage measurement windows with C, C, and G nucleotides (i.e., CCG subsequences) at positions -1, 0, and 1 in the Watson strand improves model performance, as evidenced by an increase in AUC from 0.89 to 0.94 (P-value < 0.0001, DeLong test). Using cleavage measurement windows with ACG, GCG, and TCG subsequences at positions -1, 0, and 1 in the Watson strand yields AUC values ​​of 0.86, 0.91, and 0.89, respectively. Thus, there was an improvement compared to the results in Figure 20 for the two sequence contexts (CCG and GCG), the same for TCG, but lower for ACG.

[0123] Figure 21B shows the combined data from the Watson (W) and Crick (C) strands with CCG subsequences at positions -1, 0, and 1 within the measurement window of both strands. Thus, using sequence context from both strands improves the AUC from 0.94 to 0.97. If this is extended to both strands and all four bases, there are 16 different models.

[0124] Thus, in some embodiments, the use of cleavage profiles from windows with CCG subsequences at positions -1, 0, and 1 in both the Watson and Crick strands further enhanced model performance by an AUC of 0.97. These results suggested that selectively analyzing cleavage profiles according to sequence context and a combination of Watson and Crick strand data synergistically improved model performance in distinguishing between methylated and unmethylated CpG sites.

[0125] 3. Windows of various widths In the example of Figure 20, an 11-bp measurement window was used to determine whether a given site in the genome is hypermethylated or hypomethylated. Windows of various widths, e.g., 2, 3, 4, 6, 7, 8, 9, 10, 11, and 20, can be used in any of the embodiments described herein. The results for such widths are shown in Table 1. Table 1 summarizes the AUCs of methylation predictions by SVM models with different window sizes. The AUC (overall) is analyzed by the method used in Figure 20, and the remaining four are analyzed by the method used in Figure 21A using the sequence context of the reference genome at position -1. The measurement uses fragments from both the Watson (W) strand and the Crick (C) strand, although the measurement can also use only one strand. The sequence context is defined with respect to the Watson strand, but can also be defined with respect to the Crick strand as well as both. Additionally or alternatively, other sequence contexts (other than the -1 position) can be used, e.g., at positions -3, -2, 1, 2, and 3. [Table 1]

[0126] For the 1 bp at position 0, the cleavage ratio is determined using the number of DNA fragments that terminate at the CpG site, i.e., with a C at the CpG site. For this example, the cleavage ratio is determined using the sequencing depth, as described in Section III.A. The SVM model can receive the Watson strand cleavage ratio and / or the Crick strand cleavage ratio as input. Other machine learning models, such as decision trees or neural networks, can be used.

[0127] If only one cleavage ratio is determined (e.g., for only one strand), a first cutoff / threshold can be used to distinguish between hypermethylated and non-hypermethylated, and a second cutoff / threshold can be used to distinguish between hypomethylated and non-hypomethylated. If additional cleavage ratios are used (e.g., as part of a cleavage profile), additional cutoffs / thresholds can be used. The cleavage ratio can take various forms, as described herein, such as normalization by the number of DNA fragments covering the position, the number of DNA fragments in the region surrounding the site, or the number of DNA fragments terminating at one or more specific positions around the site. Any of these normalizations can be used for the cleavage profile.

[0128] For a 1 bp at position -1, the cleavage ratio is determined at position -1 to the left of the CpG site. Such positions correspond to the NCG end motif. For the AUC column using ACG, the DNA fragment has an ACG at the end, and similarly for other AUC values ​​using each sequence context.

[0129] For 1 bp at position 1, the cleavage ratio is determined at position 1 to the right of the CpG site. For each of the remaining measurement windows, the specific position used is identified in the column next to the window size.

[0130] The data in Table 1 was generated using the 33,147 hypomethylated and 33,147 hypermethylated CpG sites described above. In this example, hypomethylation is below 20 percent for the methylation index, and hypermethylation is above 95 percent for the methylation index.

[0131] D. Results of using a convolutional neural network with array context In some embodiments, a machine learning model (e.g., a neural network) can predict methylation by considering the sequence context and cleavage profile (e.g., more than one cleavage density or cleavage ratio) surrounding a CpG site. In one embodiment, a convolutional neural network (CNN) was implemented to synergistically use the sequence context and cleavage profile present in the measurement window to determine whether a CpG site in a genome is methylated.

[0132] 22 shows a schematic diagram of methylation state prediction using a CNN model based on cleavage profile and sequence context, according to an embodiment of the present disclosure. A methylated state can be defined by a methylation index greater than 95% (or other values ​​described herein, for example), while an unmethylated state can be defined by a methylation index less than 20% (or other values ​​described herein, for example). For illustrative purposes, the 5 nt upstream (e.g., ATCTG) and 5 nt downstream (e.g., GAGTA) of the cytosine in the CpG site being analyzed are in the Watson strand. [ka] The relative positions of this sequence corresponded to -5, -4, -3, -2, -1, 0, +1, +2, +3, +4, and +5, respectively. The center position of "0" corresponded to the cytosine of the CpG site that was subjected to methylation status analysis.

[0133] The fragment cleavage ratios for each individual position can be constructed into a two-dimensional (2-D) matrix 2220 according to the sequence context. For example, for position -1, corresponding to the guanine ("G") base, the cleavage ratio associated with "G" (1.40) was entered in the corresponding cell between the "-1" column and the "G" row. The remaining rows, corresponding to "A," "C," and "T" in the Watson strand, were filled with "0." The cleavage profile and sequence context ('5-[TTACT]C[GCAGA]-3') derived from the Crick strand were similarly processed. Data matrices from the Watson and Crick strands can be combined for downstream analysis. Thus, such a data matrix can contain both sequence context and cleavage profile information. Several data matrices associating known methylation states, i.e., methylated CpG sites (output value as "1") and unmethylated CpG sites (output value as "0"), were used to train the CNN model.

[0134] As shown, the cleavage ratio was calculated for both strands. The matrix includes the base (nucleotide), position, and cleavage ratio (one-hot encoding is used in this example). For example, at -5 in the W strand, the first base is A and the cleavage ratio is approximately 0.25. The two matrices can be combined into a composite matrix 2230. This composite matrix can be the input measurement reference for a CNN model with a convolutional layout to extract features from these matrices. The extracted features can be combined into a linear level to calculate the likelihood that a CpG site is methylated or unmethylated. An AUC of approximately 0.96 can be achieved.

[0135] In one embodiment, the CNN model utilized two one-dimensional (1D) convolutional layers, each with 64 filters with a kernel size of 4; however, other hyperparameters, such as 50-100 filters or kernel sizes of 3-6, could be used. A rectified linear unit (ReLU) activation function was used in these convolutional layers. A batch normalization layer was then applied, followed by a dropout layer with a dropout rate of 0.5. A flattening layer was then added, followed by a fully connected layer containing 128 neurons using a ReLU activation function. Finally, an output layer with one neuron was applied using a sigmoid activation function to generate a probability score for each CpG site being methylated (i.e., a methylation score). The CNN model program was implemented based on the PyTorch machine learning framework (https: / / pytorch.org / ). In various embodiments, the number of convolutional layers could be 1, 3, 4, 5, 6, 7, 8, 9, 10-20, 30-40, 40-50, or more. The number of filters in each convolutional layer may be 1-10, 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, 10-20, or more. 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 the last layer of the multiple hidden layers and configured to output an output data structure.

[0136] The optimal parameters of the CNN model were obtained by iteratively adjusting the model parameters when the overall prediction error between the output score by the sigmoid function and the desired target output (binary value: 0 or 1) reached a minimum. The overall prediction error was measured by the cross-entropy loss function (pytorch.org) in the deep learning algorithm. The model parameters learned from the training dataset were used to analyze the test dataset and output a probability score (referred to as the methylation score in this study) indicating the likelihood that a CpG site is methylated.

[0137] Methylation scores can range from 0 to 1. In one embodiment, a methylation score of 1 indicates that a CpG site can be considered methylated. A methylation score of 0 can indicate that a CpG site can be considered unmethylated. In some embodiments, a cutoff value can be used to classify whether a particular CpG site is classified as "methylated" or "unmethylated" based on the methylation score cutoff. Possible cutoff values ​​include 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 55%, 60%, 65%, 70%, 75%, 80%, 85%, 90%, or 95%. Methylation scores of CpG sites above a predetermined cutoff can be classified as "methylated," while the probability of methylation for CpG sites below a predetermined cutoff can be classified as "unmethylated." The desired cutoff can be obtained from the training data set using, for example, receiver operating characteristic (ROC) curve analysis.

[0138] 23A shows the performance of predicting the methylation status of a single CpG site using a CNN based on cleavage profiles and sequence context, according to an embodiment of the present disclosure. The solid line represents the performance of a model trained on cleavage profiles and sequence context, and the dashed line represents the performance of a model trained on genomic context only. The genomic context is trained only on the reference sequence pattern without any fragmentation information.

[0139] Based on the methylation score, the CNN model was able to achieve an AUC of 0.96, which was superior to the SVM-based model (0.89) without sequence context (P-value < 0.0001, DeLong test). This AUC was for all sequence contexts. When the model did not use cleavage patterns but only sequence context, performance decreased to an AUC of 0.79. These results suggested that the combined use of cleavage profiles and sequence context significantly improved model performance in determining methylation status. In other embodiments, the cleavage profiles and sequence context can be used to infer a methylation index (e.g., a specific percentage or percentage range) instead of a binary value of hypermethylation or hypomethylation.

[0140] In some embodiments, the machine learning model may include, but is not limited to, a convolutional neural network (CNN), linear regression, 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 transform-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 for applications with noise (DBSCAN), a random forest algorithm, adaptive boosting (AdaBoost), eXtreme Gradient Boosting (XGBoost), a support vector machine (SVM), or a hybrid model including one or more of the models proposed above.

[0141] E. Results of using various positions As shown above, the cleavage distance reflects the methylation index, specifically, distances 0 (corresponding to the terminal motif CGN) and 1 (corresponding to the cleavage site between C and G), as shown, for example, in Figures 2A-2B and 3. Distance 1 corresponds to positions -1 (corresponding to the terminal motif NCG) and +1 (corresponding to the cleavage site between C and G). The cleavage ratios at various positions (e.g., determined using the amounts at specific positions) also indicate differences between methylation levels, as shown, for example, in Figures 8-10 and 17. Additional data are provided in this section showing the distinction between different methylation levels using the amount of cell-free DNA molecules terminating at the first position of a CpG site, where the first position is one of -1 and +1. As shown, the -1 position, the 0 position, or the +1 position can be used.

[0142] Figure 23B shows a comparison of the sum of the average cleavage ratios at positions -1 and +1 between hypermethylated and hypomethylated CpGs for 39 ucfDNA samples. The sum of the average cleavage ratios for 18,280,640 hypermethylated CpG sites and 2,576,102 hypomethylated CpG sites was analyzed for 39 healthy control urinary cfDNA samples with a median of 75.8 million paired-end sequencing reads (interquartile range: 60.5 million to 155 million). Each dot represents one subject. As shown, the cleavage ratios at positions -1 and +1 differ between hypermethylated and hypomethylated sites. Such data demonstrate that methylation information can be predicted without using fragment information at position 0.

[0143] Figure 24A shows the performance of methylation status prediction of a single CpG site using a CNN model based on proximal cleavage profiles and sequence context, excluding position 0. Sequence context was used as described above. Without the 0 position, the composite matrix for a + / - 5 base window has 10 columns instead of 11 columns as used in Figure 22. An AUC of 0.91, slightly lower than 0.96, was achieved using the 0 position, but still much higher than using sequence context alone (AUC = 0.79). The same dataset as Figure 23B was used.

[0144] As noted above, any of positions -1 to +1 can be used alone. Some embodiments can use a ratio (e.g., CGN / NCG ratio), or the amount of fragments terminating at a particular position relative to the CpG site, while other embodiments can use a vector of cleavage ratios at multiple positions.

[0145] Figure 24B shows the proximal cleavage profile and sequence context, excluding positions -1, 0, and 1. Figure 24B shows the performance of methylation status prediction of a single CpG site using CNN based on the text. Figure 24B shows that it is feasible to use a set of contiguous or non-contiguous positions within the truncation profile window to form the data structure that is input into the machine learning model for training. However, if some positions, such as -1, 0, and +1, significantly contribute to classification, performance may be reduced.

[0146] Figures 25A-25F are boxplots illustrating the difference in mean cleavage ratios between hypermethylated and hypomethylated CpGs at various positions in plasma and urine (ucfDNA). Figure 25A shows the difference between hypermethylated and hypomethylated CpGs at position -1 in plasma. Figure 25B shows the difference at position 0 in plasma. Figure 25C shows the difference at position +1 in plasma. The boxplots below illustrate the difference in mean cleavage ratios at positions -1 (Figure 25D), 0 (Figure 25E), and -1 (Figure 25F) in ucfDNA. The separation is very clear, with p<0.001. Position 0 provides the best discrimination. Position -1 provided better discrimination for plasma than for urine, while position +1 provided better discrimination for urine than for plasma.

[0147] F. Quantitative methylation density determination Above, examples were classifications of hypermethylation, e.g., methylation density greater than X% (e.g., 80%), or hypomethylation, e.g., less than Y% (e.g., less than 30%). Some embodiments may use a calibration function (e.g., as described in the terminology section) to determine a narrower range or to estimate a specific value. For example, a methylation recognition assay may be used to determine the methylation level of a site. Cleavage information, such as any cleavage parameter described herein, such as cleavage density or cleavage ratio, may also be determined.

[0148] This measured methylation level at a site and the corresponding cleavage information / parameter (e.g., cleavage ratio or cleavage density) can form a calibration data point, where the cleavage parameter is a calibration value corresponding to a known / measured methylation level. Given the relationship between cleavage and methylation, when the cleavage parameter changes, the methylation level also changes. A calibration function can define the proportional change in methylation level when the cleavage parameter changes. In this way, when the cleavage parameter of a new sample (e.g., sequenced without using a methylation recognition assay) is measured, the cleavage parameter can be compared to the calibration function (e.g., to a calibration value corresponding to a known methylation level). In one embodiment, the cleavage information / parameter (e.g., cleavage profile or terminal motif) can be combined with a methylation recognition assay (e.g., using a comprehensive kinetic model on data generated by single-molecule real-time sequencing, as in U.S. Publication No. 2021 / 0047679) to improve the final classification accuracy for determining DNA methylation profiles.

[0149] The calibration data points can be used to train a model, e.g., a regression model. The regression model can be used to determine a quantitative value, which can be in a range, e.g., a 5% (e.g., 40-45% density), 10%, 15%, 20%, 25%, or 30% range. The regression can provide a calibration function. In some embodiments, the likelihood from an ML model (e.g., a neural network) can be used as a methylation index.

[0150] Figures 26A-26C show a comparison of various techniques for determining quantitative values ​​of methylation levels (e.g., density) for CpGs with different methylation levels in eight plasma DNA samples. Figure 26A shows the average CGN / NCG motif ratio for various ranges of methylation density. Specifically, ranges with 20% width are shown: 0-20%, 20-40%, 40-60%, 60-80%, and 80-100%. Figure 26B shows the average cleavage ratio at position -1 for CpG sites. Figure 26C shows the average cleavage ratio at position 0 for CpG sites.

[0151] In the WGBS data of plasma DNA samples, CpGs were separated into five groups according to specific ranges of methylation index: 0-20%, 20-40%, 40-60%, 60-80%, and 80-100%. For the same samples, the average CGN / NCG motif ratio for each set of CpGs (Figure 26A), the average cleavage ratio at position -1 (Figure 26B), and the average cleavage ratio at position 0 (Figure 26C) were calculated and plotted. Each dot represents one subject.

[0152] All three parameters show a consistent relationship (increase or decrease) with methylation density. The CGN / NCG motif ratio gradually increases as methylation density increases. The cleavage ratios at positions -1 and 0 correlate negatively and positively with different methylation density ranges. These results reflect that the CGN / NCG motif ratio and cleavage ratio can distinguish between hypermethylated and hypomethylated CpGs. Furthermore, the CGN / NCG motif ratio and cleavage ratio can be used to predict the range of methylation levels. Each plot shows a calibration curve that can be used to determine a quantitative value (at least within a range) using the amount of cell-free DNA fragments terminating at position 0 or -1.

[0153] We also tested SVM regression using cleavage profiles from an 11-bp window (e.g., -5 to 5). SVM models were trained with cleavage profiles from CpGs corresponding to different methylation index ranges. For example, we trained SVM regression models based on the ranges of 0-5% (Group 1), 50-55% (Group 2), and 95-100% (Group 3), i.e., SVR 1. We trained another SVM regression model based on the ranges of 0-10% (Group 1), 45-55% (Group 2), and 90-100% (Group 3), i.e., SVR 2. We trained another SVM regression model based on the ranges of 0-20% (Group 1), 40-60% (Group 2), and 80-100% (Group 3), i.e., SVR 3.

[0154] Table 2 shows the performance of predicting the methylation status of a single CpG site within different methylation index ranges using SVM regression models based on proximal cleavage profiles. As shown in the second column (Pearson r), the predicted methylation scores correlated well with the methylation index of the tested CpGs (i.e., Pearson r = 0.73, 0.77, and 0.74). For each SVM regression model, the methylation index of each predicted group closely matches the actual methylation index range. In some embodiments, predicted values ​​may be expressed with other confidence intervals, including but not limited to, 99%, 95%, 90%, 85%, 80%, 75%, 70%, 65%, 60%, 55%, and 50%. [Table 2]

[0155] G. Methods for Determining Methylation at Sites in a Genome Provided is a method for determining methylation at a site in a genome using cleavage information.Methylation can be, for example, the characteristic of a sample at a position in a genome, as defined relative to a reference genome.Cutting information can be of various types, as described herein.Various normalizations can be used, and the cut information of more than one position can be used.

[0156] 27 is a flowchart of a method 2700 for determining methylation at sites based on fragmentation, according to an embodiment of the present disclosure. Various examples of method 2700 are described above. Method 2700 can be implemented in part or in whole using a computer system.

[0157] At block 2710, a plurality of cell-free DNA molecules from the subject's biological sample are analyzed. Various techniques can be used for such analysis in any of the methods described in this disclosure. For example, analysis can be performed using sequencing, such as massively parallel sequencing, targeted sequencing, and single molecule sequencing (e.g., using nanopores or using real-time single molecule sequencing (e.g., from Pacific Biosciences)). Exemplary PCR techniques include real-time PCR and digital PCR (e.g., droplet digital PCR). Analysis can include the physical steps of performing such assays and receiving measurement data obtained from such assays, or can simply include receiving measurement data.

[0158] Analyzing the cell-free DNA molecule can include determining a genomic location in a reference genome corresponding to at least one end of the cell-free DNA molecule. For example, one or more sequence reads of the DNA molecule (e.g., paired reads at the ends or reads for the entire molecule) can be aligned to the reference genome using any of a variety of alignment techniques understood by those skilled in the art. Alignment can be relative to part or all of the reference genome. As another example, probe-based techniques can identify a DNA molecule as originating from a particular location, for example, by emitting a particular color for a particular probe corresponding to a particular genomic location. Location determination can be relative to part or all of the reference genome, for example, if only a portion of the genome is analyzed. By way of example, the amount of genome analyzed can exceed 0.01%, 0.1%, 1%, 5%, 10%, or 50%. Such analysis can be performed using other methods described herein.

[0159] As with other methods described herein, analyzing the plurality of cell-free DNA molecules can include measuring the size of the cell-free DNA molecules, which can be performed in a variety of ways, for example, using physical separation (such as electrophoresis) and / or sequencing (such as whole molecule sequencing or alignment using paired-end reads).

[0160] In block 2720, a first amount of cell-free DNA molecules terminating at a first position within a window around the CpG site is determined. The first position can be between -1 and +1 of the window. For example, the first amount can be cell-free DNA at the -1 position, the first amount can be cell-free DNA at the 0 position, or the first amount can be cell-free DNA at the +1 position.

[0161] The first quantity can be determined in various ways. For example, if the analysis uses sequence reads, block 2720 can determine the number of reads that align to CpG sites in the reference genome and terminate with CG (having their 5' ends on the left side on the Watson strand) or GC (having their 5' ends on the right side on the Crick strand). Thus, the quantity can be a count of DNA molecules. As another example (e.g., as depicted in Figure 4), probe-based technology can be used. For example, the number of positive reactions for a particular probe can be used. As yet another example, the strength of the signal (e.g., electrical or optical) from a particular probe (e.g., from real-time PCR) can be used.

[0162] A second quantity of cell-free DNA molecules terminating at a second position (different from the first position) within the window around the CpG site can optionally be determined in block 2720. The second quantity can be determined in a similar manner to the first quantity. The quantities can be raw numbers (e.g., without normalization) to be compared to each other. In other examples, relative frequencies can be used. Examples of relative frequencies include cleavage ratios and cleavage densities, as can be determined using the normalization techniques described herein. The two quantities can be used to determine relative frequencies to be used in subsequent steps. As a further example, such normalization, or any other normalization described herein, can be performed by analyzing the same number of DNA molecules, generating the same number of reads, or using samples of the same size (e.g., a specified volume (e.g., number of milliliters)).

[0163] The second amount can be a portion of a larger amount. For example, the total number of DNA molecules covering a first position (e.g., a CpG site) can be determined (also referred to as depth). At least some of these DNA molecules terminate at a different position other than the first position. Thus, the total number covering the first position can include a second amount terminating at a second position surrounding the CpG site. Techniques that use such a total number for normalization use the second amount. Thus, when the first amount is normalized, the normalization can use the number of cell-free DNA molecules covering the CpG site. As another example, the normalization can use the number of cell-free DNA molecules terminating within a region containing a CpG site. As yet another example, the normalization can use the average or median depth of cell-free DNA molecules in a region containing a CpG site.

[0164] The second position can be any position within the measurement window, exemplary windows being described herein. For example, if the first position is position 0, the first position can be +1 or -1 from the CpG site. As another example, the measurement window can be at least -2 to +2 from the CpG site, and thus be greater than 5 bases in width.

[0165] In block 2740, the first amount is used to determine a methylation classification of the CpG site in the subject's genome. Optionally, the second amount may also be used. Various methylation classifications of the site may be determined. For example, it may be determined whether the site is hypomethylated or hypermethylated in the sample and the subject. Thus, the methylation classification may be whether the CpG site is in a hypermethylated state (also referred to as a methylation state) or a hypomethylated state (also referred to as an unmethylated state) in the subject's genome, determined by analyzing multiple cell-free DNA molecules in the sample. The methylation state may have a methylation density above a first threshold of at least 70%. The unmethylated state may have a methylation density below a second threshold of 30% or less. The classification may be any methylation level, which may be provided qualitatively (e.g., high, medium, or low, as may be defined using a certain percentage threshold) or quantitatively (e.g., a certain amount, such as a percentage, which may be provided with a given resolution / range).

[0166] The classification can be determined in various ways. For example, a separation value can be determined using a first amount and a second amount. The separation value can be compared to a calibration value. The calibration value can be determined using cell-free DNA molecules derived from one or more calibration samples and located at CpG sites with known classifications. The known classifications can be determined using a methylation-recognition assay or from the results of other assays that use a methylation-recognition assay. The calibration value is the threshold / cutoff above or below which a particular classification is determined. For example, in Figure 17, a calibration value of approximately 22 for the cleavage ratio at position 0 distinguishes between hypermethylated and hypomethylated sites.

[0167] As another example for classification, sites can be identified as having a methylation level of a particular percentage or percentage range (e.g., 5%, 10%, 15%, 20%, 25%, 30% or less). An example of a 20% range is 40%-60%. Thus, the classification can be a quantitative value. Comparing the separation value to a calibration value can include comparing the separation value to a calibration function, for example, as described above in Section E.

[0168] The classification can use at least two positions other than the first position, e.g., other than a CpG site if the first position is position 0. For each of the at least two positions within the window, the respective amounts of cell-free DNA molecules terminating at that position can be determined. As part of determining the classification, the first amount of cell-free DNA molecules terminating at a CpG site can be compared to the respective amounts of cell-free DNA molecules terminating at that position. Such a comparison can determine the difference between the amount from position -1 to position 0 and the amount from position 1 to position 0. For example, Figures 8 and 17 both have different amounts of increase from -1 to 0 and decrease from 0 to 1. Additionally, such use of two or more positions can include machine learning techniques, such as those described in sections C and D above.

[0169] When machine learning techniques are used, a feature vector, including the respective amounts and the first amount, can be input into the machine learning model as part of determining the classification. Cell-free DNA molecules located at CpG sites with known classifications can be used to train any such machine learning model described herein. The first position and at least two positions can include all positions within a window of at least +5 to -5 from the CpG site. As described in Sections C.2 and D, the machine learning model can use sequence context within the window. In one example, the machine learning model is trained on sequence context within the window (e.g., as shown in Figure 21A).

[0170] In another example, the feature vector can include sequence context (e.g., as shown in Figure 22). In such an example, the feature vector can form a matrix with each row corresponding to a base in a strand (potentially in both the Watson and Crick strands). As shown in Figure 22, the columns can include non-zero quantities in the row corresponding to the base at each position. A CNN model can be used when the feature vector forms a matrix.

[0171] More than one CpG site may be within the measurement window (e.g., as described in Figures 10, 14, and 15), and both may be classified. In such a situation, the CpG site may be a first CpG site having a first classification. The window may include a second CpG site and may include at least two positions other than the first CpG site (e.g., when the first position is position 0) and the second CpG site. A third amount of cell-free DNA molecules terminating at the second CpG site may be determined. For each of the at least two positions within the window, the respective amounts of cell-free DNA molecules terminating at that position may be determined, for example, as described above. When machine learning is used, the feature vector may include the respective amounts, the first amount, and the third amount. The feature vector may be input into a machine learning model as part of determining the classification of the first CpG site and determining the second classification of the second CpG site.

[0172] The methylation classification of more than one CpG site can be determined using a first amount and a second amount of a first CpG site at the cleavage site, for example, as described in Section II.A.3. Thus, some embodiments can use the first amount and the second amount to determine the methylation classification of different CpG sites in a subject's genome. The different CpG sites can be within a specified range, such as 50 nt, 75 nt, or 100 nt downstream of the 5' end of the CpG site.

[0173] IV. Prediction of methylation in regions based on terminal motifs In some embodiments, the methylation of a region can be predicted. The methylation classification can be, for example, hypomethylated or hypermethylated for a site, or a specific methylation level classification that can be within a range of values ​​(e.g., 1%, 2%, 3%, 4%, or 5%). Various techniques can be used, including machine learning.

[0174] A. Size of the area to be analyzed A region can be the entire genome or a specific portion of the genome, which can be on one chromosome or span multiple chromosomes. Regions can be separated from each other, i.e., composed of discrete, different windows. Thus, methylation levels can be determined aggregately for multiple regions (e.g., 2, 3, 4, 5, 10, 20, 50, etc.), which can be on multiple chromosomes (e.g., 2, 3, 4, 5, etc.). The first section below analyzes the entire genome, and subsequent sections focus on specific portions of the genome.

[0175] 1. Global Figure 28 shows the methylation index of CGN and NCG end motifs according to an embodiment of the present disclosure. Each dot represents the average methylation index of each of the four motifs within a group of CGN and NCG end motifs. N represents A, C, G, or T. Figure 28 is similar to Figure 2B, with distances 0 and 1. CGN corresponds to position 0 in Figure 7, and NCG corresponds to position -1 in Figure 7.

[0176] The plot shows that the highest methylation index (median: 88%) was observed at CpG sites associated with 5'-CGN terminal motifs (i.e., CGA, CGT, CGC, CGG). In contrast, the lowest methylation index (median: 69%) was present at CpG sites associated with 5'-NCG terminal motifs (i.e., ACG, TCG, GCG, CCG). Cleavage at methylated or unmethylated C is very different; if methylated, cleavage is more likely at C, while if unmethylated, cleavage is less likely at C. If cleavage occurs at C, the terminal motif is CGN. If cleavage occurs one base earlier, the terminal motif is NCG.

[0177] To account for this difference, some embodiments can use terminal motifs to predict local methylation density. For example, regions with more CGNs have higher methylation density. Regions with more NCGs have lower methylation density.

[0178] For example, one or more calibration samples can have a measured methylation level and an amount of a CGN terminal motif and / or an amount of an NCG terminal motif. The amount of each terminal motif can be used separately, or can be used together, for example, as a ratio of the two amounts. When used separately, the amounts can be normalized, for example, using the normalization techniques described herein. The measured methylation level (known classification) and one or more amount parameters can comprise calibration data points, and the parameters can be considered calibration values. Such calibration is described in detail in other sections, for example, in the terminology section and Section III.E.

[0179] The cell-free DNA molecules in the new biological sample can be analyzed to determine the parameters of the amount of CGN terminal motifs and / or the amount of NCG terminal motifs. This parameter can then be compared with calibration data points from one or more calibration samples, which can be done using a calibration function (e.g., a regression function fitted to the calibration data points of a set of calibration samples). If the measured parameter is similar to the calibration value of the calibration data points, the new sample and the previous calibration sample have similar methylation levels. The calibration data points can be used to train a model, such as a linear model, a nonlinear model, or other similar or more complex machine learning model.

[0180] 2. Select the area Based on the data in Figure 28, we hypothesized that the number of CGN and NCG motifs present in a specific genomic region could be used to predict local methylation regions. For this purpose, as an example, we determined the methylation density and frequency of CGN and NCG motifs for plasma DNA molecules overlapping regions near the transcription start site (i.e., ±1000 bp of the TSS, referred to as the TSS region). The molecular overlap was determined by mapping sequence reads to these regions. Each TSS region overlapped with another TSS region. The average depth of each CpG was greater than 2. More than 100 DNA fragments were mapped to each region.

[0181] TSS regions with a methylation density greater than 70% were defined as hypermethylated regions (although other percentage ranges could also be used), and regions with a methylation density less than 30% were defined as hypomethylated regions. Regions in between were excluded but may be included in other embodiments, for example, to identify partially methylated regions. More than three classifications may also be used. CGN and NCG end motif frequencies were calculated for all hypomethylated and hypermethylated regions and normalized by the expected motif frequency from eight healthy control samples (median of 381 million paired end reads (IQR: 353-404 million)).

[0182] The regions can be of various sizes (eg, 100 kb, 200 kb, 500 kb, 1 Mb, 2 Mb, 3 Mb, and 5 Mb), with a 1 Mb region being merely an example.

[0183] B. Exemplary Classification Techniques One exemplary technique uses the amount of CGN or NCG, or the ratio of the two, possibly with normalization for either. Another exemplary technique uses the amount of each motif, for example, each trimer motif. A higher k value for the k-mer motif can be used.

[0184] 1. CGN or NCG In some embodiments, the amount of CGN and / or NCG can be used. This amount can be the relative frequency of the motif within a group of trimer terminal motifs of the analyzed cell-free DNA molecules, e.g., DNA molecules from the analyzed region (potentially the entire genome). For example, the sum of the individual percentages of the terminal motifs CGA, CGC, CGG, and CGT among all trimer terminal motifs of all DNA fragments analyzed can be determined. The same can be done for NCG. Each value can be used individually, or the values ​​of CGN and NCG can be used together to determine a parameter, e.g., a ratio. Such a percentage is a type of normalization. Another example of normalization is the ratio of the amounts of CGN and NCG, e.g., CGN / NCG, NCG / CGN, CGN / (CGN+NCG), etc. Such various ratios can be used in any embodiment disclosed herein that uses a ratio of CGN and NCG. In some embodiments, such normalized values ​​can also be increased or decreased, e.g., using expected frequencies.

[0185] In some embodiments, the expected frequency of a terminal motif can be determined based on the reference sequence in the region of the reference genome, for example, the number of times that a particular terminal motif appears in the region of the reference genome.The exact expected frequency depends on the sequence of the region and can be normalized to the size of the region, for example, can be defined as the total number of k-mer terminal motifs in the region.Because a certain region may have more CpG sites than other regions, the expected frequency can provide information about whether the measured frequency is higher than expected.

[0186] In this analysis, the expected frequency of a particular terminal motif can be determined in silico to adjust the weight of the observed frequency of the corresponding terminal motif. A 3-nt interval (i.e., a 3-nt sliding window) was used to scan the region of interest with a step size of 1 nt. In other words, the 3-nt sliding window was moved along the region of interest with a transition of 1 nt each time. Therefore, the number of trimer motifs could be generated based on the movement of the 3-nt sliding window. The frequency of these trimer motifs generated in silico was considered the expected frequency of the terminal motif associated with the region of interest. Therefore, the expected frequency is what would occur for random cleavage.

[0187] In one exemplary embodiment, the ratio of the observed frequency to the expected frequency of a particular terminal motif (O / E ratio) can be used for downstream analysis. In the O / E ratio, O is the observed frequency (i.e., normalized amount) of a particular set of one or more k-mer terminal motifs. The frequency can be determined via any normalization technique described herein. For example, the observed frequency can be determined as the percentage of fragments having one of a particular set of k-mer terminal motifs among all k-mer terminal motifs (e.g., trimeric terminal motifs).

[0188] Figure 29 shows the O / E ratios of CGN and NCG terminal motifs in hypermethylated and hypomethylated regions from eight healthy control samples, according to an embodiment of the present disclosure. As shown, the O / E ratio of CGN was substantially higher in hypermethylated regions (median = 2.07) than in hypomethylated regions (median = 0.99) (P-value: 0.00016, Mann-Whitney U test). In contrast, the O / E ratio of NCG was higher in hypomethylated regions (median = 0.71) than in hypermethylated regions (median = 0.47) (P-value: 0.00016, Mann-Whitney U test). These data suggest that the use of CGN and NCG motifs can signal hypomethylation or hypermethylation.

[0189] Such data can be used to train models, for example, to determine calibration values ​​(such as cutoffs / thresholds for binary classification, or values ​​used to determine narrower quantitative values, as described herein). For example, when using O / E ratios determined from CGN frequencies, values ​​above a cutoff of about 1.7 can indicate hypermethylation, and values ​​below about 1.1 can indicate hypomethylation, as can be determined from Figure 29. When using NCGs, values ​​above about 0.6 can indicate hypomethylation, and values ​​below about 0.6 can indicate hypermethylation.

[0190] If such training data points (calibration data points) had specific values ​​of methylation level (not just hypermethylated or hypomethylated state) in addition to the calibration level (i.e., O / E ratio), such training data points could be used to determine (e.g., using regression) a calibration function that could provide a methylation level for an input O / E ratio. Such uses of calibration data points are described throughout this disclosure. Those skilled in the art will understand that the applicability of such descriptions of one technique can also be applied to other techniques.

[0191] In other embodiments, a CGN / NCG ratio can be determined, and this ratio can be normalized by the expected ratio of expected frequencies. Although described as CGN / NCG, such ratio can have various forms, such as a CGN / NCG motif ratio, or an NCG / CGN motif ratio, or CGN / (CGN+NCG), or NCG / (CGN+NCG). Such ratios can also include ratios of functions of such values. Such variations can be used in any of the embodiments described herein.

[0192] 2. k-mer feature vector Some of the examples above used quantities of only some terminal motifs, for example, all CGN terminal motifs of the trimeric terminal motif (other values ​​of k in the k-mer can be used). This section describes using quantities of all k-mers. Such quantities can form a feature vector that is input into a machine learning model, which can, for example, determine a binary classification or provide a numerical value such as a specific percentage (or equivalently, a decimal value). Regression can provide such a numerical value.

[0193] Figure 30 illustrates the use of trimer end motifs of TSS regions to determine binary classification or regression, according to an embodiment of the present disclosure. The end motif is shown in the bottom left. As shown, trimer end motifs are used, although other values ​​of k, such as 1, 2, 4, etc., can be used. For trimer motifs, there are 64 values ​​in the feature vector. For tetramer motifs, there is a frequency vector of 256 values ​​that can be input into a machine learning model.

[0194] The target determination is the classification of the methylation density of the region. Two methods are listed. The first method is a binary classification that distinguishes hypermethylated regions from hypomethylated regions. High and low were defined as methylation densities less than 30% or more than 70%. Regions with methylation densities between these two numbers were excluded in this analysis. In addition, a regression model is used to quantitatively predict the exact methylation density, which can include regions with methylation densities between the thresholds used to define hypermethylated and hypomethylated regions.

[0195] 31A-31B show an exemplary training procedure according to an embodiment of the present disclosure. FIG. 31A shows training for a single subject, labeled as the one-sample strategy. Regions within one sample are trained and tested. Over 8,000 regions that met the reference were selected. The regions were divided into a randomly selected training set, and the remaining regions were used as the test set. There is no overlap between these two sets. The technique in FIG. 31A was used to generate the data in FIGS. 32A, 32B, 33A, 33B, and 36.

[0196] Figure 31B shows training in which multiple subjects are used, labeled as the multi-sample strategy. Seven healthy control cases were merged, resulting in over 43,000 regions. An individual with approximately 8,000 regions was used to test the model. The technique in Figure 31B is used to generate the data in Figures 34A, 34B, and 38.

[0197] SVM was implemented to predict whether a TSS region has a methylation density above 70% or below 30% by using the local 5' k-mer end motif frequency (e.g., the first three nucleotides at the 5' end of a cfDNA fragment and a total of 64 different trimer motifs).

[0198] Figures 32A-32B and 33A-33B show binary classification results for different k-mer motifs according to embodiments of the present disclosure. These results used a single-sample strategy. We observed that the prediction accuracy increased from the monomer motif to the tetramer motif. This increase was most pronounced from the dimer motif to the trimer motif. The AUC for the trimer terminal motif achieved approximately 0.97, demonstrating that regional methylation density can be predicted from terminal motif profiles.

[0199] Figure 34A shows the performance of local methylation status prediction using SVM based on the local terminal motif frequency or O / E ratio of terminal motifs according to an embodiment of the present disclosure. The local terminal motif frequency or O / E ratio of terminal motifs from TSS regions (i.e., ±1000 bp of the TSS) with methylation densities above 70% and below 30% from seven healthy control samples was used to train the model. The TSS region from one independent healthy control sample was used to perform ROC analysis for distinguishing the methylated and unmethylated states of the genomic region. The solid and dashed lines represent model performance based on the terminal motif frequency and O / E ratio of the terminal motif, respectively.

[0200] SVM models trained on the 5' trimer end motif frequency and the O / E ratio of the 5' trimer end motif achieved AUCs of 0.97 and 0.94, respectively. These data suggested that the use of 5' end motifs is informative for predicting methylation density in specific regions of interest. Note that the use of 5' end motifs is merely an example. 3' end motifs may be used in addition to or instead of any of the embodiments described herein.

[0201] In another embodiment, support vector regression (SVR) was implemented to predict the methylation density of TSS regions based on the local 5' trimer end motif frequency. SVR can perform multidimensional linear regression. The exact number of methylation densities was used to train a model and predict the percentage of methylation density.

[0202] Figure 34B shows the performance of local methylation density estimation using SVR based on local end motif frequency according to an embodiment of the present disclosure. A model was trained to predict the methylation density of the TSS region from one independent healthy control sample using local end motif frequencies from the TSS region (i.e., ±1000 bp of the TSS) from seven healthy control samples. The scatter plot shows the correlation between the predicted methylation density and the observed methylation density detected by bisulfite sequencing of the test sample.

[0203] The predicted methylation density correlated well with that inferred by bisulfite sequencing (Pearson r: 0.93). Therefore, the data suggested that trimer terminal motif profiles enable the inference of the methylation density of genomic regions of interest. In another embodiment, local terminal motif frequencies can be used to predict the methylation density of other regions, such as CpG islands (CGIs), long interspersed nucleotide sequences (LINEs), Alu repeats, microsatellites, tandem repeats, enhancers, transposons, and transcription factor binding sites. In another embodiment, terminal motifs can involve, but are not limited to, 1 nt, 2 nt, 3 nt, 4 nt, 5 nt, 6 nt, 7 nt, 8 nt, 9 nt, and 10 nt or more at each end of cell-free DNA molecules. In yet another embodiment, other machine learning models, such as decision trees, Bayesian classifiers, support vector regression, recurrent neural networks, and other neural network-based models, can be implemented, as will be appreciated by those skilled in the art.

[0204] In some embodiments, only certain terminal motifs are used, for example, those corresponding to CGN or NCG, or those that include CG (i.e., including CGN and NCG, as well as larger motifs depending on the value of k).

[0205] C. Normalization The sequence context itself may play a role in predicting methylation status. Various types of normalization can be performed using the sequence context of a region. Exemplary types of normalization can use expected frequencies per region (described above) or per fragment normalization.

[0206] 1. Normalized motif frequency by region 35 illustrates normalization using the expected frequency of regions of each motif (local normalization) according to an embodiment of the present disclosure. The normalized motif frequencies of different motifs can then be used to generate feature vectors that are input into an ML model.

[0207] For example, as described above, a sliding window can be used to calculate the expected frequency of a given motif in a region. The expected frequency can provide a measure of what frequency would be expected if cleavage were random. As shown, the observed frequency is divided by the expected frequency of a given region, and the normalized value of CC is 2, indicating a higher-than-expected frequency. The normalized frequency of TT is lower than expected. This normalization can eliminate the influence caused by sequence context. In this way, only the influence of cleavage preference is avoided.

[0208] Figure 36 shows the results of binary classification using an SVM model with context-normalized terminal motifs (local normalization) according to an embodiment of the present disclosure. The results are worse than without normalization, as shown in Figures 32A-33B. Decreased prediction accuracy is observed for trimer and tetramer motifs. This decrease is due to the fact that the sequence context carries some information, and normalization effectively loses some of this information. Another reason may be that for tetramer motifs, there is a diversity of 256 different types. However, most regions have only 100-200 types mapped. Therefore, the features of tetramer motifs are relatively sparse. Larger regions may solve this problem, since more motifs are found.

[0209] 2. Normalized motif frequency per molecule Molecule-wise normalization was also investigated.

[0210] 37 illustrates per-molecule contextual normalization using expected motif frequencies in molecules, according to an embodiment of the present disclosure. Although the fragments are plotted against a reference genome, normalization is performed using the sequence of each molecule. The sequence of each molecule can be determined using the reference, for example, if the entire fragment is not sequenced and the two ends of the fragment are mapped to the reference. The reference sequence can be used to fill in the portions of the molecule that have not been sequenced.

[0211] The exemplary fragment shown has a CGT terminal motif. Because the observed motif occurs, it can be considered 1. The expected motif frequency can be determined, for example, using a 3-base sliding window across the fragment, as described herein. For example, if the fragment is 150 bp, there are 148 possibilities for cleavage. If only one of those possibilities is CGT, the expected frequency is 1 / 148, thereby providing a normalized motif frequency of 148. However, if there were five instances of CGT, the expected frequency would be 5 / 148, resulting in a normalized frequency of 29.6. In this way, the expected frequency can act as a weight for each fragment, determining how much the fragment can contribute to the normalized motif frequency.

[0212] 38 shows trimer motif results for no normalization, local normalization, and molecule-by-molecule normalization, according to an embodiment of the present disclosure. Results using raw trimer motif frequencies performed best.

[0213] D. Which terminus provides the greatest differentiation? We investigated which terminus provides the best discrimination of the predictive models.

[0214] Figure 39 shows the normalized frequency difference (using local normalization) of different dimer end motifs according to an embodiment of the present disclosure. The Y-axis shows the normalized motif frequency. If the cleavage was random, one would expect the motif frequency to be approximately 1 after normalization. Two outliers are observed, with CC and CG. There are many CC ends, whether hypomethylated or hypermethylated. In hypermethylated regions, there is a further increase in CC ends.

[0215] For CG, there is no preference for cleavage in hypomethylated regions because it is near the dashed line at 1. However, in hypermethylated regions, cleavage at this CG is preferred. This is consistent with other results herein. Also, the CG terminal motif shows the greatest difference between hypomethylated and hypermethylated regions, which is consistent with the results above.

[0216] Given the difference between CC and CG normalized motif ratios (frequencies), such ratios can be used to distinguish (classify) hypomethylated and hypermethylated regions. Such classification using calibration values ​​can be performed, for example, as described above for Figure 29. Using local normalization, CG terminal motifs provide the best separation.

[0217] Figure 40 shows the difference in normalized frequency (local normalization) of different trimer end motifs according to an embodiment of the present disclosure. The CGN end motif shows the largest difference, which is consistent with the results above. Again, we examine motifs starting from CC and CG regions. We can see that the motifs CA and CT do not show very significant differences in hypomethylated and hypermethylated regions, while the CC and CG motifs show some differences, with the CG motif showing more differences than the CC motif. Furthermore, the differences coincide with hypermethylated regions having a greater amount of CN motifs.

[0218] E. Methods for Determining Methylation in a Region Methods are provided for determining the methylation of regions in a genome using cleavage information. Methylation can be a property of a sample or a property of a particular region in a genome, and the location of the region can be defined relative to a reference genome. The cleavage information can be of various types as described herein, such as sequence end motifs.

[0219] 41 is a flowchart illustrating a method 4100 for determining methylation in a region according to an embodiment of the present disclosure. As with other methods described herein, method 4100 can be implemented in part or in whole using a computer system. Blocks performing similar functions to other methods can be implemented similarly to what is described for those blocks in other flowcharts. A region may include discrete subregions (i.e., not contiguous), which may be on the same chromosome or on different chromosomes.

[0220] In block 4110, a plurality of cell-free DNA molecules from the subject's biological sample are analyzed. Block 4110 can be performed similarly to block 2710, or vice versa. Similar to block 2710, analyzing the cell-free DNA molecules can include determining the location of the cell-free DNA molecules in a reference genome, for example, when the region is smaller than the entire genome. The plurality of cell-free DNA molecules can be located in a region of the reference genome.

[0221] The location can be determined by various methods described herein, such as aligning sequence reads to a reference genome or using location-based probes. Location-based probes are those whose sequences are primarily or entirely based on sequences present in a reference genome. Thus, the binding of location-based probes can be used to indicate the genomic location of a bound piece of DNA. The bound piece of DNA can be a cell-free DNA molecule found in bodily fluids, such as plasma, serum, saliva, urine, peritoneal fluid, cerebrospinal fluid, sweat, semen, vaginal secretions, cervical mucus, etc.

[0222] The region can be the entire reference genome or a portion of the subject's genome. The portion of the genome can be 10 Mb or less of a particular chromosome, and other sizes of regions are described herein. Thus, multiple cell-free DNA molecules can be derived from one or more specific regions of the reference genome. At least one of the one or more specific regions can correspond to a particular genotype or haplotype containing one or more alleles. In this manner, the genotype, haplotype, and methylation of a particular chromosome copy can be determined. If a subject has a certain methylation level in a genotype or haplotype, such a characteristic can indicate a pathology (e.g., via genomic imprinting, as described in more detail elsewhere in this disclosure). For example, a certain methylation level (e.g., a certain methylation classification) can silence genes associated with pathologies such as those resulting from, or related to, disease (e.g., cancer, inflammation) or abnormal conditions / disorders in physiological processes (e.g., during cell differentiation, development, or aging). Copy number abnormalities can be such pathologies. Embodiments may also determine the normal level of a physiological process, for example, an individual's stage of pregnancy or aging, or monitor the type or level of an individual's genomic imprinting.

[0223] The analysis can include determining a terminal sequence motif (also referred to as a terminal motif) at at least one end of the cell-free DNA molecule. The terminal sequence motif can be determined from the outermost base in the sequence read. If the sequence read includes only one end of a DNA fragment / molecule (e.g., as one of paired reads covering both ends) rather than the entire DNA fragment / molecule, the base corresponding to the end of the DNA fragment is used, and the base toward the center of the fragment is not used. The end of the cell-free DNA molecule can have a first position at the outermost position, a second position adjacent to the first position, and a third position adjacent to the second position.

[0224] In block 4120, a first amount of a first set of one or more terminal sequence motifs having a C at a first position and a G at a second position (e.g., the amount of CGN or any one of these motifs) is determined. The first set of one or more terminal sequence motifs can include one or more of CGA, CGC, CGG, and CGT, with CGN representing all. In other embodiments, the first amount can be of a first set of one or more terminal sequence motifs having a C at a second position and a G at a third position (e.g., NCG). In this example, the first set of one or more terminal sequence motifs can include one or more of ACG, CCG, GCG, and TCG, with NCG representing all. Such exemplary terminal motifs can be used in various methods described herein to determine amounts and / or normalized parameters (e.g., frequencies) for use in performing classification. Additional normalization based on expected frequencies per region or per fragment can be used.

[0225] Block 4120 can be performed similarly to block 2720. For example, the first amount can be determined using sequencing or probe-based technology, such as PCR. The amount of sequence reads having a particular terminal motif can be counted. As another example (e.g., as depicted in FIG. 4), a probe-based technology can be used. For example, the number of positive reactions for a particular probe can be used. As another example, an intensity signal (e.g., electrical or optical) from a probe specific to a particular terminal motif can be used to determine the first amount.

[0226] In block 4130, the methylation classification of the region in the subject's genome is determined by comparing the first amount to a calibration value. The calibration value can be determined using cell-free DNA molecules derived from one or more calibration samples and located in regions (each containing one or more CpG sites) with known classifications. The known classifications can be determined using a methylation-recognition assay or from the results of other assays using a methylation-recognition assay.

[0227] Other amounts can be determined, for example, for normalization. For example, when the first amount is determined using one or more CGN motifs, a second amount of a second set of one or more terminal sequence motifs (e.g., NCG) having a C at the second position and a G at the third position can be determined. Then, determining the classification can include normalizing the first amount with the second amount to obtain a normalized first amount that is compared with the calibration value. The second set of one or more terminal sequence motifs can include all or one or more of ACG, CCG, GCG, and TCG.

[0228] Another example of quantity is, as described above, expected quantity.Based on the reference sequence of reference genome in the region, can determine the expected quantity of the first set of one or more end sequence motifs.Then, determining classification can comprise: normalizing the first quantity with expected quantity to obtain the normalized first quantity (for example, O / E ratio) that is compared with calibration value.

[0229] As another example of normalization, the total amount of a sequence motif in a plurality of cell-free DNA molecules (e.g., in a region) can be determined. Determining the classification can then include normalizing the first amount by the total amount to obtain a normalized first amount (e.g., motif ratio or relative frequency) that is compared to a calibration value.

[0230] As with other techniques described herein, a feature vector of multiple amounts can be provided to a machine learning model, for example, as described in the sections above. The different amounts can correspond to the selection of terminal motifs, e.g., those included in the first set and additional terminal motifs. Such terminal motifs included in the feature vector can be all of a particular k or only a subset of k-mers. Thus, the respective amounts of each trimeric terminal sequence motif having a C at the first position and a G at the second position (i.e., four amounts of CGN) can be determined. A feature vector including each amount, including the first amount, can be generated and then input into a machine learning model as part of determining the methylation classification of a region in a subject's genome. The machine learning model can be trained using cell-free DNA molecules from one or more calibration samples. Additionally or alternatively, the feature vector can include the respective amounts of each trimeric terminal sequence motif having a C at the second position and a G at the third position (i.e., NCG). As described above, the feature vector can include the respective amounts of all trimeric terminal sequence motifs.

[0231] V. Predicting methylation of sites in specific fragments Some embodiments can use terminal motif patterns associated with a single plasma DNA molecule, combined with other features such as fragment size and sequence context, to predict its methylation state. The methylation state can be fully methylated, fully unmethylated, or partially methylated. Various types of data besides terminal motifs can be used to determine the methylation levels of CpG sites on individual fragments.

[0232] A. Input Figure 42 shows an exemplary workflow of a machine learning model for predicting methylation status at the single molecule level, according to an embodiment of the present disclosure. A reference genome 4210 is shown along with aligned fragments 4205. For each fragment, three types of information are shown: terminal motifs, genomic (sequence) content, and size. Such data can be formed into various data structures to create a feature vector. An example of a feature vector is shown in Figure 42.

[0233] On the left side, the terminal motif is converted into a matrix with one-hot encoding for each position. For example, the terminal motif is CGCA. Because the first base is C, the first row has 1 for C and 0 for the other bases. Because the second base is G, the second row for G has 1. The third base is C, and the fourth base is A. Thus, we coded the terminal motif from the 5' end as a matrix based on the one-hot encoding rules. A four-dimensional vector was used to encode the nucleotides. For example, [1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], and [0, 0, 0, 1] represent A, C, G, and T, respectively. Thus, for the "CGCA" terminal motif, the matrix is ​​represented as shown in Figure 42 (represented by feature I). Terminal motifs can be represented in other ways, for example, using values ​​1 to 4 assigned to each base type, e.g., A=1, C=2, G=3, and G=4.

[0234] The center shows the use of size. In this example, the size of the fragment is determined based on paired-end sequencing and alignment of the sequence read to the reference genome. If the entire molecule is sequenced, alignment is not required. Because most cell-free DNA fragments are less than 600 bp, normalization can be performed, for example, by dividing by 600. Therefore, the fragment size can be determined by the outermost genomic coordinate of the aligned paired-end read and normalized by a factor of 600 bp (represented by feature II).

[0235] On the right, the genomic context is a matrix containing dinucleotide patterns. The first base of a dimer is in the row, and the second base is in the column. For each specific dimer, the matrix element stores a count of the number of instances of that dimer. Thus, sequence context can be encoded into a matrix containing counts of dinucleotide types along a single plasma DNA molecule. A 2-nt sliding window was used to scan a single plasma DNA molecule to determine the dinucleotide frequency of a single plasma DNA molecule. A two-dimensional (2-D) matrix was used to represent the dinucleotide frequency (represented by feature III). For example, if the frequency of the "AT" dinucleotide is 2, a "2" is entered in the cells of the raw "A" and "T" columns. Other k-mer arrays (vectors, matrices, or higher-order tensors) can be used. For trinucleotide patterns, there is a third dimension to the array. The genomic (sequence) context can also be the entire sequence. Another example of genomic context is GC percentage.

[0236] Two or more of these features can be combined (or only the terminal motifs can be used alone) and input into a machine learning model to predict whether the molecule is fully methylated, fully unmethylated, or partially methylated. A composite feature matrix containing the aforementioned features (I, II, and III) was used to train a CNN model to determine whether a DNA fragment is fully methylated or fully unmethylated.

[0237] B. Results Figure 43A shows the performance of single-molecule methylation status analysis using CNN based on terminal motifs, sequence context, and fragment size, according to an embodiment of the present disclosure. A model for distinguishing between methylated and unmethylated molecules can be obtained with an AUC of 0.85. The results suggest that the use of fragmentomics features with other properties derived from a single plasma DNA molecule potentially enables the determination of the methylation status of that plasma DNA molecule.

[0238] Figure 43B shows the performance of single-molecule methylation status analysis using molecules with CGN and NCG end motifs based on a CNN model, according to an embodiment of the present disclosure. Therefore, accuracy was determined for a subset of DNA fragments. Interestingly, focusing on those molecules beginning with CGN or NCG end motifs significantly improved model performance (i.e., AUC of 0.85-0.94, P value <0.0001, DeLong test). Therefore, selective analysis of a subset of DNA molecules with certain motifs allows for enhanced classification power in inferring the methylation status of single molecules.

[0239] The CGN and NCG motifs can act like tags in DNA fragments, allowing for more precise differentiation between fully methylated and fully unmethylated DNA fragments at the individual molecule level. Thus, on a fragment-by-fragment basis, embodiments can determine how methylated a particular fragment is.

[0240] C. Methods for predicting methylation of fragment sites Also provided are methods for determining the methylation status at a single fragment site using cleavage information, as described above. The cleavage information can be of various types as described herein, such as sequence end motifs.

[0241] 44 is a flowchart illustrating a method 4400 for predicting the methylation of DNA fragments, according to an embodiment of the present disclosure. As with other methods described herein, method 4400 can be implemented in part or in whole using a computer system. Blocks performing similar functions to other methods can be implemented similarly to what is described for those blocks in other flowcharts.

[0242] At block 4410, one or more sequence reads for a test cell-free DNA molecule can be received. The test cell-free DNA molecule can be generated in a variety of ways, for example, using the techniques described in blocks 2710 and 4110. The test cell-free DNA molecule can include one or more CpG sites.

[0243] In block 4420, the sequence context of the test cell-free DNA molecule is determined using one or more sequence reads. Examples of sequence context include consecutively occurring base combinations (e.g., various k-mers, also referred to as K-mers). For example, the sequence context can include a K-mer matrix that specifies the number of instances of each K-mer in the test cell-free DNA molecule, where K is an integer. When K is equal to 2, the K-mer matrix is ​​a dinucleotide matrix that specifies the number of instances of dinucleotide pairs in the test cell-free DNA molecule. As another example, the sequence context can include sequence information from each nucleotide (i.e., the entire sequence) of the test cell-free DNA molecule.

[0244] In block 4430, a first end sequence motif at a first end of the test cell-free DNA molecule is determined using one or more sequence reads. The number of bases in the first end sequence motif can vary. For example, different k-mers can be used. If one sequence read is present for the entire molecule, the outermost k bases can be selected from either end of the sequence read. If two sequence reads are obtained from both ends, the bases corresponding to the ends of the DNA molecule can be identified based on the positional information obtained during the sequencing process. If probes are used, the signal corresponding to a particular probe provides the order from the outermost toward the center of the DNA molecule. A second end sequence motif at a second end of the test cell-free DNA molecule can also be determined.

[0245] At block 4440, method 4400 includes loading into the memory of the computer system a machine learning model trained using a training set of training cell-free DNA molecules having CpG sites with known methylation status. The training set can include sequence context and terminal sequence motifs of the training cell-free DNA molecules. By way of example, the methylation level of the training cell-free DNA molecules can be determined using methylation-aware sequencing or by methylating or unmethylating all CpG sites. For example, whole genome amplification can unmethylate newly synthesized DNA, and CpG methyltransferase (M.SssI) can methylate input DNA samples. Training can involve a variety of techniques and models, including ensemble models. For example, different subsets of training samples can be used, and different subsets can be used for validation.

[0246] In block 4450, the machine learning model can determine the methylation level of one or more CpG sites of the test cell-free DNA molecule. The sequence context and first terminal sequence motif of the test cell-free DNA molecule (e.g., as part of a feature vector) are input into the machine learning model. Such a feature vector can use other features. For example, a second terminal sequence motif at the other end can also be input into the machine learning model. Another exemplary feature is the size of the test cell-free DNA molecule.

[0247] If the test cell-free DNA molecule contains multiple CpG sites, the methylation level can be the percentage of the multiple CpG sites that are methylated. As another example, the methylation level can be whether a particular CpG site is methylated.

[0248] Various formats of the feature vector can be used. For example, the first terminal sequence motif can be one-hot encoded for input to the machine learning model. The format of the sequence context (or genome context) can be a two-dimensional or higher-dimensional matrix, for example, based on the value of K used for the K-mer.

[0249] As described above, size can be input into a machine learning model as part of determining the methylation level of one or more CpG sites of a test cell-free DNA molecule. Size can have various formats, such as the number of bases, or a ratio to a reference size, such as 600 bp. Size can be determined in various ways. For example, if one sequence read spans the test cell-free DNA molecule (e.g., as a result of single-molecule sequencing), the size of the test cell-free DNA molecule can be determined by counting the number of bases in the one sequence read. As another example, if two sequence reads (paired ends) corresponding to both ends of the test cell-free DNA molecule are obtained, the size can be determined by aligning the two sequence reads to a reference genome.

[0250] Various types of machine learning models can be used, such as convolutional neural networks, recurrent neural networks, transformer models (encoder-decoder models), random forest classification, and others described elsewhere in this disclosure. A machine learning model can also be a combination of models, thereby an ensemble model.

[0251] VI. Detection of 5hmc-enriched or -depleted regions Most of the above discussion has discussed methylation generally, rather than any specific type of methylation, because methylation is generally applicable. Techniques can be used to detect certain types of methylation. For example, hMe-Seal can be used to detect 5hmC methylation. Differential fragmentation can be identified in 5hmC-enriched regions (where a higher proportion of methylation comes from 5hmC than in other regions) and 5hmC-depleted regions (where a lower proportion of methylation comes from 5hmC than in enriched regions). In this way, some embodiments can detect certain regions that have a higher or lower proportion of a certain type of methylation than other regions of the genome. Based on the fragmentation, some embodiments can determine whether a region is 5hmC-enriched or depleted.

[0252] A. Relationship between terminal motifs and 5hmc-enriched and -depleted regions 5-Hydroxymethylcytosine (5hmC) is an important mammalian DNA epigenetic modification implicated in gene regulation and cancer pathogenesis. In bisulfite sequencing, both 5hmC and 5hmC are read as cytosine and therefore cannot be distinguished. We reasoned that the presence of 5hmC may alter the frequency of CG-containing motifs. To test this hypothesis, we obtained selective chemical labeling-based 5hmC sequencing (hMe-Seal data) from 10 plasma DNA samples from healthy subjects (Song et al. Cell Res 2017;27(10):1231-1242).

[0253] hMe-Seal is a method that selectively labels 5hmC with biotin via azide-modified glucose using β-glucosyltransferase (βGT) to enrich 5hmC-containing DNA fragments for sequencing. Such biotin can be captured by several beads, capturing 5hmC fragments but not 5hmC. Higher sequencing depths around a CpG site represent greater amounts of 5hmC present in the sample. Regions of higher sequencing depth (higher coverage) that form "peak" signals across the genome were defined as "5hmC-enriched regions." Using peak-calling methods (e.g., MACS2), we identified 38,168 common "peak" regions across 10 healthy subjects. We analyzed 4,631,823 CpG sites, which were determined (in-house) to be commonly methylated across eight healthy samples by bisulfite sequencing. Among them, 25,102 CpGs overlapped with the 5hmC-enriched peak regions (referred to as group I, 5hmC-enriched regions) and 3,409,329 CpGs (referred to as group II, 5hmC-depleted regions) that were not present in any of the peak regions of the 10 healthy subjects.

[0254] We reasoned that those CpG sites in group I have relatively more 5hmC modifications, while those in group II have fewer 5hmC modifications. The frequencies of CG-containing terminal motifs were obtained for these selected CpGs in the same eight healthy plasma samples.

[0255] To compare motif frequencies between Group I and Group II, the expected terminal motif frequencies of CG-containing motifs within these two groups were calculated from the reference genome, which was used to normalize the observed terminal motif frequencies. The normalized values ​​were calculated using the observed / expected ratio (O / E ratio). The expected frequencies can be determined, for example, by using a sliding window of size k (corresponding to the k-mer terminal motif being used) to determine the frequency of each terminal motif being analyzed, as described herein. In other embodiments, only the observed frequencies can be used. The frequencies of certain terminal motifs between the two types of regions (5hmC-enriched and 5hmC-depleted) were then compared.

[0256] Figures 45A-45B are boxplots showing the O / E ratios of CGN (Figure 45A) and NCG (Figure 45B) motifs between 5hmC-enriched and 5hmC-depleted regions. As shown in Figures 45A-45B, compared with cfDNA molecules associated with CpG4510 (depleted) in group II, CpG4520 (enriched) in group I showed significantly higher O / E ratios in CGA, ACG, and TCG motifs. For CGC, CGG, CCG, and GCG motifs, significantly lower O / E ratios were observed in 5hmC-enriched region 4520 (i.e., CpG sites in group I). These results indicated that the presence of 5hmC affects plasma DNA cleavage patterns.

[0257] We have shown above that the O / E ratio of CG-containing motifs differs between regions containing 5hmC and regions without 5hmC. The presence of 5hmC can alter the frequency and O / E ratio of CG-containing motifs. Abnormalities in CG-containing motifs can be used to predict the presence of 5hmC in a genomic region.

[0258] To demonstrate the predictive power, we used the common "peak" regions identified above (n = 38,168). We then standardized the length of the peak regions to 200 bp (i.e., peak center + / - 100 bp), and further selected peak regions containing at least one CpG site for each NCG context (i.e., ACG, CCG, GCG, TCG) (n = 24,221). These regions were designated as 5hmC-containing regions (also designated as 5hmC-enriched) in healthy controls.

[0259] To define regions lacking 5hmC (also referred to as 5hmC depletion) in healthy controls, we divided the reference genome into 200-bp bins and identified bins that did not overlap with 5hmC-containing regions and contained at least one CpG site for each NCG context (i.e., ACG, CCG, GCG, TCG) (n=992,991). As an example, we pooled WGBS reads of plasma DNA from eight healthy control samples and performed classification analysis of regions containing or not containing 5hmC. For this purpose, approximately 2,000 regions containing 5hmC and 2,000 regions not containing 5hmC, which were covered by more than 20 CG-containing ends, were used for downstream analysis.

[0260] An SVM model was trained using the O / E ratio of CG-containing motifs (i.e., ACG, TCG, CCG, GCG, CGA, CGT, CGC, and CGG) on a dataset containing 1,000 regions containing 5hmC and 1,000 regions not containing 5hmC, respectively. The trained model was used to analyze another independent dataset containing 1,000 regions containing 5hmC and 1,000 regions not containing 5hmC, respectively.

[0261] Figures 46A-46B show the performance of 5hmC prediction based on the 5' trimer CG-containing terminal motif. Figure 46A shows a boxplot showing the probability of 5hmC presence between regions containing 5hmC and regions not containing 5hmC. Figure 46B shows the ROC analysis for distinguishing regions containing 5hmC from regions not containing 5hmC. As shown, regions containing 5hmC showed significantly higher prediction probability than regions not containing 5hmC (P<0.001). The ROC analysis showed that the SVM model could reach an AUC of approximately 0.90 for distinguishing regions containing 5hmC from regions not containing 5hmC, i.e., for distinguishing regions that are 5hmC-enriched from regions that are 5hmC-depleted.

[0262] In addition to the O / E ratio, other techniques described herein can be used, either individually or using feature vectors of CG-containing motifs (e.g., the SVM model shown above). For example, the CGN / NCG ratio can be used to select terminal motifs, such as those with an inverse relationship. Terminal motif frequency can also potentially be used as a feature vector of CG-containing motifs (e.g., 1 to 8 values ​​for different trimer terminal motifs). Feature vectors can be used in machine learning models, as described herein. Such quantities can be used directly without normalization of expected quantities. Other normalizations can be used, for example, using the total quantity of a particular group of terminal motifs (e.g., CGN, NCG, or both) or all terminal motifs of a desired length (e.g., trimer or tetramer).

[0263] B. Methods using terminal motifs to detect 5hmc enriched and depleted regions Figure 47 is a flowchart illustrating a method for measuring 5hmC methylation of regions in a subject's genome using cell-free DNA molecules. As with other methods described herein, method 4700 can be implemented in part or in whole using a computer system. Blocks that perform similar functions to other methods can be implemented similarly to what is described for those blocks in other flowcharts.

[0264] In block 4710, a plurality of cell-free DNA molecules from the subject's biological sample are analyzed. Block 4710 can be performed similarly to blocks 2710 and 4110. Similar to blocks 2710 and 4110, analyzing the cell-free DNA molecules can include determining the location of the cell-free DNA molecules in a reference genome, for example, when the region is smaller than the entire genome. The plurality of cell-free DNA molecules can be located in a region of the reference genome.

[0265] The analysis may include determining a terminal sequence motif (also referred to as a terminal motif) of at least one end of the cell-free DNA molecule, which may be performed similarly to block 4110. The terminal sequence motif may be determined from the outermost bases in the sequence read. If the sequence read includes only one end rather than the entire DNA fragment / molecule (e.g., as one of paired reads covering both ends), the bases corresponding to the end of the DNA fragment are used, and bases toward the center of the fragment are not used. The ends of the cell-free DNA molecule may have a first position at the outermost position, a second position adjacent to the first position, and a third position adjacent to the second position.

[0266] In block 4720, a first amount of a first set of one or more terminal sequence motifs having a C at a first position and a G at a second position (e.g., the amount of CGN or any one of these motifs) is determined. The first set of one or more terminal sequence motifs can include one or more of CGA, CGC, CGG, and CGT, with CGN representing all. In other embodiments, the first amount can be of a first set of one or more terminal sequence motifs having a C at a second position and a G at a third position (e.g., NCG). In this example, the first set of one or more terminal sequence motifs can include one or more of ACG, CCG, GCG, and TCG, with NCG representing all. Such exemplary terminal motifs can be used in various methods described herein to determine amounts and / or normalized parameters (e.g., frequencies) for use in performing classification. Additional normalization based on expected frequencies per region or per fragment can be used. Block 4720 can be performed similarly to blocks 2720 and 4120.

[0267] In block 4730, the 5hmC methylation classification of the region in the subject's genome is determined by comparing the first amount with a calibration value. The calibration value can be determined using cell-free DNA molecules derived from one or more calibration samples and located at CpG sites with known classifications of 5hmC methylation levels. The known classification can be determined using a 5hmC recognition assay (e.g., as described above) or from the results of other assays using a 5hmC recognition assay.

[0268] Other amounts can be determined, for example, for normalization. For example, when the first amount is determined using one or more CGN motifs, a second amount of a second set of one or more terminal sequence motifs (e.g., NCG) having a C at the second position and a G at the third position can be determined. Then, determining the classification can include normalizing the first amount with the second amount to obtain a normalized first amount that is compared with the calibration value. The second set of one or more terminal sequence motifs can include all or one or more of ACG, CCG, GCG, and TCG.

[0269] Another example of quantity is, as described above, expected quantity.Based on the reference sequence of reference genome in the region, can determine the expected quantity of the first set of one or more end sequence motifs.Then, determining classification can comprise: normalizing the first quantity with expected quantity to obtain the normalized first quantity (for example, O / E ratio) that is compared with calibration value.

[0270] As another example of normalization, the total amount of a sequence motif in a plurality of cell-free DNA molecules (e.g., in a region) can be determined. Determining the classification can then include normalizing the first amount by the total amount to obtain a normalized first amount (e.g., motif ratio or relative frequency) that is compared to a calibration value.

[0271] As with other techniques described herein, a feature vector of multiple amounts can be provided to a machine learning model, for example, as described in the sections above. The different amounts can correspond to the selection of terminal motifs, e.g., those included in the first set and additional terminal motifs. Such terminal motifs included in the feature vector can be all of a particular k or only a subset of k-mers. Thus, the respective amounts of each trimeric terminal sequence motif having a C at the first position and a G at the second position (i.e., four amounts of CGN) can be determined. A feature vector including each amount, including the first amount, can be generated and then input into a machine learning model as part of determining the methylation classification of a region in a subject's genome. The machine learning model can be trained using cell-free DNA molecules from one or more calibration samples. Additionally or alternatively, the feature vector can include the respective amounts of each trimeric terminal sequence motif having a C at the second position and a G at the third position (i.e., NCG). As described above, the feature vector can include the respective amounts of all trimeric terminal sequence motifs.

[0272] VII. Clinically Relevant Fractional Concentrations of DNA Fragmentation in different types of tissues was also investigated. The terminal motif groups CGN and NCG, as well as cleavage profiles, were analyzed in regions / sites with tissue-specific methylation patterns. Clinically relevant DNA can be from any tissue type in which differentially methylated regions, such as hypermethylated, hypomethylated, 5hmc-enriched, or 5hmc-depleted regions, can be identified for the specific tissue from which the clinically relevant DNA is derived. Exemplary tissue types include fetal, tumor, or transplant tissue, or tissue from a specific organ, such as liver tissue. For example, tissue-specific alleles can be used to detect tissue-specific DNA to measure calibration values ​​for use in a calibration function, which can be used to provide fractional concentrations using fragmentomics parameters. DNA fragments from specific regions / sites can be obtained through targeted techniques (also called enrichment), which can be performed by amplifying certain regions or capturing DNA from certain regions. Further details on enrichment can be found in other sections of this disclosure.

[0273] In this disclosure, paired-end bisulfite sequencing (75 bp x 2 (i.e., paired-end sequencing), Illumina) was performed. Genotypes for maternal buffy coat and placenta / chorionic villus tissue samples were obtained using microarray-based genotyping technology (HumanOmni2.5 Genotyping Array Illumina) to identify informative SNPs (i.e., the mother was homozygous (represented as AA genotype) and the fetus was heterozygous (represented as AB genotype)). Fetal-specific DNA fragments were identified according to DNA fragments carrying fetal-specific alleles at the informative SNP site. In this scenario, the B allele was fetal-specific, and it was inferred that DNA fragments carrying the B allele originated from fetal tissue. The number of fetal-specific molecules (p) carrying the fetal-specific allele (B) was determined. The number of molecules (q) carrying the shared allele (A) was determined. The fetal DNA fraction across all cell-free DNA samples was calculated by 2p / (p+q)*100%. The amount of fetal-specific alleles varied between samples (median: 198,248, range: 147,614–207,735).

[0274] A. Tissue-specific methylation of CGN and NCG Because the number of DNA molecules with CGN and NCG ends was found to correlate with the methylation status of CpGs at or near the ends of cell-free DNA molecules, we further explored the possibility of using the relative amount of plasma DNA molecules terminating in CGN and / or NCG motifs to reflect methylation patterns. Relative amounts can be measured in various ways. For example, relative amounts can be measured as ratios (e.g., CGN / NCG motif ratio, or NCG / CGN motif ratio, or CGN / (CGN+NCG), or NCG / (CGN+NCG)). Comparison of CGN or NCG with other terminal motifs can also be used.

[0275] As another example, the percentage of cell-free DNA molecules terminating in CGN relative to the total number of cell-free DNA molecules or cell-free DNA ends analyzed (e.g., sequenced) can be measured. Similarly, the percentage of cell-free DNA molecules terminating in NCG relative to the total number of cell-free DNA molecules or cell-free DNA ends analyzed (e.g., sequenced) can also be measured. Furthermore, such normalization, or any other normalization described herein, can be performed by analyzing the same number of DNA molecules, generating the same number of reads, or using samples of the same size (e.g., a specified number of ml). Thus, if the same number of DNA molecules are sequenced from one run to another, that amount can be used directly, since normalization is effectively performed. Such various options for normalization can be used with any of the methods described herein.

[0276] The results below show that the relative amounts of plasma DNA molecules terminating in CGN or NCG motifs in genomic regions that exhibit differential DNA methylation patterns across different tissue types can allow for the determination of the probability of tissue of origin of the plasma DNA molecules and the determination of the fractional concentration of specific types of cell-free DNA molecules.

[0277] 1. Motif frequency of fetal-specific alleles Maternal plasma samples were analyzed for fetal hypermethylated and hypomethylated regions. Total trimer motifs were calculated for selected regions, and motif frequency was determined as the amount of one motif (e.g., CGN or NCG) per total amount of trimer motifs in the region.

[0278] 48A and 48B show motif frequencies of fetal-specific hypermethylated and hypomethylated regions, according to embodiments of the present disclosure. CGN and NCG terminal motif frequencies differ significantly between DNA molecules carrying fetal-specific and shared alleles depending on the methylation status.

[0279] We analyzed fetal-specific and shared cfDNA from the third trimester (sequencing depth: 270X). For the fetal-specific hypermethylated region (chr3: 187,767,589–187,768,381, containing nine CpG sites), the frequency of CGN terminal motifs in fetal-specific DNA molecules was 5.2%, higher than that of DNA molecules carrying the allele shared by the fetus and mother (1.3%). In contrast, the frequency of NCG terminal motifs in fetal-specific DNA molecules was 1.7%, lower than that of DNA molecules carrying the allele shared by the fetus and mother (3.2%). For the fetal-specific hypomethylated region (chr7: 25,889,274–25,890,315, containing nine CpG sites), we observed a different pattern from that of the hypermethylated region. Fetal-specific DNA molecules showed a lower frequency of the CGN end motif but a higher frequency of the NCG end motif compared to corresponding DNA molecules carrying alleles shared by the fetus and mother (CGN end motif: 1.7% vs. 3.7%, NCG end motif: 3.3% vs. 0.5%). These results further demonstrated that the use of CGN and NCG end motif frequencies can be informative for predicting methylation changes across several genomic regions of interest.

[0280] This data indicates that CGN and NCG correlate with the amount of fetal DNA. Because fetal DNA terminates at CGN at different rates than shared alleles, the amount of DNA molecules terminating at CGN can reflect the fractional concentration of fetal DNA. The same applies to DNA molecules terminating at NCG. Furthermore, because this behavior occurs in both hypermethylated and hypomethylated regions, both types of regions and both CGN and NCG can be used, thereby providing four types of measurements or performing only one type of measurement.

[0281] 2. Ratio of NCG and CGN To test the ability to determine the fractional concentration of fetal DNA, fetal-specific sites were identified and examined. Using bisulfite sequencing results from buffy coat (sequencing depth: 75x haploid genome coverage) and placental tissue (sequencing depth: 94x haploid genome coverage), placenta-specific hypermethylated (n = 188,978) and hypomethylated (n = 2,013,795) CpG sites were identified. Placenta-specific hypermethylated sites were defined as CpG sites with a methylation density greater than 70% in placental tissue but less than 30% in buffy coat samples. Placenta-specific hypomethylated sites were defined as CpG sites with a methylation density less than 30% in placental tissue but greater than 70% in buffy coat samples. The ratios of plasma DNA molecules terminated with CGN to NCG terminal motifs at placenta-specific hypermethylated and hypomethylated CpG sites, respectively, were determined. A total of 30 plasma DNA samples from pregnant women were analyzed using massively parallel paired-end sequencing (median paired-end reads: 206 million, IQR: 142-232 million).

[0282] In addition, fetal-specific alleles, specifically single nucleotide polymorphisms, were identified, and the number of reads with fetal-specific alleles was used to determine the SNP-based fetal fraction.

[0283] 49A-49B show the correlation between fetal DNA fraction and CGN / NCG ratio based on plasma DNA molecules derived from placenta-specific hypermethylated CpG sites (FIG. 49A) and placenta-specific hypomethylated CpG sites (FIG. 49B), according to embodiments of the present disclosure. Measurements were performed for various pregnancies at different gestational ages. The shape and color of the data points indicate whether the pregnancy was in the first, second, or third trimester. The amount of CGN only or NCG only (or any one terminal motif within these sets) can also be used.

[0284] The amount of sequence reads with CG at the end covering hypermethylated or hypomethylated sites was used to determine the amount of CGN. The amount of sequence reads with NCG at the end was used to determine the amount of NCG, where CG covered hypermethylated or hypomethylated sites.

[0285] As shown in Figures 49A and 49B, the CGN / NCG ratios from placenta-specific hypermethylated and hypomethylated CpG sites correlated positively and negatively, respectively, with the fetal fraction inferred by fetal-specific SNPs (Pearson r: 0.90 and -0.86, P-values ​​<0.0001 for both correlations). The data indicate that the relative amounts of plasma DNA molecules terminated with CGN and NCG (e.g., the CGN / NCG ratio) can be used to infer the likelihood of the tissue of origin of plasma DNA molecules and determine the fractional concentration of cell-free DNA of clinically relevant DNA, such as fetal DNA, tumor DNA, or transplant DNA.

[0286] The lines in Figures 49A-49B are examples of calibration functions. For any CGN / NCG ratio determined for a new sample, that value can be input into a linear function to provide an output on the X-axis of fetal fraction. Such a calibration function can be considered to include calibration values ​​for the calibration data points, including the corresponding fractional concentrations. Determining the fractional concentration can include comparing the new CGN / NCG ratio (or other characteristic described herein) to a calibration value for which the corresponding fractional concentration is known. The known fractional concentration can be determined via some other technique, such as SNP analysis. Because such a calibration function needs to be determined only once, such SNP-based analysis (which may not be applicable in all situations or may require more time, assays, and effort to identify SNPs) does not need to be performed on the new sample. Calibration functions, such as machine learning techniques (e.g., support vector regression) that still involve comparison to a calibration value, can be more complex.

[0287] 3. Independently, NCG or CGN In addition to using the CGN / NCG ratio, CGN and NCG can be used independently. In one embodiment, a measurement window (e.g., 11 bp) centered on each placenta-specific hypermethylated or hypomethylated CpG site can be defined as a placenta-specific hypermethylated and hypomethylated region. The CGN and NCG motif frequencies are calculated within the placenta-specific hypermethylated and hypomethylated regions. Each region has one CpG site. The motif frequency is determined by normalizing the amount of CGN or NCG with the total amount of terminal motifs from the region, for example, (CGN or NCG) / total amount of terminals in the region).

[0288] Figures 50A-50B show correlations between fetal DNA fraction and the frequency of NCG and CGN motifs based on plasma DNA molecules derived from placenta-specific hypermethylated CpG sites (50A) and placenta-specific hypomethylated CpG sites (50B), according to embodiments of the present disclosure. As shown in Figure 50A, the frequencies of NCG and CGN motifs derived from placenta-specific hypermethylated regions were negatively and positively correlated with the fetal fraction inferred by fetal-specific SNPs, respectively (Pearson r: -0.84 and 0.87, P-values ​​<0.0001 for both correlations). In contrast, the frequencies of NCG and CGN motifs derived from placenta-specific hypomethylated regions were positively and negatively correlated with the fetal fraction inferred by fetal-specific SNPs, respectively (Figure 50B, Pearson r: 0.87 and -0.69, P-values ​​<0.0001 for both correlations).

[0289] The CGN motif frequency was defined as the number of terminal motifs in the region with CGN. The NCG motif frequency was defined as the number of terminal motifs in the region with NCG. Such normalization can be performed in various ways. For example, instead of limiting the normalization factor to only the DNA within the measurement window around the CpG site as described above, the total number of cell-free DNA molecules or the total number of cell-free DNA ends analyzed (e.g., sequenced) can be used.

[0290] Each of these plots shows a calibration function that can be used in accordance with embodiments of the present disclosure, e.g., as described herein. For example, new values ​​of CGN motif frequency in hypermethylated regions can be compared to a calibration value (e.g., by inputting values ​​into the linear calibration function shown in the right plot of Figure 50A).

[0291] 4. CGN / NCG of hypo- and hypermethylated regions in ML model In another embodiment, the CGN / NCG ratios derived from placenta-specific hypermethylated CpG sites and hypomethylated CpG sites, respectively, can be integrated together as two independent features for machine learning. In other embodiments, individual amounts, such as CGN or NCG, can be used as independent features. Examples of such features are described above.

[0292] Such features can be input into a machine learning model, which can determine the fractional concentration of clinically relevant DNA. For example, a support vector regression (SVR) model can be used, which can receive values ​​ranging from 2 (e.g., any variation in the CGN / NCG ratio at hyper- and hypomethylated CpG sites) to 4 (e.g., the CGN and NCG amounts at hyper- and hypomethylated CpG sites). In other embodiments, other algorithms can be used, including, but not limited to, linear / nonlinear regression, decision trees, lasso regression, random forests, elastic networks, Bayesian networks, support vector machines, convolutional neural networks, recurrent neural networks, logistic regression, etc. As an example, an SVR model can be used to facilitate the determination of fetal DNA contribution in maternal plasma DNA of pregnant women.

[0293] Figure 51 shows the correlation of fetal fraction inferred by a SNP-based approach (x-axis) with the CGN / NCG ratio (y-axis) from placenta-specific hypermethylated and hypomethylated CpGs, according to an embodiment of the present disclosure. In this example, the SVR model was used.

[0294] As shown in Figure 51, the fetal DNA contribution from pregnancy samples estimated by the CGN / NCG ratio correlated with the fetal fraction estimated by the SNP-based approach (Pearson r: 0.95, P value < 0.0001). This correlation applied to all three gestational ages. The data suggested that the combined analysis of CGN / NCG ratios from regions showing tissue-specific hypomethylation and hypermethylation enhances the determination of the cfDNA contribution from the tissue of origin of plasma DNA molecules, for example, placental tissue.

[0295] The calibration data points shown in Figure 51 are examples of training data (e.g., calibration data points) used to determine an SVR calibration function. Such training data is derived from calibration samples, as described herein for other examples. Figure 51 illustrates the accuracy of the SVR calibration function.

[0296] 5. Examples of other tissue types Other tissue types besides fetus / placenta can be analyzed using loci with tissue-specific methylation status. For example, loci that are hypermethylated or hypomethylated in liver cells, blood cells, or colon cells can be used to determine the percentage (fractional concentration / contribution) of DNA fragments in samples (e.g., plasma or serum) derived from each tissue. To illustrate the use of such measurements in different tissue types, the relative changes in different tissue types in different types of subjects (implantation and pregnancy) are analyzed relative to healthy subjects.

[0297] As shown above (e.g., Figures 49A-49B), the CGN / NCG motif ratio from placenta-specific hypermethylated and hypomethylated CpGs correlated well with the fetal fraction in maternal plasma DNA. In addition to determining fractional concentrations, the terminal position of a given molecule at a tissue-specific locus (e.g., hypermethylated or hypomethylated in a given tissue) can be used to identify DNA fragments most likely to originate from a particular tissue. The likelihood that a given DNA fragment originates from a particular tissue is higher if the DNA fragment has a CGN motif at a hypermethylated locus in that tissue (e.g., indicating a likelihood of being methylated). In this way, terminal motifs can be used to enrich a sample for DNA fragments from a particular tissue type.

[0298] Figure 52 shows a schematic diagram of CGN / NCG motif ratio analysis based on tissue-specific methylated CpGs in blood and liver. Figure 52 shows cfDNA fragments terminating around CpG site 5105, which is hypermethylated in blood cells, CpG site 5115, which is hypermethylated in liver, CpG site 5125, which is hypomethylated in blood cells, and CpG site 5135, which is hypomethylated in liver. The schematic diagram shows how CGN / NCG motif ratio analysis can be performed on a set of tissue-specific methylated CpGs across different tissue types.

[0299] Tissue-specific hypermethylated CpGs were defined as CpGs with a methylation density of more than 70% in the target tissue (e.g., liver) and less than 30% across the remaining tissues analyzed. Tissue-specific hypomethylated CpGs were defined as CpGs with a methylation density of less than 30% in the target tissue and more than 70% across the remaining tissues. The CGN / NCG motif ratio was calculated for each set of tissue-specific CpGs in the plasma DNA samples.

[0300] For tissue-specific hypermethylated CpGs, the target tissue contributed more DNA molecules cleaving 5' to 'C' sites in the CpG context to the plasma DNA pool, thus resulting in a higher CGN / NCG motif ratio. Other tissues had the opposite preference for those CpG sites. Therefore, the CGN / NCG ratio of molecules derived from those tissue-specific hypermethylated CpG sites was positively proportional to the amount of cfDNA contributed by the target tissue. In contrast, for tissue-specific hypomethylated CpGs, the target tissue contributed fewer DNA molecules cleaving 5' to 'C' sites in the CpG context to the plasma DNA pool, resulting in a lower CGN / NCG motif ratio. The CGN / NCG ratio of molecules derived from those tissue-specific hypomethylated CpG sites was negatively proportional to the amount of cfDNA contributed by the target tissue.

[0301] Because the fractional concentration of DNA from blood cells is higher in plasma than that of DNA from the liver, the CGN / NCG ratio should be higher at blood cell-specific hypermethylated loci than at liver-specific hypermethylated loci. This relationship is shown at the bottom of the panel for hypermethylated CpGs. Conversely, the CGN / NCG ratio should be lower at blood-specific hypomethylated loci than at liver-specific hypomethylated loci, as shown at the bottom of the panel for hypomethylated CpGs.

[0302] As an example, based on bisulfite sequencing results of buffy coat, liver tissue, colon tissue, and placenta, we identified 7,467, 3,756, 7,726, and 1,000 hypermethylated CpG sites specific to blood cells, liver, colon, and placenta, respectively, along with 40,788, 19,619, 93,510, and 631,734 hypomethylated CpG sites specific to the corresponding tissues. We determined the ratios of plasma DNA molecules terminated with CGN to NCG end motifs at these tissue-specific hypermethylated and hypomethylated CpG sites, respectively. Massively parallel paired-end sequencing was used to analyze plasma DNA samples from one healthy control case (number of paired-end reads: 391 million), one liver transplant case (number of paired-end reads: 35 million), and one pregnant woman (number of paired-end reads: 238 million).

[0303] Figures 53A-53B show CGN / NCG motif ratios from different tissue-specific hypermethylated (Figure 53A) and hypomethylated (Figure 53B) CpGs in plasma DNA of healthy controls, liver transplant cases, and pregnant cases. For each tissue type (blood, colon, liver, and placenta), motif ratios are provided for three different subjects (healthy controls, liver transplants, and pregnant women). The relative values ​​of the motif ratios indicate increasing and decreasing contributions for each of the different subject tissue types.

[0304] As shown in Figure 53A, in healthy control cases, the CGN / NCG motif ratio derived from blood cell-specific hypermethylated CpG sites corresponds to the highest value, followed by liver, and the lowest value from placenta.This is because liver transplant recipients contribute more liver DNA, thereby reducing the contribution from blood cells.Similarly, pregnant women have cell-free fetal DNA, so that the contribution from blood is reduced.

[0305] For the tissue-specific hypomethylated CpG sites in Figure 53B, the CGN / NCG motif ratios appeared in reverse order. A higher CGN / NCG motif ratio in tissue-specific hypermethylated CpGs reflects a higher methylation density of plasma DNA at those tissue-specific CpG sites of interest, indicating a higher contribution from that tissue. A lower CGN / NCG motif ratio in tissue-specific hypomethylated CpGs reflects a lower methylation density of plasma DNA at those tissue-specific CpG sites of interest, indicating a higher contribution from that tissue.

[0306] These results are consistent with the notion that blood cells are the primary source of cell-free DNA contributing to plasma (Sun et al. Proc Natl Acad Sci USA. 2015;112:E5503-5512). In liver transplant patients (SNP-based approach), in whom the liver cfDNA fraction was approximately 40%, the CGN / NCG motif ratios derived from liver-specific hypermethylated and hypomethylated CpG sites showed significant increases and decreases, respectively, compared with plasma DNA in healthy controls. The liver contribution is approximately 10% in healthy controls.

[0307] In pregnancy cases in which the fetal cfDNA fraction was approximately 39% (SNP-based approach), the CGN / NCG motif ratios derived from placenta-specific hypermethylated and hypomethylated CpG sites showed a significant increase and decrease, respectively, compared with plasma DNA of healthy controls. The colonic contribution remained approximately the same across different subjects.

[0308] The data suggested that the relative amounts of plasma DNA molecules terminated with CGN and NCG (e.g., CGN / NCG ratio) can be used to infer the tissue of origin of plasma DNA molecules from multiple tissues. Thus, the relative changes (increases or decreases) in the contributions of different tissues can be determined using motif ratios. For example, baseline motif ratios of healthy controls can be used to determine whether a particular tissue contribution has significantly changed by comparing the new motif ratios of a subject whose motif ratio is relative to a tissue-specific locus. The specific fractional contribution of any particular tissue type can be determined using the respective calibration functions of that particular tissue type.

[0309] B. Tissue-specific methylation patterns of cleavage profiles In addition to or instead of using terminal motifs (e.g., CGN and NGC), cleavage profiles can be used. The use of cleavage profiles can be similar to those described above. One or more quantities (e.g., cleavage ratio / density) can be determined. Such quantities can be used with one or more calibration values ​​(e.g., as a calibration function) whose fractional concentrations are known and form calibration data points, as described herein. Thus, the amount of cleavage at a given position or the cleavage profiles of two or more positions can be used.

[0310] 1. Cleavage profile of hyper- and hypomethylated sites Figures 54A-54B show that fetal-specific methylation is reflected in the cleavage profile of cfDNA fragmentation, according to embodiments of the present disclosure. In this plot, the black line 5405 represents the cleavage profile of fetal-specific DNA, and the gray line 5410 represents the cleavage profile of shared DNA. Figure 54A shows the cleavage profile of fetal-specific and shared DNA molecules derived from placenta-specific hypermethylated CpG sites. The x-axis represents position relative to the CpG site of interest. The y-axis represents the average cleavage ratio. Figure 54B shows the cleavage profile of fetal-specific and shared DNA molecules derived from placenta-specific hypomethylated CpG sites.

[0311] The cleavage profiles of fetal-specific and shared plasma DNA molecules from one pregnancy cfDNA sample were determined based on placenta-specific hypermethylated (n=188,978) and hypomethylated (n=2,013,795) CpG sites, respectively. Placenta-specific hypermethylated and hypomethylated CpG sites were identified using bisulfite sequencing results of buffy coat (sequencing depth: 75X) and placental tissue (sequencing depth: 94X). Placenta-specific hypermethylated sites were defined as CpG sites with a methylation density of greater than 70% in placental tissue but less than 30% in maternal buffy coat samples. Placenta-specific hypomethylated sites were defined as CpG sites with a methylation density of less than 30% in placental tissue but greater than 70% in maternal buffy coat samples.

[0312] For placenta-specific hypermethylated CpG sites, fetal-specific fragments showed higher cleavage ratios at CpG sites than shared fragments (mainly of hematopoietic origin). For placenta-specific hypomethylated CpG sites, fetal-specific fragments showed lower cleavage ratios at CpG sites than shared fragments. These results indicated that differences in tissue-specific cfDNA methylation can be reflected in cfDNA fragmentation patterns. Methylation patterns can be used to infer DNA contributions from different tissues / organs to the plasma DNA pool (Sun et al. Proc Natl Acad Sci USA. 2015;112:E5503-5512). Therefore, we speculated that cleavage profiles may be useful for inferring DNA contributions from different tissues to the plasma DNA pool, such as placental DNA contributions in pregnant women's plasma DNA.

[0313] 2. SVR Implementation Using pregnancy samples as an example, cleavage profiles were determined from placenta-specific hypermethylated and hypomethylated CpG sites. Such cleavage profiles from 30 pregnancy samples (median number of paired-end sequencing reads: 206 million (IQR: 142-232 million)) were used to train an SVR model to predict fetal DNA contribution based on a leave-one-out strategy. In other embodiments, SVR can use the abundance at position 0 (e.g., cleavage ratio), which can be complemented by abundances at other positions.

[0314] Figures 55 and 56 show marker selection, data collection, and training of an SVR model to determine fetal fraction in cell-free samples (e.g., plasma) according to an embodiment of the present disclosure, using bisulfite sequencing data from buffy coat and placenta. Marker references are the same as those described in the previous section.

[0315] For data collection, individual samples are sequenced and the cleavage profile for each CpG site is calculated. For hypermethylated CpGs, a merged cleavage profile is calculated, which is the average cleavage ratio for each position. For a single position, all counts from individual sites are averaged to create a single count. The same is done for hypomethylated sites.

[0316] For each patient, there are two vectors: one for hypermethylated CG sites and another for hypomethylated CG sites. Each vector contains 11 numbers reflecting the average cleavage ratio at each position. Other numbers of values ​​can be used for different widths of the cleavage profiles described herein, such as widths of 4, 5, 6, 7, 8, 9, and 10. For each training sample, the fetal fraction is determined using fetal-specific alleles.

[0317] The SVR model is trained using the leave-one-out strategy, which uses 29 samples to train the model, and then predicts the fetal fraction of one sample.Then, train with 29 different samples, and test another sample until all 30 samples are included in the test set and the fetal fraction is predicted.The predicted fetal fraction is compared with the SNP-based analysis.

[0318] Figure 57 shows the correlation between the fetal fraction inferred by the proximal cleavage profile and the SNP-based approach, according to embodiments of the present disclosure. The estimated fetal DNA contribution from pregnancy samples by the cleavage profile correlated well with the fetal fraction inferred by the fetal-specific SNPs (Pearson r: 0.97, P-value < 0.0001). In various embodiments, the cleavage profile can be used to predict the DNA contribution of other tissues, including but not limited to, liver, brain, T cells, B cells, neutrophils, muscle, heart, placenta, ovary, breast, testis, etc.

[0319] The tissue fraction may be of clinically relevant DNA, for example, tumor DNA.

[0320] 3. Cleavage profile of individual sites The above example combines measurements across CpG sites, and then uses the resulting cleavage profile.For example, the ratio of the amount of DNA molecules that end at position 0 can be compared with the amount of DNA molecules that end at position -1 or 1 (or the sum of both).As described above, other positions can be used as well, or instead.As another example, the cleavage profile can be input into a machine learning model.

[0321] However, in other embodiments, a separate amount can be determined for each CpG site and compared with each calibration value separately.Thus, if N CpG sites are analyzed, N amounts can be determined, each amount corresponding to the amount of DNA fragment at each CpG site.In addition, a separate cleavage profile can be determined for each CpG site (or a subset of sites), and these individual cleavage sites can be used, for example, as input into a machine learning model.Each cleavage profile can be input into a machine learning model to determine the fractional concentration of that particular CpG site.Such fractional concentrations can be averaged (including weighted average based on sequencing depth).

[0322] In another example, all of the cleavage profiles can be input into the machine learning model simultaneously. In this way, more features are input into the machine learning model. Such a feature matrix can be the size of the number of CpG sites times the number of positions in the measurement window.

[0323] In a different example, a separate cleavage ratio can be determined for each cleavage profile of a different subset of CpG sites and the individual fractional concentrations can be averaged.

[0324] C. Determining the fractional concentration of DNA from the first tissue using tissue-specific methylation sites The fractional concentration of cell-free DNA molecules (cell-free DNA fragments) from a particular tissue type (e.g., clinically relevant DNA) can be determined in a variety of ways. For example, the amount of DNA fragments terminating at the 0 or -1 position of one or more CpG sites in a new sample and calibration information (e.g., from a calibration sample with known or measured fractional concentrations by other means) can be used to determine the fractional concentration in the new sample.

[0325] 58 is a flowchart illustrating a method 5800 for determining the fractional concentration of clinically relevant DNA according to an embodiment of the present disclosure. The biological sample includes cell-free DNA. As with other methods described herein, method 5800 can be implemented in part or in whole using a computer system. Blocks performing similar functions to other methods can be implemented similarly to what is described for those blocks in other flowcharts.

[0326] At block 5810, method 5800 includes analyzing a plurality of cell-free DNA molecules from the subject's biological sample. Analyzing the cell-free DNA molecules can include determining a genomic location in the reference genome corresponding to at least one end of the cell-free DNA molecule. Such analysis of each of the plurality of cell-free DNA molecules at block 5810 can be performed using any of the techniques described herein, as described with respect to other methods herein, such as at block 2710. In some embodiments, analyzing the cell-free DNA molecules can include determining a terminal sequence motif of at least one end of the cell-free DNA molecule, wherein the ends of the cell-free DNA molecule have a first position at the outermost position, a second position adjacent to the first position, and a third position adjacent to the second position.

[0327] The analysis can be performed in a variety of ways (e.g., as described herein) and can include physical steps that can be controlled by a computer. For example, the analysis can be performed by sequencing, PCR (e.g., droplet digital PCR), mass spectrometry, or electrophoresis.

[0328] At block 5820, method 5800 includes identifying a first set of CpG sites in the reference genome that all have a first differential methylation for the first tissue type. As an example, the first differential methylation can be that all of the first set of CpG sites are hypomethylated, or that all of the first set of CpG sites are hypermethylated. Another example is that all of the first set of CpG sites are 5hmC-enriched, or that all of the first set of CpG sites are 5hmC-depleted. In some embodiments, both types of CpG sites can be used, for example, when the amount of each is used as a separate feature for a machine learning model. In such an embodiment, a second set of CpG sites of the other type (e.g., hypermethylated if the first set is hypomethylated) can be identified, and separate amounts can be determined for the second set of sites.

[0329] The first set of CpG sites may include one or more CpG sites, for example, one CpG site or multiple CpG sites.

[0330] In block 5830, method 5800 includes determining a first amount of cell-free DNA molecules terminating at position 0 or -1 of any of the first set of CpG sites. Such an example corresponds to using a CGN motif at position 0. The first amount terminating at a CpG site at position 0 or -1 can be further filtered so that only a subset of CGN motifs, such as CGA, CGT, CGC, or CGG, or a subset of these four terminal motifs, is used. Such combinations can also be used for other techniques described in other sections herein. Other amounts can be determined, for example, at other positions surrounding the CpG site, for example, to determine a CGN / NCG ratio or cleavage profile.

[0331] In another embodiment, or as part of determining the differential features for the model, an NCG terminal motif corresponding to the -1 position can also be used. For example, a first amount of a first set of one or more terminal sequence motifs having a C at the second position and a G at the third position. The first set of terminal motifs can include one or more of the trimer motifs ACG, CCG, GCG, and TCG.

[0332] Thus, a first amount of a first set of one or more terminal sequence motifs can be determined. The first set of one or more sequence motifs can have a C at a first position and a G at a second position. As another example, the first set of one or more sequence motifs can have a C at a second position and a G at a third position.

[0333] The first amount can be normalized before being compared with the calibration value. Normalization can be performed in various ways. For example, normalization can use the number of cell-free DNA molecules that end within the region containing the CpG site (e.g., local normalization as described above). As another example, normalization can use the number of cell-free DNA molecules that cover the CpG site (e.g., using the sequencing depth as described herein). As yet another example, normalization can use the average or median depth of cell-free DNA molecules in the region containing the CpG site.

[0334] When more than one CpG site is analyzed, instead of a single amount for each CpG site, a respective amount can be determined at each CpG site (e.g., as described in section B.3 above). Thus, determining a first amount can include determining the respective amounts of cell-free DNA molecules terminating at the 0 or 1 position of each of the first set of CpG sites. Then, determining the fractional concentration of DNA from the first tissue type can include comparing each of the respective amounts to a respective calibration value. Such respective amounts can be used in the model's feature vector. Thus, comparing each of the respective amounts to a respective calibration value can be performed by inputting each amount into a machine learning model as part of a feature vector. The entire cleavage profile of each site can be used (see B.3 above). Thus, for each CpG site, a cleavage profile (e.g., for each of at least two positions within a window around the CpG site) can be determined, for example, by determining the respective amount of cell-free DNA molecules terminating at that position. The respective amounts of the at least two positions and the CpG site can be included in the feature vector.

[0335] At block 5840, method 5800 includes determining a fractional concentration of DNA from the first tissue type in the biological sample by comparing the first amount to a calibration value. The calibration value can be determined from one or more calibration samples having known fractional concentrations of DNA from the first tissue type. The fractional concentration of DNA can have various resolutions, for example, above or below a certain percentage, or within a range. Such ranges can have various resolutions, such as 5%, 10%, 15%, 20%, 25%, and 30% or less resolution. As an example, 5% resolution can correspond to 35% to 40%.

[0336] In various embodiments, such techniques may be used to determine the DNA contribution of a variety of tissues, including, but not limited to, liver, brain, T cells, B cells, neutrophils, muscle, heart, placenta, ovary, breast, testis, tumor tissue (e.g., hepatocellular carcinoma, lung cancer, colon cancer, ovarian cancer), diseased tissue (e.g., tissue involved in inflammation (such as liver tissue involved in hepatitis, peritumoral tissue, and others), and tissue with increased cell death (such as ischemic or necrotic tissue, or tissue involved in traumatic injury), and tissue infected with one or more infectious agents (such as viruses or bacteria).

[0337] As described above, other quantities can be determined. For example, if the first quantity uses the 0 position, a second quantity of cell-free DNA molecules terminating at a first position within a window surrounding any one of the first set of CpG sites, such as the -1 position corresponding to NGC, can be determined. Determining the fractional concentration can then include normalizing the first quantity by the second quantity to obtain a normalized first quantity (e.g., CGN / NGC) that is compared to a calibration value. For example, the first position can be +1 or -1 from one of the one or more CpG sites. For another example, the window can be at least -2 to +2 from the CpG site, or at least -4 to +4 from the CpG site.

[0338] In some embodiments, a cleavage profile can be determined within the window, for example, as described above. The cleavage profile can be for at least two positions (e.g., a first position) within the window other than one of the CpG sites. For each of the at least two positions within the window, the respective amounts (including the second amount) of cell-free DNA molecules terminating at that position can be determined. As part of determining the fractional concentration, the first amount of cell-free DNA molecules terminating at position 0 or -1 of the CpG site can be compared with the respective amounts of cell-free DNA molecules terminating at that position. For example, separate ratios of each of the respective amounts and the first amount can be determined.

[0339] In other embodiments, the comparison of the respective amounts can be performed as part of a machine learning model (e.g., support vector regression). For example, a feature vector can include the respective amounts and the first amount. The feature vector can then be input into a machine learning model to determine the fractional concentration. The machine learning model can be trained using cell-free DNA molecules from one or more calibration (training) samples having known (e.g., measured) fractional concentrations of DNA from the first tissue type. Training can be performed using a variety of techniques, as will be understood by those skilled in the art.

[0340] When a first set of one or more CpG sites are all hypomethylated for a first tissue type in the reference genome, a second set of one or more CpG sites that are all hypermethylated can be used (e.g., as depicted in Figures 45-47). A first feature vector can be generated for the first set, and a second feature vector can be generated for the second set. Both feature vectors can be input into a machine learning model as part of determining the fractional concentration. In any of these exemplary techniques using a window, the at least two positions can include all positions within the window. By way of example, the window can be at least +2 to -2 from the CpG site, or at least +4 to -4 from the CpG site.

[0341] As described in Section A above, multiple tissue types can be analyzed to determine fractional concentrations or changes therein. In such an example, a second set of one or more CpG sites can be identified in the reference genome that are all hypomethylated or all hypermethylated for the second tissue type. A second amount of cell-free DNA molecules terminating within a window surrounding any one of the second set of CpG sites can then be determined. The second amount can then be compared to another calibration value (e.g., a baseline in healthy controls) to determine the change in the fractional concentration of DNA from the second tissue type in the biological sample. Thus, another calibration value can be determined from one or more other calibration samples.

[0342] D. Determining fractional concentrations in multiple tissues As noted above, some embodiments can perform deconvolution to determine fractional concentrations in multiple tissues simultaneously. Furthermore, such techniques do not require identification of sites that are hypermethylated or hypomethylated in a particular tissue type. Cleavage profiles and / or terminal motif profiles can be used to estimate fractional concentrations in different tissue types.

[0343] The reference matrix A can include values ​​at a set of sites, such as CpG sites and surrounding sites, for example, when cleavage profiles are used. Terminal motif values ​​can also be used. Each column in the reference matrix can correspond to a different tissue type or a different template sample (e.g., designated contributions for different tissues). Measurements of the value b (measurement vector) in a cell-free mixture (e.g., plasma, urine, etc.) can be used with A to determine the contribution x of each tissue type for each template sample. This can be done by solving Ax = b.

[0344] The contribution vector x provides the percentage of each tissue type when the columns in A correspond to different tissue types. In such embodiments, reference values ​​in the reference matrix from cells of tissues can be prepared using various DNA nucleases (e.g., DNASE1L3, DNASE1, DFFB, etc.) to digest nuclei from their corresponding tissues. For example, nuclei from liver tissue can be subjected to DNASE1L3-based digestion. The sequencing data of the digested DNA is analyzed to obtain reference values ​​(e.g., cleavage profiles and / or terminal motif profiles) at various sites to obtain tissue-specific profiles of fragmentomics features against which the plasma fragmentomics features of cfDNA can be compared.

[0345] In another example, references can be generated through the analysis of tissue-specific cfDNA molecules.For example, liver transplant cases can be used to identify liver-specific DNA (for example, using tissue-specific alleles), and liver-specific sequence reads can be used to determine tissue-specific profiles.Similar techniques can be used for other tissues.

[0346] When template samples are used, each template sample can correspond to a set of percentages that add up to 100%.For example, a liver transplant subject can be used as a template, with a specified percentage for each of a set of tissues.Other templates can be for each organ pathology, for example, so that there is a template for each tissue type whose contribution is determined.A healthy subject can also be used.To determine the contribution rate of each tissue, the contribution of the template is used in combination with the percentage of that tissue type for that template sample, and this is performed for each template sample.

[0347] VIII. Pathological Detection In addition to determining clinically relevant DNA methylation and fractional concentrations, cleavage profiles and terminal motif groups (CGN and NCG) were used to determine disease detection (e.g., cancer). For terminal motifs, global analysis and quantification were performed at specific CpG sites, e.g., sites that are hypermethylated in specific tissue types and / or sites that are hypomethylated in specific tissue types.

[0348] Cleavage profiles were also used at specific CpG sites, e.g., sites that are hypermethylated in a particular tissue type and / or sites that are hypomethylated in a particular tissue type, as well as CpG sites that do not have differential methylation between tissues but have different cleavage profiles for different levels of pathology. Different types of pathology (e.g., different types of cancer) can be determined using informative sites specific to different types of pathology (e.g., different types of cancer). Surprisingly accurate results were obtained using cleavage information (e.g., terminal motifs and cleavage profiles) around CpG sites relative to using other terminal motifs, e.g., CCA or all-trimer terminal motifs).

[0349] Viral and human DNA were used in different techniques. Exemplary pathologies include cancer and autoimmune diseases (e.g., systemic lupus erythematosus (SLE)). The classification of pathology can be the presence or absence of pathology. In another example, the classification can be of the severity of the pathology described herein, such as, for example, the size of one or more tumors or the surface area of ​​one or more tumors. Such severity can be determined based on a range of parameters above a threshold / cutoff or relative to a calibration level from subjects known to have a particular severity (e.g., cancer stage).

[0350] A. Use of CGN and / or NGC terminal motifs Cancers are typically hypomethylated when methylation is determined genome-wide. This global hypomethylation is described in PCT Publication WO2014 / 043763, which is incorporated by reference in its entirety for all purposes. Figure 28 above shows that fragmentation at CpG sites reflects methylation. The following analysis demonstrates that pathologies such as cancer can be determined using fragmentation, for example, across the entire genome or across repeated sections of the genome. Such analysis can also be performed on all or a large portion of the genome, for example, at least 10 Mb, 20 Mb, 50 Mb, 100 Mb, 500 Mb, 1 Gb, or more. The portion of the genome need not be contiguous; for example, 50 Mb can be divided among five discrete (i.e., non-contiguous) 10 Mb regions.

[0351] 1. CGN / NCG ratio The whole-genome CGN / NCG ratio is used to reflect the whole-genome methylation level and can be used for disease detection. Global methylation density was analyzed for six healthy controls and eight patients with hepatocellular carcinoma (HCC) with a median of 424 million paired-end bisulfite sequencing reads (IQR: 352-519 million).

[0352] Figures 59A-59C show boxplots illustrating the differences in genome-wide methylation density (Figure 59A), genome-wide CGN / NCG ratio (Figure 59B), and CGN / NCG ratio in Alu regions (Figure 59C) between healthy controls and HCC cases, according to embodiments of the present disclosure. As shown in Figure 59A, HCC cases showed significantly lower methylation levels compared to healthy controls (P value < 0.001, Wilcoxon test). The genome-wide CGN / NCG ratio was analyzed for each patient and showed a significant reduction in the HCC group compared to the control group (Figure 59B, P value < 0.001, Wilcoxon rank-sum test).

[0353] In addition to genome-wide analysis, we hypothesized that changes in the CGN / NCG ratio across Alu regions or other repetitive elements could be useful to reflect corresponding methylation changes. Over 90% of methylated CpG sites in the human genome occur in DNA repetitive elements such as Alu and LINE-1 (Zheng et al. Nucleic Acids Research. 2017;45:8697-8711). Alu regions can effectively represent the entire genome. Other repetitive regions can also be used. As an example, as shown in Figure 59C, the CGN / NCG ratio from Alu regions showed a significant decrease in HCC cases compared to controls (P value <0.001, Wilcoxon rank-sum test). These results demonstrate that the use of genome-wide CGN / NCG ratios enables the determination of genome-wide global methylation and cancer detection.

[0354] Therefore, global and repeat-specific fragmentation patterns specific to CGN and NCG (e.g., as a ratio or individually) can provide accurate disease detection. The separation between healthy and HCC is surprising for the use of fragmentation values ​​that do not require the use of methylation-aware assays.

[0355] 2. CGN or NCG The CGN and NGC terminal motifs can also be used individually. CGN and NCG motif frequencies were calculated across the genome and for Alu regions.

[0356] Figures 60A-60D show boxplots illustrating the difference in CGN and NCG motif frequencies from the whole genome and Alu regions between healthy controls and HCC cases, according to an embodiment of the present disclosure. As shown in Figures 60A-60D, the CGN frequencies from the whole genome and Alu regions were significantly reduced in the HCC group compared to the control group (P value < 0.0047, Wilcoxon rank sum test). Furthermore, the NCG frequencies from the whole genome and Alu regions were significantly increased in the HCC group compared to the control group (whole genome: P value = 0.05, Alu region: P value = 0.0019, Wilcoxon rank sum test).

[0357] Separation is better using the ratio of CGN / NCG (or its variants), but the individual amounts of these terminal motifs can also be used.

[0358] 3.Multiple cancer analysis Although the above example was for HCC, other cancers can also be detected. Because ALU has better separation, we used the ALU region for the CGN / NCG ratio for the analysis of different cancer types. A decrease in the CGN / NCG ratio in the ALU region can facilitate the detection of various cancers (i.e., pan-cancer detection). To this end, we analyzed six healthy controls, as well as eight patients with HCC, ten patients with head and neck squamous cell carcinoma (HNSCC), ten patients with colorectal cancer (CRC), ten patients with lung cancer (LC), and ten patients with NPC, using a median of 87 million paired-end sequencing reads (IQR: 73–127 million).

[0359] Figure 61 shows the Alu region CGN / NCG ratios from healthy controls, HCC, HNSCC, CRC, LC, and NPC according to an embodiment of the present disclosure. The P value between the two groups was calculated using the Wilcoxon rank-sum test. As shown in Figure 61, the CGN / NCG ratios from the Alu region from cancer subjects, including but not limited to HCC, HNSCC, CRC, LC, and NPC, were significantly lower than those from control subjects (P values: HCC = 0.00067, HNSCC = 0.042, CRC = 0.0047, LC = 0.003, NPC = 0.042, Wilcoxon rank-sum test). Therefore, the CGN / NCG ratio can serve as a biomarker for pan-cancer analysis. A better reduction in the separation value was observed for CRC, LC, and HCC.

[0360] 4. PCR analysis PCR can be used for any of the techniques described herein, for example, as described above in Section I. Such PCR analysis can be used to measure global CGN / NCG ratios for cancer diagnosis, as well as the methylation and fractional concentration of DNA from specific tissue types. Thus, digital or single-molecule PCR assays can be used to determine CGN / NCG ratios by utilizing repetitive regions, such as Alu regions. In various examples, digital PCR can be droplet digital PCR assays or other types of partition-based (e.g., beads or surfaces (e.g., on glass or in flow cells) for carrying out the reaction). In another embodiment, digital PCR can be enabled using microfluidic chips, for example, from Fluidigm.

[0361] Figure 62 shows an exemplary technique using droplet PCR to determine NCG and CGN motifs in a region of interest, as well as exemplary structures of probes, according to embodiments of the present disclosure. For example, Figure 62 shows probe 6210 targeting DNA template sequence 6220 ligated with common adaptor 6230. PCR can use a probe that covers the junction of common adaptor 6230 and DNA template sequence 6220.

[0362] In the example shown, different probes are used for CGN and NCG. In this way, NCG can be distinguished from CGN. For example, for the junction, the probe can have a size of approximately 15 bp. For example, one portion of the probe hybridizes to the common adapter region, for example, 7-8 bp, while the other portion of the probe, for example, 7-8 bp, hybridizes to the target Alu sequence. As will be understood by those skilled in the art, other sizes of hybridized portions can be used. For potential experimental design, we focused on the most frequently occurring sequences for the top CG probe and top NCG probe, i.e., the 7-8 bp probe in the Alu region.

[0363] Figures 63A-63B show the top 10 CG probes and the top 10 NCG probes for potential experimental designs according to embodiments of the present disclosure. Different combinations were analyzed to see which probes provided the best results. Different combinations were performed using one probe from CGN and one probe from NCG to determine the ratio.

[0364] Using the CG probe "CGCCTGT" and the NCG probe "TCGCTTGA" as examples, in silico estimation of the potential performance of ddPCR was performed using the number of cfDNA fragments ending in "CGCCTGT" and "TCGCTTGA" (determined using sequencing results) in healthy controls and cancer patients. The CGN / NCG ratio in the target Alu region to be read by the ddPCR assay is reflected as the amount of "CGCCTGT"-terminated fragments divided by the amount of "TCGCTTGA"-terminated fragments in the sequencing results. In silico estimation uses sequencing data to identify sequences and then select the corresponding sequences as if they were detected using PCR probes.

[0365] Figure 64 shows an in silico analysis of the CGN / NCG ratio by ddPCR assay using probes containing the Alu-related sequences "CGCCTGT" and "TCGCTTGA" across healthy controls and subjects with HCC, HNSCC, CRC, LC, or NPC, according to an embodiment of the present disclosure. The P value between the two groups was calculated by Wilcoxon rank-sum test. As shown in Figure 64, the CGN / NCG ratio from cancer subjects was significantly lower than that from control subjects (P values: HCC = 0.0027, HNSCC = 0.0047, CRC = 0.0005, LC = 0.001, NPC = 0.001, Wilcoxon rank-sum test). As shown above, the separation was as good for all of the cancers or, for some cancers, better than using sequencing trimers for the CGN / NCG ratio.

[0366] Figures 65A-65E show ROC analysis of using the CGN / NCG ratio from an in silico ddPCR assay to distinguish patients with HCC, HNSCC, CRC, LC, or NPC from healthy controls, according to embodiments of the present disclosure. The analysis used the same two sequences used in Figure 64. The CGN / NCG ratio inferred from the target Alu region assayed by ddPCR was able to distinguish HCC (AUC: 0.96), HNSCC (AUC: 0.92), CRC (AUC: 0.98), LC (AUC: 0.97), and NPC (AUC: 0.97) from healthy controls. The ROC was surprising for the use of these specific terminal motifs, compared to using all trimer motifs in HCC, which provided an ROC of 0.872 in a previous study.

[0367] Probes with different sizes and / or multiplexes can be used for cancer detection. Probes targeting other regions of interest, such as LINE-1, can be used for cancer detection and monitoring. In another example, mass spectrometry (e.g., using a time-of-flight mass spectrometer) can be used to detect reaction products, including DNA molecules.

[0368] A further example is the use of qPCR (real-time PCR) to detect CGN and NCG frequencies from specific regions.

[0369] Figure 66 shows an exemplary method using qPCR to determine NCG and CGN motifs in a region of interest. This method is similar to the method described in Figure 4, so the description of Figure 4 can be applied to this technique. As shown, cell-free DNA molecules can undergo DNA end pairing, A-tailing, and common adapter ligation as optional steps. A pair of PCR primers can be designed so that one primer (e.g., common forward primer 6645) binds to the common adapter region 6615, and the other primer (e.g., region-specific reverse primer 6650) binds to a specific region of interest. The DNA molecules are amplified by the pair of PCR primers, which hydrolyze a fluorescent probe specific to a particular terminal motif (such as a CGN or NCG terminal motif) to release a fluorescent signal, thus enabling detection of the presence of the specific motif as well as quantification of the specific motif.

[0370] Figures 67A-67C show examples of primer and probe design. The amplification region (i.e., between the two primers) can be a repetitive region, including, but not limited to, LINEs (long interspersed nucleotide sequences) and SINEs (short interspersed nucleotide sequences). For illustrative purposes, a reverse primer (i.e., CCACTGCACTCCAGCCTG) was designed to target the consensus sequence of an Alu element, a type of SINE (Figure 67A). This primer pair, along with a forward primer targeting an added common adapter, can be used to amplify Alu regions with a conserved terminal motif at the 5' end of cfDNA. Two probe hybridization sites (i.e., CGCCTGT and TCGCTTGA) were selected from the desired Alu consensus sequence (i.e., 200 bp before the reverse primer) (Figure 67A). A CGN probe (i.e., CTCTTCCGATCTCGCCTGT) and an NCG probe (i.e., CTCTTCCGATCT TCGCTTGA) were designed with a 12 nt sequence from the common adapter (Figure 67B).

[0371] To test the specificity of these probes, based on the sequencing read of cfDNA, determine the possible hybridization sites in the whole genome.As shown in Figure 67C, most of the probe hybridization sites are located in Alu region, and only about 0.3% are off-target.Therefore, these two probes can provide good coverage and provide good results as described below.

[0372] Such PCR-based methods were tested on one mechanically fragmented gDNA sample, one healthy control cfDNA sample, and one early-stage HCC cfDNA sample. For qPCR testing, the relative fold change in the CGN / NCG motif ratio was calculated based on the C values ​​determined from the qPCR assay.

[0373] Figures 68A-68B show the test results from qPCR (68A) and ddPCR (68B). gDNA represents the mechanically fragmented gDNA sample, used as a control in the qPCR assay. For the mechanically fragmented gDNA sample, the fold change in the CGN / NCG motif ratio was compared to the control (i.e., fragmented gDNA sample) (CT CGN control ) and test sample (CT CGN testing ) from the CGN probe, as well as the control (CT NCG control ) and test sample (CT NCG testing ) was calculated based on the CT values ​​determined from the NCG probe as follows:

number

[0374] For ddPCR tests, the CGN / NCG motif ratio was calculated as the amount of CGN probe-positive droplets divided by the amount of NCG probe-positive droplets.

[0375] Because the gDNA samples were mechanically sheared to 200 bp, the terminal motifs present in the gDNA samples tended to be random (i.e., the CGN / NCG motif ratio was close to 1). In healthy control cfDNA samples, Alu regions are typically hypermethylated, so there were more CGN terminal motifs than NCG terminal motifs (i.e., CGN / NCG motif ratios >1). Indeed, both qPCR and ddPCR test results indicated that the healthy control cfDNA samples exhibited substantially higher CGN / NCG motif ratios than mechanically fragmented gDNA. Furthermore, both qPCR and ddPCR test results indicated that the CGN / NCG motif ratios from HCC cfDNA samples were lower than those from healthy control cfDNA, reflecting reduced methylation levels (i.e., hypomethylation) in the HCC samples.

[0376] Tumor-derived DNA was shorter than background DNA, primarily of hematopoietic origin. Because the CGN probe was 86 nt away from the reverse primer, while the NCG probe was only 45 nt away, the amplicon size corresponding to the CGN probe was longer than that corresponding to the NCG probe. Therefore, the amplicon size corresponding to the NCG probe increased in patients with cancer compared with the CGN probe due to a higher probability of having both the NCG probe and the reverse primer on the same fragment, i.e., the shorter distance. The result was an extra reduction in the CGN / NCG motif ratio due to size characteristics. Therefore, the CGN / NCG motif ratio targeting the Alu region could synergistically enhance the cancer signal associated with hypomethylation and size shortening of tumor-derived DNA molecules, leading to better performance in cancer detection, monitoring, and prognosis.

[0377] 5. Terminal motifs of fragments of specific sizes In some embodiments, the methylation-aware fragmentation analysis described in this disclosure can be performed using long-read sequencing technologies, including, but not limited to, nanopore sequencing (e.g., Oxford Nanopore Technologies) and single-molecule real-time sequencing (e.g., Pacific Biosciences). Plasma DNA samples from 15 healthy subjects were sequenced using single-molecule real-time sequencing (Pacific Biosciences) (median reads: 970,633, IQR: 407,695-1,399,709). Putative hypermethylated (4,631,823) and hypomethylated (307,831) CpGs in plasma were identified based on bisulfite sequencing of plasma DNA samples from eight healthy subjects. Putative hypermethylated CpGs were defined as CpGs with a methylation density greater than 70% across all cases. Putative hypomethylated CpGs were defined as CpGs with a methylation density less than 30% across all cases. The CGN / NCG motif ratios of putative hypermethylated and hypomethylated CpGs in plasma were analyzed by pooling all single molecule real-time sequencing reads from these healthy subjects.

[0378] Figures 69A-69B show the CGN / NCG motif ratios in putative hypermethylated and hypomethylated CpGs across different size ranges of pooled single-molecule real-time sequencing reads from healthy subjects. As shown in Figure 69A, the CGN / NCG motif ratios of putative hypermethylated CpGs decreased with increasing cfDNA fragment size. As shown in Figure 69B, the CGN / NCG motif ratios of putative hypomethylated CpGs decreased with increasing cfDNA fragment size for the size range of ≦1000 bp, while the values ​​increased with increasing cfDNA fragment size for the size range of >1000 bp.

[0379] Figures 69A-69B show that cleavage preferences differ when the size range is different. For long molecules, the CGN / NCG motif ratios of hypermethylated and hypomethylated CpG sites are closer than when the size is lower than 200. Therefore, we investigated whether analyzing DNA fragments of a specific size range could provide increased accuracy.

[0380] To illustrate the usefulness of taking fragment size into account during methylation-aware fragmentation analysis, for illustrative purposes, single-molecule real-time sequencing (Pacific Biosciences) was used to sequence plasma DNA from 15 healthy subjects, 13 patients with chronic HBV infection (HBV carriers), and 45 patients with HCC. In one embodiment, the overall CGN / NCG motif ratio was calculated across healthy subjects, HBV, and HCC cases using short (≦200 bp) and long (≧1000 bp) plasma cfDNA molecules separately. In addition to 200 bp, other short cutoff values ​​can be used, such as 150 bp, 175 bp, 225 bp, and 250 bp. In addition to 100 bp, other long cutoff values ​​can be used, such as 6000 bp, 700 bp, 800 bp, 900 bp, 1,110 bp, 1,500 bp, and 2,000 bp.

[0381] Figures 70A-70C are boxplots showing the global CGN / NCG motif ratios of short cfDNA fragments (≦200 bp, 70A) and long cfDNA fragments (≧1000 bp, 70B), as well as the difference in CGN / NCG motif ratios between short cfDNA fragments (≦200 bp) and long cfDNA fragments (≧1000 bp) across healthy subjects (healthy), HBV carriers (HBV), and HCC cases (70C). Due to global hypomethylation in HCC, the global CGN / NCG motif ratios of HCC cases in short DNA fragments were significantly lower than those of healthy and HBV subjects (Figure 70A). For long DNA fragments, patients with HCC showed a higher global CGN / NCG motif ratio compared to healthy and HBV subjects (Figure 70B).

[0382] In some embodiments, a first amount (or ratio) of short DNA fragments containing CGN or NCG and a second amount (or ratio) of long DNA fragments containing CGN or NCG can be used together. The use of a separation value (e.g., difference or ratio) between the two ratios can provide increased accuracy, for example, because the changes are in opposite directions for short and long DNA fragments. As another example, the first amount and the second amount can be used as inputs (potentially four amounts of CGN and NCG for both short and long DNA fragments) to a machine learning model such as a decision tree, SVM, regression, or neural network.

[0383] Therefore, the motif ratios from cfDNA fragments with different lengths can be combined, but not limited to, by calculating the difference or ratio between the CGN / NCG motif ratios of short cfDNA fragments and long cfDNA fragments.As shown in Figure 70C, the difference in the CGN / NCG motif ratio between short cfDNA fragments and long cfDNA fragments is significantly lower in HCC subjects compared with non-HCC subjects.

[0384] Figures 71A-71B show the performance of HCC detection based on the CGN / NCG motif ratio together with size information. Figure 71A shows ROC analysis for distinguishing HCC from non-HCC (healthy and HBV subjects) using the overall CGN / NCG motif ratio of short cfDNA fragments (≦200bp) and long cfDNA fragments (≧1000bp), and the difference in the CGN / NCG motif ratio between short cfDNA fragments (≦200bp) and long cfDNA fragments (≧1000bp) (combined analysis). Figure 71B shows ROC analysis for distinguishing HCC from HBV using the overall CGN / NCG motif ratio of short cfDNA fragments (≦200bp) and long cfDNA fragments (≧1000bp), and the difference in the CGN / NCG motif ratio between short cfDNA fragments (≦200bp) and long cfDNA fragments (≧1000bp) (combined analysis).

[0385] The use of short DNA-derived CGN / NCG motif ratio measurement references resulted in an AUC of 0.89 for distinguishing HCC cases from non-HCC cases (healthy and HBV subjects, Figure 71A), and an AUC of 0.79 for HCC vs. HBV carriers (Figure 71B). The use of long DNA-derived CGN / NCG motif ratio measurement references resulted in an AUC of 0.72 for distinguishing HCC cases from non-HCC cases (healthy and HBV carriers, Figure 71A), and an AUC of 0.79 for HCC vs. HBV subjects (Figure 71B).

[0386] An AUC of 0.95 was achieved for distinguishing HCC cases from non-HCC cases (healthy and HBV carriers, Figure 71A), and an AUC of 0.90 for HCC vs. HBV subjects (Figure 71B). Such performance was better than the corresponding AUC values ​​using short or long cfDNA fragments alone. In another embodiment, short cfDNA fragments can be defined as cfDNA fragments having lengths less than, but not limited to, 50bp, 60bp, 70bp, 80bp, 90bp, 100bp, 200bp, 300bp, 400bp, 500bp, 600bp, 1000bp, etc., and long cfDNA fragments can be defined as cfDNA fragments having lengths greater than, but not limited to, 300bp, 400bp, 500bp, 600bp, 700bp, 800bp, 900bp, 1000bp, 2000bp, 3000bp, 4000bp, 5000bp, etc. In yet another embodiment, cfDNA fragments having different sizes can be obtained from an Illumina sequencing platform, long-read sequencing technology, or other platforms. Mass can also be used as a measure of size instead of length.

[0387] 6. Methods using CGN and / or NCG terminal motifs Pathology can be determined in a variety of ways. For example, the amount of DNA fragments terminating in CpG sites (e.g., CGN terminal motifs, or a subset thereof) in a new sample. The amount can be determined using a group of DNA fragments within a particular size range. Classification can be determined by comparing a single amount to a threshold, cutoff, or reference value, or using various models, such as machine learning models. Such models can be trained (e.g., via using different techniques) using reference / calibration samples with known or measured classifications. As with training other models, different accuracy goals can be selected, e.g., the desired tradeoff between sensitivity and specificity. Such training can be performed for other techniques described herein.

[0388] 72 is a flowchart illustrating a method 7200 for determining a level of pathology according to an embodiment of the present disclosure. A biological sample containing cell-free DNA. As with other methods described herein, method 7200 can be implemented in part or in whole using a computer system. Blocks performing similar functions to other methods can be implemented similarly to those described for those blocks in other flowcharts.

[0389] At block 7210, method 7200 includes analyzing a plurality of cell-free DNA molecules from the biological sample to obtain sequence reads. The sequence reads may include end sequences corresponding to the ends of the plurality of cell-free DNA molecules. Block 7210 may be performed similarly to block 2710 and other analysis steps described herein. In some embodiments, analyzing the cell-free DNA molecules may include determining a location of the cell-free DNA molecules in a reference genome, for example, when the analysis is of a region smaller than the entire genome. Such a region may be a chromosome. The plurality of cell-free DNA molecules may be located in a region of the reference genome. The location may be determined by various methods described herein, such as aligning the sequence reads to the reference genome or using position-based probes.

[0390] In block 7220, method 7200 includes, for each of the plurality of cell-free DNA molecules, determining a sequence motif for each of one or more ends of the cell-free DNA molecule. Block 7220 can be implemented via any of the techniques described herein for determining end motifs, for example, via sequencing or probe-based techniques including PCR. Block 7220 can be implemented similarly to block 4120. The end sequence motif can be determined from the outermost bases in the sequence read. If the sequence read includes only one end rather than the entire DNA fragment / molecule (e.g., as one of the paired reads covering both ends), the bases corresponding to the end of the DNA fragment are used, and bases toward the center of the fragment are not used. The ends of the cell-free DNA molecule can have a first position at the outermost position, a second position adjacent to the first position, and a third position adjacent to the second position.

[0391] At block 7230, method 7200 includes determining a first amount of a first set of one or more end sequence motifs having a C at a first position and a G at a second position (e.g., the amount of CGN or any one of these motifs). For example, the first set of one or more end sequence motifs can include one or more of CGA, CGC, CGG, and CGT, where CGN represents any nucleotide at the third position. The first amount can be normalized in various ways.

[0392] In other embodiments, the first amount may be of a first set of one or more terminal sequence motifs (e.g., NCG) having a C at the second position and a G at the third position. In this example, the first set of one or more terminal sequence motifs may include one or more of ACG, CCG, GCG, and TCG, with NCG representing all. Such exemplary terminal motifs may be used in various methods described herein to determine amounts and / or normalized parameters for use in performing classification. Block 7230 may be performed similarly to block 2730 and similar steps described herein.

[0393] Other amounts can be determined and used, for example, other terminal motifs, such as NCG to determine the CGN / NCG ratio. Thus, when the first amount uses CGN, a second amount of a second set of terminal sequence motifs having C at the second position and G at the third position corresponding to NCG or any subset thereof can be determined. Then, classification can include normalizing the first amount with the second amount to obtain a normalized first amount that is compared with a reference value. As another example, the total amount of sequence motifs (e.g., in one or more specific regions) can be determined and used for normalization. The normalized first amount (e.g., motif ratio or relative frequency) can be compared again with the reference value.

[0394] As another example of normalization, the total amount of a sequence motif in a plurality of cell-free DNA molecules (e.g., in a region) can be determined. Then, determining the classification can include normalizing the first amount by the total amount to obtain a normalized first amount (e.g., motif ratio or relative frequency) that is compared to a reference value.

[0395] At block 7240, method 7200 includes determining a classification of the level of pathology of the subject based on a comparison of the first amount to a reference value. The reference value can be determined as part of training a model, for example, using a reference sample with a known classification. The comparison to the reference value can be performed in various ways, for example, by direct comparison to the reference value, where values ​​below or above indicate a different classification. As another example, the comparison can be performed via a machine learning model.

[0396] As described above, the plurality of cell-free DNA molecules may be derived from one or more specific regions of the reference genome. Pathologies may be present in these one or more specific regions. For example, at least one of the one or more specific regions corresponds to a particular genotype or haplotype containing one or more alleles. If a subject has a certain amount of fragmentation of a particular terminal motif (e.g., CGN and NCG) in the genotype or haplotype, such a characteristic may indicate a level of pathology resulting from a disease (e.g., cancer or inflammation), an abnormal state / disorder of a physiological process, or an abnormal state / disorder associated with a physiological process (e.g., aging, or a developmental process, or via genomic imprinting, as described in more detail elsewhere in this disclosure). Copy number abnormalities may be such pathologies. Embodiments may also be used to determine the normal level of a physiological process, such as monitoring the stage of pregnancy, the aging of an individual, or the type or level of genomic imprinting of an individual.

[0397] Similar to other techniques described herein, a feature vector of multiple quantities can be provided to a machine learning model, for example, as described in the sections above. The different quantities can correspond to the selection of terminal motifs, e.g., those included in the first set and additional terminal motifs. Such terminal motifs included in the feature vector can be all of a particular k or only a subset of k-mers. Thus, the respective quantities of each trimeric terminal sequence motif having a C in the first position and a G in the second position (i.e., four quantities) can be determined. A feature vector including each quantity, including the first quantity, can be generated and then input into a machine learning model as part of determining a classification of the subject's level of pathology (e.g., the level of abnormality in a physiological process). Additionally or alternatively, the feature vector can include the respective quantities of each trimeric terminal sequence motif having a C in the second position and a G in the third position.

[0398] The analysis of the cell-free DNA molecules can include measuring the size of the cell-free DNA molecules. A first amount can then be determined for a first group of cell-free DNA molecules that fall within a first size range, for example, as described in Section 5. By way of example, the first size range can correspond to sizes that are less than a size cutoff or sizes that are greater than the first size cutoff.

[0399] A second amount can be determined for a second group of cell-free DNA molecules within a second, different size range, for example, corresponding to a size less than a second size cutoff when the first range is greater than the first size cutoff. The second amount can be one or more terminal motifs of CGN or NCG, as described above. Classification can then be determined using the first amount and the second amount. When CGN is used for long and short fragments, for example, a third amount of NCG long DNA fragments and a fourth amount of NCG short DNA fragments, defined using a first size cutoff and a second size cutoff, can be used. For example, classification can use the ratio of the first amount and the second amount, the difference between the first amount and the second amount, or a machine learning model that receives the first amount and the second amount as separate inputs.

[0400] B. Tissue-specific hypo- or hypermethylated regions / sites In addition to analyzing CGNs and NCGs generally (e.g., globally or across a large portion of the genome), specific regions can be analyzed. For example, hypermethylated or hypomethylated regions for a specific tissue can be analyzed. Such analysis can determine whether a specific type of cancer (e.g., tissue origin / source) is present. By way of example, the size of the region can be at least 2 bp, 10 bp, 100 bp, 200 bp, 500 bp, 1 kb, 2 kb, 3 kb, 4 kb, 5 kb, 10 kb, 50 kb, 100 kb, 500 kb, 1 Mb, etc. DNA fragments from specific regions / sites can be obtained through targeted techniques (also referred to as enrichment), which can be performed by amplifying a specific region or capturing DNA from a specific region. Further details about enrichment can be found in other sections of this disclosure.

[0401] Samples from HCC patients were analyzed. Differentially methylated regions between the liver and neutrophils (the most common white blood cells), representing the DNA background of blood cells, were identified. Each region was 500 bp in length. In this way, liver-specific hypermethylated and liver-specific hypomethylated regions were obtained compared to neutrophils. Differentially methylated regions may contain one or more CpG sites depending on the size of the region.

[0402] For liver hypermethylated regions where enzyme (e.g., DNASE1L3) activity is comparable between disease and control groups, the amount of CGN terminal motifs should increase and the amount of NCG terminal motifs should decrease. Therefore, the CGN / NCG ratio should increase. However, if enzyme activity is downregulated, such a ratio may decrease. For liver hypomethylated regions, the amount of CGN should decrease and the amount of NCG should increase. Therefore, the CGN / NCG ratio should decrease.

[0403] When enzyme (e.g., DNASE1L3) activity was comparable between disease and control groups, HCC subjects had higher hepatocyte contributions to plasma than healthy subjects. From hypermethylated liver regions, we expected an increase in CGN and a decrease in NCG compared to healthy subjects. From hypomethylated liver regions, we expected a decrease in CGN and an increase in NCG. This causes an increase in the CGN / NCG ratio in hypermethylated regions, but a decrease in the CGN / NCG ratio in hypomethylated regions.

[0404] Additionally, cleavage profiles or other methylation-related cfDNA fragmentomic features from liver- or HCC-specific hypomethylated regions can be used to detect HCC. And methylation-related cfDNA fragmentomic features from other tissue-specific hypomethylated or hypermethylated regions can be used to detect other cancers. For example, methylation-related cfDNA fragmentomic features from colon-, breast-, or prostate-specific hypomethylated or hypermethylated regions can be used to diagnose colon cancer, breast cancer, or prostate cancer, respectively. Such methylation-related cfDNA fragmentomic features from specific regions can be inferred from targeted sequencing results. Targeted sequencing can be based on, but is not limited to, probe-based hybridization, amplicon-based sequencing, immunoprecipitation followed by sequencing, etc.

[0405] 1. CGN and / or NCG Terminal motif frequencies from tissue-specific hypomethylation or hypermethylation are used to detect and monitor diseases such as cancer. Eight healthy controls, 17 patients with chronic hepatitis B (HBV), and 34 patients with hepatocellular carcinoma (HCC) with a median of 483 million paired-end sequencing reads (IQR: 395-549 million) were analyzed.

[0406] Figures 73A and 73B show the performance of HCC detection based on CGN / NCG motif ratios from liver-specific hypermethylated and hypomethylated regions according to embodiments of the present disclosure. Figure 73A is a boxplot showing CGN / NCG motif ratios from liver-specific hypermethylated and hypomethylated regions across healthy controls (CTRs), patients with chronic hepatitis B (HBV), and patients with HCC (HCC). Figure 73B is a ROC analysis of the performance of CGN / NCG motif ratios from liver-specific hypomethylated and hypermethylated regions, respectively, in distinguishing patients with and without HCC.

[0407] When comparing non-HCC subjects with HCC subjects, a significantly higher CGN / NCG motif ratio (i.e., CGN motif frequency / NCG motif frequency) was observed in liver-specific hypomethylated regions (P < 0.001, Kruskal-Wallis test), whereas no significant difference was observed in liver-specific hypermethylated regions. The CGN / NCG motif ratio from liver-specific hypomethylated regions may serve as a biomarker for HCC, enabling the detection of HCC patients with an area under the curve (AUC) of 0.85 (P = 0.003, DeLong test).

[0408] These results indicate that the CGN / NCG motif ratio associated with liver-specific hypomethylated CpG sites exhibits better discrimination ability between HCC patients and non-HCC individuals than liver-specific hypermethylated CpG sites. Such enhanced cancer detection is due to the selective analysis of CGN / NCG motif ratios from hypomethylated regions, which may result in the synergistic effects of DNA methylation and downregulated DNASE activity on characteristic cfDNA breaks (e.g., both epigenetic abnormalities and abnormal DNASE activity result in a CGN / NCG motif ratio).

[0409] 2. Use of vectors for each trimer or higher motif Because liver-specific hypomethylated region may contain more information, all trimer motifs in this region are further used to train SVM model.Therefore, feature vector has 64 values ​​of each amount of trimer that occurs in hypomethylated region.Other machine learning models can be used, for example, as described herein.

[0410] Figures 74A and 74B show the performance of HCC diagnosis using SVM based on trimer end motif frequency from liver-specific hypomethylated regions according to an embodiment of the present disclosure. Figure 74A is a boxplot showing the probability of determining HCC based on the trimer end motif frequency of plasma DNA molecules associated with liver-specific hypomethylated regions (CTR: healthy controls, HBV: patients with chronic hepatitis B, HCC: patients with HCC). Figure 74B is an ROC analysis of the use of trimer end motif frequency from liver-specific hypomethylated regions to distinguish between patients with and without HCC. The trimer end motif frequency can be determined as the amount of a particular end motif divided by the total number of DNA fragments or end motifs detected or used.

[0411] To further maximize the utility of terminal motifs from liver-specific hypomethylated regions, we constructed an SVM model using 5' trimer terminal motifs from liver-specific hypomethylated regions and performed HCC detection. The model performance could reach an AUC of approximately 0.91 for distinguishing HCC from non-HCC.

[0412] Therefore, the frequency of terminal motifs from tissue-specific hypomethylated or hypermethylated regions can be used to detect and monitor diseases such as cancer. For example, an SVM model was constructed using all 5' trimer terminal motifs from liver-specific hypomethylated regions to detect HCC. However, fewer than all 64 trimer terminal motifs can be used. For example, only eight motifs containing CG can be used. These eight trimer motifs provide increased accuracy compared to using all 64 trimer terminal motifs. These eight motifs are specifically CGA, CGT, CGC, CGC, ACG, TCG, CCG, and GCG. Other subsets of terminal motifs, such as a subset that does not contain CG, can also be used.

[0413] Instead of using all 64 trimer terminal motifs, we further tested the feasibility of detecting HCC by using CG-containing trimer terminal motifs (a total of eight trimer terminal motifs).Based on the bisulfite sequencing results of buffy coat and HCC tissue, we identified 116,245 and 802,393 HCC-specific hypermethylated and hypomethylated CpGs, respectively.

[0414] In various embodiments, any one or more of the CGN and / or NCG amounts may be used, which may be normalized, for example, as frequency. The frequency may be determined in various ways. For example, normalization may be performed as if all 64 terminal motifs were used. As another example, normalization may be performed within a subset of terminal motifs, for example, by the number of terminal motifs within the subset. In one embodiment, an NCG-related data vector, a CGN-related data vector, and a composite data vector (i.e., a data vector of CG-containing terminal motifs) were used, which may be obtained from HCC-specific hypermethylated and hypomethylated CpGs for 55 non-HCC cases and 34 HCC cases, respectively.

[0415] For the NCG-related data vector, only the NCG motif frequency was used. For example, the frequency of ACG was calculated as the amount of ACG motifs divided by the total amount of NCG motifs. The sum of the frequencies of ACG, TCG, CCG, and GCG is 100%. Similar calculations were applied to the CGN-related data vector. For example, the frequency of CGA was calculated as the number of CGA motifs divided by the total amount of CGN motifs. The sum of the frequencies of CGA, CGT, CGC, and CGG is 100%. The composite data vector was calculated for all CGN and NCG motifs. For example, the frequency of CGA was calculated as the number of CGA motifs divided by the total amount of CGN and NCG motifs. The sum of the frequencies of ACG, TCG, CCG, GCG, CGA, CGT, CGC, and CGG is 100%. These datasets were each used to build support vector machine (SVM) models.

[0416] Figures 75A-75B show the performance of HCC detection based on CG-containing trimer terminal motifs from HCC-specific hypermethylated and hypomethylated CpGs. As shown in Figure 75A, SVM models trained on NCG-associated data vectors, CGN-associated data vectors, and composite data vectors from HCC-specific hypermethylated CpGs achieved AUCs of 0.85, 0.89, and 0.98, respectively, for distinguishing HCC patients from non-HCC cases. As shown in Figure 75B, for HCC-specific hypomethylated CpGs, the models achieved AUCs of 0.94, 0.90, and 0.98, respectively. The NCG-associated data vectors and composite data vectors performed better than the accuracy of Figure 74B.

[0417] Some embodiments can combine feature vectors from hypermethylated and hypomethylated sites for use in training a machine learning model, for example, an SVM.

[0418] Figure 76 shows the ROC analysis for distinguishing patients with and without HCC using a composite data vector from HCC-specific hypermethylated and hypomethylated CpGs. To further maximize the utility of terminal motifs from HCC-specific methylated CpGs, the composite vectors from HCC-specific hypermethylated and hypomethylated CpGs were combined and used to train an SVM model that achieved an AUC of 0.99 for distinguishing HCC patients from non-HCC cases.

[0419] 3. Using cutting profiles We also investigated whether cleavage profiles from tissue-specific hypermethylated and hypomethylated CpG sites can be used to detect subjects with pathological or physiological conditions. As explained above, the cleavage profile includes two or more positions, for example, position 0, which contains the terminal motif of CGN.

[0420] a) Example workflow For illustrative purposes, pregnancy was used as a model. Pregnancy can be viewed as, or resembles, a pathological or physiological state in that a different tissue type with different properties than the host is growing. For example, both fetuses and tumors are globally hypermethylated, with certain sites in tumor suppressor genes being hypermethylated, for example. Thus, if a subject can be correctly classified as carrying a fetus, and such tissues have specific methylation patterns, embodiments can also be used to detect pathologies associated with other tissue types using hypo- or hypermethylated CpG sites. Another analysis can be a group of CpG sites (e.g., CpG islands). A region can be or include one or more such CpG sites.

[0421] Figure 77 shows exemplary marker selection and data collection according to an embodiment of the present disclosure. Placenta-specific hypermethylated sites were identified, defined by CpG sites with a methylation density greater than 70% in placental tissue but less than 30% in buffy coat samples. In other embodiments, these methylation density thresholds may be modified to achieve different diagnostic sensitivity and / or specificity, as described in various embodiments of the present disclosure. Hypomethylation may be defined as, but is not limited to, less than 10%, less than 20%, less than 30%, less than 40%, less than 50%, etc. Hypermethylation may be defined as, but is not limited to, greater than 40%, greater than 50%, greater than 60%, greater than 70%, greater than 80%, greater than 90%, etc. Thresholds may be in ranges, for example, but are not limited to, 20%-30%, 30%-40%, 40%-50%, 60%-70%, 70%-80%, etc. Placenta-specific hypomethylated sites were defined by CpG sites with a methylation density below 30% in placental tissue but above 70% in buffy coat samples.

[0422] For data collection, cleavage profiles were determined from placenta-specific hypermethylated and hypomethylated CpG sites. The cleavage profiles were determined using the techniques described herein. Such cleavage profiles were determined from 30 pregnancy samples (median paired-end sequencing reads: 26 million (IQR: 142-232 million)) and 9 non-pregnant healthy female cases (median paired-end sequencing reads: 397 million (IQR: 350-424 million)). In this exemplary embodiment, cleavage profiles can be generated by merging all data for a particular type of site, for example, merging data for all hypomethylated sites for one cleavage profile and all hypermethylated sites for another cleavage profile.

[0423] Figure 78 shows a workflow that involves generating cleavage profiles for various subjects to train a machine learning model, according to an embodiment of the present disclosure. In the example shown with measurement windows and using hypermethylated and hypomethylated sites, there are a total of 24 features. The machine learning model can then distinguish subjects based on the features. Such results are provided below.

[0424] b) Clustering the results In the results below, clustering was performed using principal component analysis (PCA). Other embodiments may use other types of machine learning, such as supervised or unsupervised, or a combination thereof, as described elsewhere in this disclosure.

[0425] Figures 79A, 79B, and 80 show PCA analysis of pregnant and non-pregnant subjects by using cleavage profiles from placenta-specific hypermethylated sites (Figure 79A), placenta-specific hypomethylated sites (Figure 79B), or a combination of placenta-specific hypermethylated and hypomethylated sites (Figure 80), according to embodiments of the present disclosure. PCA analysis identifies the linear combination of features that provides the best separation between the two types of subjects. Support vector machines can provide similar results. PCA allows visualization along the two main dimensions identified.

[0426] As shown in Figure 79A, principal components (PCs) (i.e., PC1 and PC2) based on cleavage profiles from placenta-specific hypermethylated sites could classify pregnant and non-pregnant subjects into two clusters, with only two non-pregnant subjects overlapping with pregnant subjects. Principal components (i.e., PC1 and PC2) based on cleavage profiles derived from placenta-specific hypomethylated sites showed better discrimination between pregnant and non-pregnant cases, with no overlap between the two groups (Figure 79B).

[0427] Furthermore, PCs based on cleavage profiles from both placenta-specific hypermethylated and hypomethylated sites (i.e., PC1 and PC2) showed complete separation between pregnant and non-pregnant cases (Figure 80). The data suggested that selective analysis of cleavage profiles derived from hypomethylated and / or hypermethylated CpG sites could improve classification power for classifying different pathological and physiological conditions.

[0428] c) Illustrative use case Cleavage profiles from tissue-specific hypomethylated or hypermethylated CpG sites are useful for detecting various cancers or other diseases. For example, cleavage profiles from colon-, breast-, or prostate-specific hypomethylated or hypermethylated CpG sites can be used to diagnose colon cancer, breast cancer, or prostate cancer, respectively. Therefore, cleavage profiles are useful for the prognosis and / or monitoring of cancer patients.

[0429] For example, proximal cleavage profiles from tissue-specific CpGs can be useful for inferring DNA contributions from different tissues to plasma cfDNA pools, such as tumor DNA fractions in HCC patients, which can be used for HCC diagnosis. Cleavage profiles were determined from HCC-specific hypermethylated and hypomethylated CpG sites. These cleavage profiles were used to train an SVM model to perform HCC detection.

[0430] Figure 81 shows the performance of HCC detection based on cleavage profiles. The performance of the model can reach AUCs of 0.915, 0.939, and 0.947 for distinguishing HCC patients from non-HCC controls when trained with cleavage profiles from HCC-specific hypermethylated or hypomethylated, or a combination of hypermethylated and hypomethylated CpGs.

[0431] d) Cleavage profiles of individual regions / sites The above examples combine measurements across regions (e.g., to determine the CGN / NCG ratio) or across CpG sites (e.g., to determine the cleavage ratio). However, in other embodiments, separate amounts can be determined for each region or CpG site (e.g., CGN and / or NCG amounts) and compared to their respective reference values ​​separately. Thus, if N regions or CpG sites are analyzed, N amounts can be determined, each corresponding to the amount of DNA fragments in the respective region or CpG site. These N amounts can be a feature vector input into a machine learning model. Additionally or alternatively, separate cleavage profiles can be determined for each CpG site (or a subset of sites), and these individual cleavage sites can be used, for example, as input into a machine learning model. Each cleavage profile can be input into a machine learning model to determine the level of pathology.

[0432] In another example, all of the cleava...

Claims

1. 1. A method for measuring methylation of CpG sites in a subject's genome using cell-free DNA molecules, comprising: analyzing a plurality of cell-free DNA molecules from a biological sample of the subject, the analyzing the cell-free DNA molecules comprising: determining a genomic location in a reference genome corresponding to at least one end of the cell-free DNA molecule; determining a first amount of cell-free DNA molecules that terminate at a first position within a window around the CpG site, the first position being between position -1 and position +1 of the window; and using the first amount to determine a methylation class of the CpG sites in the genome of the subject.

2. 2. The method of claim 1, further comprising determining a second amount of cell-free DNA molecules that terminate at a second position within the window around the CpG site, the second position being different from the first position, and wherein the classification is determined using the first amount and the second amount.

3. 3. The method of claim 2, wherein the first position is position 0 and the second position is +1 or -1 from the CpG site.

4. The method of claim 2, wherein the window is at least -2 to +2 from the CpG site.

5. determining said classification, determining a separation value using the first amount and the second amount; and and comparing the separation value to a calibration value, the calibration value being determined using cell-free DNA molecules derived from one or more calibration samples and located at CpG sites with known classification.

6. The method of claim 5 , wherein the classification is a quantitative value and comparing the separation value to the calibration value comprises comparing the separation value to a calibration function.

7. The method of claim 6 , wherein the quantitative value is in the range of 30% or less.

8. the CpG site is a first CpG site, the classification is a first classification, the window includes a second CpG site and at least two positions other than the first CpG site and the second CpG site, and the method further comprises: determining a third amount of cell-free DNA molecules terminating at the second CpG site; and For each position of the at least two positions within the window, determining a respective amount of cell-free DNA molecules terminating at said location, thereby determining a respective amount including said second amount, said at least two locations including said first location; generating a feature vector including the respective quantities, the first quantity, and the third quantity; 3. The method of claim 2, further comprising, as part of determining the classification of the first CpG site and determining a second classification of the second CpG site, inputting the feature vector into a machine learning model, the machine learning model being trained using cell-free DNA molecules that are located within a window around a CpG site with a known classification.

9. the window includes at least two positions other than the first position, and the method further comprises: For each position of the at least two positions within the window, determining an amount of each of the cell-free DNA molecules that terminate at said positions, said at least two positions including said first position; 2. The method of claim 1, further comprising, as part of determining the classification, comparing a first amount of the cell-free DNA molecules that terminate at the CpG site to a respective amount of the cell-free DNA molecules that terminate at the position.

10. the window includes at least two positions other than the first position, and the method further comprises: For each position of the at least two positions within the window, determining the amount of each of the cell-free DNA molecules that terminate at said position, thereby determining the amount of each of the cell-free DNA molecules that terminate at said position; generating a feature vector including the respective quantities and the first quantity; 2. The method of claim 1, further comprising, as part of determining the classification, inputting the feature vector into a machine learning model, the machine learning model being trained using cell-free DNA molecules that are located within a window around a CpG site with a known classification.

11. 11. The method of claim 10, wherein the feature vector forms a matrix with each row corresponding to a base of a strand, and columns containing non-zero quantities in the row corresponding to the base at each position.

12. The method of claim 10 , wherein the machine learning model is a convolutional neural network.

13. 11. The method of claim 10, wherein the first position and the at least two positions include all positions within the window, the window being at least +4 to -4 from the CpG site.

14. The method of claim 10 , wherein the machine learning model uses sequence context within the window.

15. The method of claim 14 , wherein the machine learning model is trained on the sequence context within the window.

16. The method of claim 14 , wherein the feature vector includes the sequence context.

17. 2. The method of claim 1, wherein the methylation classification indicates that the CpG site is in a hypermethylated or hypomethylated state for the cell-free DNA molecule at the CpG site, the hypermethylated state indicating a methylation density above a first threshold of at least 70% and the hypomethylated state indicating a methylation density below a second threshold of 30% or less.

18. The method of claim 1 , wherein the first amount is normalized.

19. The method of claim 18, wherein the normalization uses the number of cell-free DNA molecules that terminate within a region that contains the CpG site.

20. The method of claim 18, wherein the normalization uses the number of cell-free DNA molecules that cover the CpG site.

21. The method of claim 18, wherein the normalization uses the average or median depth of cell-free DNA molecules in the region containing the CpG site.

22. 3. The method of claim 1 or 2, further comprising using the first amount and optionally the second amount to determine another classification of methylation of different CpG sites in the genome of the subject, the different CpG sites being within 600 nt downstream from the 5' end of the CpG site.

23. 1. A method for measuring methylation of a region in a genome of a subject, comprising: analyzing a plurality of cell-free DNA molecules from a biological sample of the subject, the analyzing the cell-free DNA molecules comprising: determining a terminal sequence motif of at least one end of the cell-free DNA molecules, wherein the end of the cell-free DNA molecule has a first position at an outermost position, a second position adjacent to the first position, and a third position adjacent to the second position, and the plurality of cell-free DNA molecules are located in the region of a reference genome; determining a first amount of a first set of one or more end sequence motifs, the first set of one or more terminal sequence motifs has a C at the first position and a G at the second position; or determining that a first set of one or more terminal sequence motifs has a C at the second position and a G at the third position; determining a methylation classification of said region in the genome of said subject by comparing said first amount to a calibration value, said calibration value being determined using cell-free DNA molecules derived from one or more calibration samples and located at CpG sites with known classification.

24. 24. The method of claim 23, wherein the region is a portion of the genome of the subject, and the plurality of cell-free DNA molecules are derived from one or more specific regions of the reference genome.

25. the region comprises a sequence variant or copy number abnormality determined from the plurality of cell-free DNA molecules, the method comprising:

25. The method of claim 24, further comprising determining whether the sequence variant or the copy number abnormality is derived from a tumor based on the methylation classification of the region.

26. 25. The method of claim 24, wherein at least one of the one or more particular regions corresponds to a particular genotype or haplotype comprising one or more alleles.

27. the methylation classification of the region is a first methylation level corresponding to a first haplotype in a woman carrying a fetus, the method comprising: determining a second methylation level corresponding to a second haplotype of the female using a second amount of the first set of one or more terminal sequence motifs for cell-free DNA molecules from the second haplotype; 27. The method of claim 26, further comprising determining a genetic haplotype of the fetus based on the first methylation level and the second methylation level.

28. 27. The method of claim 26, further comprising determining a level of pathology of the subject in the one or more specified regions based on the classification.

29. 29. The method of claim 28, wherein the level of the pathology comprises an abnormal level of a physiological process or a copy number abnormality.

30. a first set of one or more terminal sequence motifs having a C at the first position and a G at the second position, said method comprising: determining a second amount of a second set of one or more end sequence motifs having a C at the second position and a G at the third position; 24. The method of claim 23, wherein determining the classification comprises normalizing the first amount by the second amount to obtain a normalized first amount that is compared to the calibration value.

31. 31. The method of claim 30, wherein the first set of one or more terminal sequence motifs comprises CGA, CGC, CGG, and CGT, and the second set of one or more terminal sequence motifs comprises ACG, CCG, GCG, and TCG.

32. determining an expected amount of the first set of one or more end sequence motifs based on a reference sequence of the reference genome in the region; 24. The method of claim 23, wherein determining the classification comprises normalizing the first amount with the expected amount to obtain a normalized first amount that is compared to the calibration value.

33. determining a total amount of sequence motifs for the plurality of cell-free DNA molecules; 24. The method of claim 23, wherein determining the classification comprises normalizing the first amount by the total amount to obtain a normalized first amount that is compared to the calibration value.

34. 24. The method of claim 23, wherein the region is the reference genome.

35. 24. The method of claim 23, wherein the region is 10 Mb or less of a particular chromosome.

36. a first set of one or more terminal sequence motifs having a C at the first position and a G at the second position, said method comprising: determining the amount of each trimer end sequence motif having a C at the first position and a G at the second position, thereby determining the amount of each; generating a feature vector including each of the quantities, including the first quantity; 24. The method of claim 23, further comprising, as part of determining the methylation classification of the region in the genome of the subject, inputting the feature vector into a machine learning model, wherein the machine learning model is trained using cell-free DNA molecules derived from the one or more calibration samples.

37. 37. The method of claim 36, wherein the feature vector comprises the abundance of each of all trimer end sequence motifs.

38. Analyzing cell-free DNA molecules 24. The method of claim 23, comprising determining the location of the cell-free DNA molecule in the reference genome.

39. 1. A method for measuring methylation of CpG sites in test cell-free DNA molecules from a biological sample of a subject, comprising: receiving one or more sequence reads for the test cell-free DNA molecule, the test cell-free DNA molecule including one or more CpG sites; determining a sequence context of the test cell-free DNA molecule using the one or more sequence reads, the sequence context comprising sequence information from each nucleotide of the test cell-free DNA molecule; determining a first end sequence motif at a first end of the test cell-free DNA molecule using the one or more sequence reads; loading into a memory of a computer system a machine learning model that is trained using a training set of training cell-free DNA molecules having CpG sites with known methylation status, the training set including sequence context and end sequence motifs of the training cell-free DNA molecules; determining a methylation level of the one or more CpG sites of the test cell-free DNA molecule using the machine learning model, wherein (1) the sequence context of the test cell-free DNA molecule and (2) the first terminal sequence motif are input into the machine learning model.

40. 40. The method of claim 39, further comprising determining a second end sequence motif at a second end of the test cell-free DNA molecule using the one or more sequence reads, wherein the second end sequence motif is further input into the machine learning model.

41. 40. The method of claim 39, wherein the first terminal sequence motif is one-hot encoded prior to inputting it into the machine learning model.

42. 40. The method of claim 39, wherein the sequence context comprises the entire sequence of the test cell-free cell DNA molecule.

43. 40. The method of claim 39, wherein the sequence context comprises a K-mer matrix that specifies the number of instances of each K-mer in the test cell-free DNA molecule, where K is an integer.

44. 44. The method of claim 43, wherein K is equal to 2 and the K-mer matrix is ​​a dinucleotide matrix that identifies the number of instances of dinucleotide pairs in the test cell-free DNA molecules.

45. 40. The method of claim 39, wherein the machine learning model comprises a convolutional neural network.

46. 40. The method of claim 39, wherein the test cell-free DNA molecule comprises a plurality of CpG sites and the methylation level is the percentage of the plurality of CpG sites that are methylated.

47. determining the size of the test cell-free DNA molecule; 40. The method of claim 39, further comprising inputting the size of the test cell-free DNA molecule into the machine learning model as part of determining the methylation level of the one or more CpG sites of the test cell-free DNA molecule.

48. 48. The method of claim 47, wherein the one or more sequence reads is a single sequence read that spans the test cell-free DNA molecule, and determining the size of the test cell-free DNA molecule comprises counting the number of bases in the single sequence read.

49. 48. The method of claim 47, wherein the one or more sequence reads include two sequence reads corresponding to opposite ends of the test cell-free DNA molecule, and determining the size of the test cell-free DNA molecule includes aligning the two sequence reads to a reference genome.

50. 1. A method of analyzing a biological sample to determine a level of pathology in said biological sample of a subject, said biological sample comprising cell-free DNA; analyzing a plurality of cell-free DNA molecules from the biological sample to obtain sequence reads, the sequence reads including terminal sequences corresponding to ends of the plurality of cell-free DNA molecules; determining, for each of the plurality of cell-free DNA molecules, a sequence motif for each of one or more ends of the cell-free DNA molecule, wherein the ends of the cell-free DNA molecule have a first position at an outermost position, a second position adjacent to the first position, and a third position adjacent to the second position; determining a first amount of a first set of one or more end sequence motifs of the plurality of cell-free DNA molecules, the first set of one or more terminal sequence motifs has a C at the first position and a G at the second position; or determining that a first set of one or more terminal sequence motifs has a C at the second position and a G at the third position; and determining a classification of the level of the pathology in the subject based on a comparison of the first amount to a reference value.

51. 51. The method of claim 50, wherein analyzing the plurality of cell-free DNA molecules comprises measuring a size of each of the plurality of cell-free DNA molecules, and the first amount is determined for a first group of cell-free DNA molecules that are within a first size range.

52. 52. The method of claim 51, wherein the first size range corresponds to a size that is less than a size cutoff.

53. 52. The method of claim 51, wherein the first size range corresponds to sizes greater than a first size cutoff.

54. determining a second amount of a second set of one or more terminal sequence motifs of the plurality of cell-free DNA molecules; a second set of one or more terminal sequence motifs having a C at the first position and a G at the second position; or a second set of one or more terminal sequence motifs having a C at the second position and a G at the third position; 54. The method of claim 53, wherein the second amount is determined for a second group of cell-free DNA molecules that are within a second size range corresponding to a size that is less than a second size cutoff, and the classification is determined using the first amount and the second amount.

55. the first amount is determined using a first set of one or more end sequence motifs having a C at the first position and a G at the second position, and the second amount is determined using a second set of one or more end sequence motifs having a C at the first position and a G at the second position, the method comprising: determining a third amount of the first group of cell-free DNA molecules having a third set of terminal sequence motifs having a C at the second position and a G at the third position; 55. The method of claim 54, further comprising determining a fourth amount of the second group of cell-free DNA molecules having a fourth set of terminal sequence motifs having a C at the second position and a G at the third position, wherein the classification is further determined using the third amount and the fourth amount.

56. 55. The method of claim 54, wherein the classification uses a ratio of the first amount and the second amount, uses a difference between the first amount and the second amount, or uses a machine learning model that receives the first amount and the second amount as separate inputs.

57. The method of claim 50, wherein the first amount is determined for a first group of cell-free DNA molecules each located within one or more regions that are each 5hmC enriched or each 5hmC depleted for a particular tissue type, and the pathology is of the particular tissue type.

58. 58. The method of claim 57, wherein the specific tissue type is HCC cancer tissue.

59. a first group of cell-free DNA molecules each located within one or more regions that are each enriched in 5hmC, the method comprising:

58. The method of claim 57, further comprising determining a second amount of the first set of one or more terminal sequence motifs of the plurality of cell-free DNA molecules, each located within one or more regions that are each 5hmC-depleted for the particular tissue type, wherein the classification is determined using the first amount and the second amount.

60. 51. The method of claim 50, wherein the first amount is determined for a first group of cell-free DNA molecules each located within one or more regions that are each hypermethylated or hypomethylated for a particular tissue type, and the pathology is of the particular tissue type.

61. the plurality of cell-free DNA molecules are derived from a first region of a plurality of regions of a reference genome, and the comparison of the first amount to the reference value indicates whether the first amount in the first region is abnormal, and the method further comprises: determining a number of the plurality of regions that are abnormal; and 51. The method of claim 50, further comprising comparing the number of abnormal regions to a threshold to determine the classification of the level of the pathology.

62. 62. The method of claim 61 , wherein for each of the plurality of regions, the amount of each of the first set of one or more end sequence motifs is compared to a region-specific reference value.

63. 51. The method of claim 50, wherein the plurality of cell-free DNA molecules is derived from one or more specific regions of a reference genome, and the pathology is in the one or more specific regions.

64. 64. The method of claim 63, wherein at least one of the one or more particular regions corresponds to a particular genotype or haplotype comprising one or more alleles.

65. 64. The method of claim 63, wherein the level of pathology comprises a copy number abnormality.

66. 51. The method of claim 50, wherein the first set of one or more terminal sequence motifs comprises CGN, wherein N comprises any nucleotide at the third position.

67. a first set of one or more terminal sequence motifs having a C at the first position and a G at the second position, said method comprising: determining a second amount of a second set of terminal sequence motifs having a C at the second position and a G at the third position; 51. The method of claim 50, wherein determining the classification comprises normalizing the first amount by the second amount to obtain a normalized first amount that is compared to the reference value.

68. 68. The method of claim 67, wherein the second set of terminal sequence motifs comprises NCG, where N comprises any nucleotide at the first position.

69. determining a total amount of sequence motifs for the plurality of cell-free DNA molecules; 51. The method of claim 50, wherein determining the classification comprises normalizing the first amount with the total amount to obtain a normalized first amount that is compared to the reference value.

70. 51. The method of claim 50, wherein the pathology is cancer.

71. 51. The method of claim 50, wherein the pathology is an abnormal level of a physiological process, the physiological process being pregnancy, aging, development, or growth.

72. a first set of one or more terminal sequence motifs having a C at the first position and a G at the second position, said method comprising: determining the amount of each trimer end sequence motif having a C at the first position and a G at the second position, thereby determining the amount of each; generating a feature vector including each of the quantities, including the first quantity; 51. The method of claim 50, further comprising, as part of determining the classification of the level of the pathology of the subject, inputting the feature vector into a machine learning model, the machine learning model being trained using cell-free DNA molecules in a training sample having known classifications.

73. 73. The method of claim 72, further comprising determining an additional amount of each of the trimer end sequence motifs having a C at the second position and a G at the third position, thereby determining each additional amount, wherein the feature vector comprises the each additional amount.

74. 73. The method of claim 72, wherein the feature vector comprises the amount of each of all trimer end sequence motifs.

75. wherein the first set of one or more terminal sequence motifs has a C at the second position and a G at the third position, and the method comprises: determining the amount of each trimer end sequence motif having a C at the second position and a G at the third position, thereby determining the amount of each; generating a feature vector including each of the quantities, including the first quantity; 51. The method of claim 50, further comprising, as part of determining the classification of the level of the pathology of the subject, inputting the feature vector into a machine learning model, the machine learning model being trained using cell-free DNA molecules in a training sample having known classifications.

76. 1. A method of analyzing a biological sample to determine a level of pathology in said biological sample of a subject, said biological sample comprising cell-free DNA; (a) analyzing a plurality of cell-free DNA molecules from the biological sample of the subject, wherein analyzing the cell-free DNA molecules comprises: determining locations of the cell-free DNA molecules in a reference genome, wherein the plurality of cell-free DNA molecules are located in a first set of one or more regions of the reference genome, and each of the first set of one or more regions is hypomethylated or each is hypermethylated in a first tissue type; determining a terminal sequence motif of at least one end of the cell-free DNA molecule, the end of the cell-free DNA molecule having a first position at an outermost position, a second position adjacent to the first position, and a third position adjacent to the second position; (b) determining a first amount of a first set of one or more end sequence motifs, the first set of one or more terminal sequence motifs has a C at the first position and a G at the second position; or determining that a first set of one or more terminal sequence motifs has a C at the second position and a G at the third position; (c) determining a classification of the level of the pathology in the first tissue type of the subject based on a comparison of the first amount to a reference value.

77. 77. The method of claim 76, wherein the first tissue type is a cancer tissue type.

78. 78. The method of claim 77, wherein the biological sample is urine and the cancer tissue type is selected from bladder cancer, kidney cancer, and prostate cancer.

79. 78. The method of claim 77, wherein the biological sample is plasma or serum and the cancer tissue type is selected from liver cancer, colon cancer, lung cancer, and breast cancer.

80. wherein said classification uses differentially methylated regions of a plurality of tissue types including said first tissue type, said method comprising:

77. The method of claim 76, further comprising analyzing another set of cell-free DNA molecules from the biological sample, wherein each set of cell-free DNA molecules maps to a respective set of one or more regions of the reference genome, and each of the respective sets of one or more regions is hypomethylated or hypermethylated in a respective tissue type of the plurality of tissue types.

81. 81. The method of claim 80, wherein the particular set of one or more regions comprises a plurality of regions of a plurality of region types, a first region type that is differentially methylated relative to other tissues of the plurality of tissue types, and a second region type that is differentially methylated relative to at least one other tissue of the plurality of tissue types, but not all of the other tissues.

82. 82. The method of claim 81 , wherein another specific set of one or more regions are all hypomethylated or all hypermethylated in the cancer tissue type of the first tissue and the healthy first tissue.

83. 81. The method of claim 80, wherein the classification of the level of pathology uses a first machine learning model, and the amounts of the first set of one or more terminal sequence motifs are used to train one or more other machine learning models using training samples having the pathology in one or more other tissue types of the plurality of tissue types.

84. 81. The method of claim 80, wherein the classification of the level of pathology uses multi-class machine learning that provides a probability of each of the multiple tissue types having the pathology.

85. 81. The method of claim 80, wherein the multiple tissue types include cancerous and non-cancerous tissue types.

86. identifying a second set of CpG sites that are all hypomethylated or all hypermethylated across multiple tissue types in the reference genome; determining a second amount of cell-free DNA molecules that terminate around any one of the second set of CpG sites; 77. The method of claim 76, wherein determining the classification of the level of the pathology comprises normalizing the first amount by the second amount to obtain a normalized first amount that is compared to the reference value.

87. said first set of one or more end sequence motifs defining a first set of one or more CpG sites within said first set of one or more regions, said method comprising: For each position of at least two positions within a window surrounding any one of the first set of one or more CpG sites, determining the amount of each of the cell-free DNA molecules that terminate at said position, thereby determining the amount of each of the cell-free DNA molecules that terminate at said position; generating a feature vector including the respective quantities and the first quantity; 77. The method of claim 76, further comprising, as part of determining the classification of the level of the pathology in the first tissue type of the subject, inputting the feature vector into a machine learning model, wherein the machine learning model is trained using cell-free DNA molecules from a training sample having a known classification.

88. the first set of one or more CpG sites are all hypomethylated for the first tissue type, the feature vector is a first feature vector, and the method comprises: generating a second feature vector using a second set of one or more CpG sites that are all hypermethylated; 88. The method of claim 87, further comprising inputting the second feature vector and the first feature vector into the machine learning model as part of determining the classification of the level of the pathology in the first tissue type of the subject.

89. said first set of one or more end sequence motifs defining a first set of CpG sites within said first set of one or more regions; and determining said first amount; 77. The method of claim 76, comprising determining respective amounts of cell-free DNA molecules terminating at each of the first set of CpG sites, and wherein determining the classification of the level of the pathology comprises comparing each of the respective amounts to a respective reference value.

90. 90. The method of claim 89, wherein comparing each of the respective amounts to the respective reference values ​​is performed by inputting the respective amounts as part of a feature vector into a machine learning model.

91. for each CpG site of the first set of one or more CpG sites, For each position of at least two positions within a window around the CpG site, determining the amount of each of the cell-free DNA molecules that terminate at said position, thereby determining the amount of each of the cell-free DNA molecules that terminate at said position; 91. The method of claim 90, further comprising including the respective amounts of the at least two positions and the respective amounts of the CpG sites in the feature vector.

92. 77. The method of claim 76, wherein the first set of one or more regions is a plurality of regions.

93. 77. The method of claim 76, wherein the first set of one or more terminal sequence motifs comprises CGN, wherein N comprises any nucleotide at the third position.

94. determining a second amount of a second set of terminal sequence motifs having a C at the second position and a G at the third position; 77. The method of claim 76, wherein determining the classification comprises normalizing the first amount by the second amount to obtain a normalized first amount that is compared to the reference value.

95. 95. The method of claim 94, wherein the second set of terminal sequence motifs comprises NCG, where N comprises any nucleotide at the first position.

96. determining a total amount of sequence motifs for the plurality of cell-free DNA molecules; 77. The method of claim 76, wherein determining the classification comprises normalizing the first amount with the total amount to obtain a normalized first amount that is compared to the reference value.

97. 77. The method of claim 76, wherein the pathology is cancer.

98. 77. The method of claim 76, wherein the pathology is an abnormal level of a physiological process, the physiological process being pregnancy, aging, development, or growth.

99. 77. The method of Claim 76, further comprising repeating (a), (b), and (c) for one or more additional sets of one or more regions of the reference genome, wherein the one or more additional sets are all hypomethylated or all hypermethylated in one or more other tissue types, thereby determining the cancer type of the subject.

100. wherein the first set of one or more terminal sequence motifs comprises all sequence motifs having a C at the first position and a G at the second position, and the method comprises: determining the amount of each trimer end sequence motif having a C at the first position and a G at the second position, thereby determining the amount of each; generating a feature vector including each of the quantities, including the first quantity; 77. The method of claim 76, further comprising, as part of determining the classification of the level of the pathology in the first tissue type of the subject, inputting the feature vector into a machine learning model, the machine learning model being trained using cell-free DNA molecules in a training sample having known classifications.

101. 101. The method of claim 100, further comprising determining an additional amount of each of the trimer end sequence motifs having a C at the second position and a G at the third position, thereby determining each additional amount, wherein the feature vector comprises the each additional amount.

102. The method of claim 100, wherein the feature vector comprises the respective amounts of all trimer end sequence motifs.

103. wherein the first set of one or more terminal sequence motifs comprises all sequence motifs having a C at the second position and a G at the third position, and the method comprises: determining the amount of each trimer end sequence motif having a C at the second position and a G at the third position, thereby determining the amount of each; generating a feature vector including each of the quantities, including the first quantity; 77. The method of claim 76, further comprising, as part of determining the classification of the level of the pathology of the subject, inputting the feature vector into a machine learning model, the machine learning model being trained using cell-free DNA molecules in a training sample having known classifications.

104. wherein said classification uses differentially methylated regions of a plurality of tissue types including said first tissue type, said method comprising: analyzing another set of cell-free DNA molecules from the biological sample, each set of cell-free DNA molecules mapping to a respective set of one or more regions of the reference genome, each of the respective sets of one or more regions being hypomethylated or hypermethylated in a respective tissue type of the plurality of tissue types; 104. The method of claim 100 or 103, wherein the feature vector forms a matrix with each row corresponding to one of a respective set of the one or more regions of the reference genome and columns corresponding to the respective abundance of each trimer end sequence motif of the first set, and wherein the machine learning model is a convolutional neural network.

105. 1. A method for determining a level of pathology in a subject from whom said biological sample was obtained by analyzing a biological sample comprising a mixture of cell-free DNA molecules, said mixture comprising cell-free DNA molecules derived from said subject and derived from a virus, (a) analyzing a plurality of cell-free DNA molecules from the biological sample of the subject, wherein analyzing each of a group of viral cell-free DNA molecules of the virus comprises: determining locations of the viral cell-free DNA molecules in a reference genome of the virus, wherein the group of viral cell-free DNA molecules are located at a first set of one or more CpG sites of the reference genome; determining a terminal sequence motif of at least one end of the viral cell-free DNA molecule, thereby determining a terminal sequence motif, wherein an end of the viral cell-free DNA molecule has a first position at an outermost position, a second position adjacent to the first position, and a third position adjacent to the second position; (b) using the end sequence motif to determine a first quantity of the population of viral cell-free DNA molecules of the virus that terminate at any one or more of positions −1 to +1 of any one of the first set of one or more CpG sites; (c) determining a classification of the level of the pathology in the subject based on a comparison of the first amount to a reference value.

106. 106. The method of claim 105, wherein each of the first set of one or more CpG sites has a methylation level within a specified range in a subject with the pathology.

107. 106. The method of claim 105, wherein each of the first set of one or more CpG sites is identified using a first training set of samples from subjects with the pathology and a second training set of samples from subjects without the pathology, and wherein the first amount at each of the first set of one or more CpG sites differs by at least a specified percentage between the first training set and the second training set.

108. 106. The method of claim 105, wherein each of the first set of one or more CpG sites is differentially methylated in subjects with the pathology relative to subjects without the pathology.

109. 106. The method of claim 105, wherein the first amount is of a first set of one or more terminal sequence motifs having a C at the first position and a G at the second position.

110. 110. The method of claim 109, wherein the first set of one or more terminal sequence motifs comprises CGN, wherein N comprises any nucleotide at the third position.

111. determining a second amount of a second set of terminal sequence motifs having a C at the second position and a G at the third position; 110. The method of claim 109, wherein determining the classification comprises normalizing the first amount by the second amount to obtain a normalized first amount that is compared to the reference value.

112. 112. The method of claim 111, wherein the second set of terminal sequence motifs comprises NCG, where N comprises any nucleotide at the first position.

113. 106. The method of claim 105, wherein the first amount is of a first set of one or more terminal sequence motifs having a C at the second position and a G at the third position.

114. 106. The method of claim 105, wherein the first amount is the sum of the viral cell-free DNA molecules that terminate at the -1 or +1 position of the first set of one or more CpG sites.

115. 115. The method of claim 114, wherein the first set of one or more CpG sites are each hypomethylated.

116. For each position of at least two positions within a window surrounding any one of the first set of one or more CpG sites, determining the amount of each of the cell-free DNA molecules that terminate at said position, thereby determining the amount of each of the cell-free DNA molecules that terminate at said position; generating a feature vector including the respective quantities and the first quantity; 106. The method of claim 105, further comprising, as part of determining the classification of the level of pathology, inputting the feature vector into a machine learning model, the machine learning model being trained using cell-free DNA molecules from a training sample having a known classification.

117. 117. The method of claim 116, wherein each of the first set of one or more CpG sites is hypomethylated in a subject with the pathology.

118. the feature vector is a first feature vector, generating a second feature vector using a second set of one or more CpG sites that are all hypermethylated; 118. The method of claim 117, further comprising inputting the second feature vector and the first feature vector into the machine learning model as part of determining the classification of the level of the pathology.

119. wherein the first set of one or more CpG sites is a first set of CpG sites, and determining the first amount determining a respective amount of cell-free DNA molecules that terminate at each of the first set of CpG sites; 106. The method of claim 105, wherein determining the classification of the level of the pathology comprises comparing each of the respective amounts to a respective reference value.

120. 120. The method of claim 119, wherein comparing each of the respective amounts to the respective reference values ​​is performed by inputting the respective amounts as part of a feature vector into a machine learning model.

121. for each CpG site of the first set of one or more CpG sites, For each position of at least two positions within a window around the CpG site, determining the amount of each of the cell-free DNA molecules that terminate at said position, thereby determining the amount of each of the cell-free DNA molecules that terminate at said position; and including the respective amounts of the at least two positions and the respective amounts of the CpG sites in the feature vector.

122. determining a total amount of sequence motifs for the plurality of viral cell-free DNA molecules; 106. The method of claim 105, wherein determining the classification comprises normalizing the first amount by the total amount to obtain a normalized first amount that is compared to the reference value.

123. 106. The method of claim 105, wherein the pathology is cancer.

124. determining the amount of each trimer end sequence motif having a C at the first position and a G at the second position, thereby determining the amount of each; generating a feature vector including each of the quantities, including the first quantity; 106. The method of claim 105, further comprising, as part of determining the classification of the level of the pathology, inputting the feature vector into a machine learning model, the machine learning model being trained using cell-free DNA molecules in a training sample having known classifications.

125. 125. The method of claim 124, further comprising determining an additional amount of each of the trimer end sequence motifs having a C at the second position and a G at the third position, thereby determining each additional amount, wherein the feature vector comprises the each additional amount.

126. 125. The method of claim 124, wherein the feature vector comprises the respective amounts of all trimer end sequence motifs.

127. Analyzing the plurality of viral cell-free DNA molecules includes determining a size of each of the plurality of viral cell-free DNA molecules, the method comprising: determining a size distribution statistics of the plurality of viral cell-free DNA molecules; 106. The method of claim 105, wherein determining the classification of the level of the pathology in the subject further comprises comparing the statistical value to a size reference value.

128. determining a quantity of said plurality of viral cell-free DNA molecules; 128. The method of claim 105 or 127, wherein determining the classification of the level of the pathology in the subject further comprises comparing the amount to a count reference value.

129. 1. A method for measuring the fractional concentration of DNA from a first tissue type in a biological sample from a subject, the biological sample comprising: analyzing a plurality of cell-free DNA molecules from the biological sample of the subject, wherein analyzing the cell-free DNA molecules comprises: determining a genomic location in a reference genome corresponding to at least one end of the cell-free DNA molecule; identifying a first set of CpG sites in the reference genome that all have a first differential methylation for the first tissue type, the first set of CpG sites including one or more CpG sites; determining a first amount of cell-free DNA molecules that terminate at one of the 0 or -1 positions of any one of the first set of CpG sites; determining a fractional concentration of the DNA from the first tissue type in the biological sample by comparing the first amount to a calibration value, the calibration value being determined from one or more calibration samples having known fractional concentrations of DNA from the first tissue type.

130. 130. The method of claim 129, wherein the first differential methylation is that all of the first set of CpG sites are hypomethylated or the first differential methylation is that all of the first set of CpG sites are hypermethylated.

131. identifying a second set of CpG sites that are all hypomethylated or all hypermethylated across multiple tissue types in the reference genome; determining a second amount of cell-free DNA molecules that terminate within a window around any one of the second set of CpG sites; 131. The method of claim 130, wherein determining the fractional concentration comprises normalizing the first amount by the second amount to obtain a normalized first amount that is compared to the calibration value.

132. identifying a second set of CpG sites that are all hypomethylated or all hypermethylated for a second tissue type in the reference genome, the second set of CpG sites comprising one or more CpG sites; determining a second amount of cell-free DNA molecules that terminate at any one of the second set of CpG sites; 131. The method of claim 130, further comprising determining a change in fractional concentration of DNA from the second tissue type in the biological sample by comparing the second amount to another calibration value, the other calibration value being determined from one or more other calibration samples.

133. 130. The method of claim 129, wherein the first differential methylation is that all of the first set of CpG sites are 5hmC enriched or the first differential methylation is that all of the first set of CpG sites are 5hmC depleted.

134. Determining the first amount comprises: determining a respective amount of cell-free DNA molecules that terminate at position 0 of each of the first set of CpG sites; 130. The method of claim 129, wherein determining the fractional concentration of the DNA from the first tissue type comprises comparing each of the respective amounts to a respective calibration value.

135. 135. The method of claim 134, wherein comparing each of the respective amounts to the respective calibration values ​​is performed by inputting the respective amounts as part of a feature vector into a machine learning model.

136. For each CpG site of the one or more CpG sites: For each position of at least two positions within a window around the CpG site, determining the amount of each of the cell-free DNA molecules that terminate at said position, thereby determining the amount of each of the cell-free DNA molecules that terminate at said position; and including the respective amounts of the at least two positions and the respective amounts of the CpG sites in the feature vector.

137. 130. The method of claim 129, wherein the first set of CpG sites is a plurality of CpG sites.

138. the first amount being of cell-free DNA molecules terminating at position 0, and the method comprising: determining a second amount of cell-free DNA molecules that terminate at the first position within a window around any one of the first set of CpG sites; 130. The method of claim 129, wherein determining the fractional concentration comprises normalizing the first amount by the second amount to obtain a normalized first amount that is compared to the calibration value.

139. 139. The method of claim 138, wherein the first position is at +1 or -1 from the CpG site.

140. 139. The method of claim 138, wherein the window is at least -2 to +2 from the CpG site.

141. For each position of the at least two positions within the window, determining an amount of each of the cell-free DNA molecules that terminate at said positions, said at least two positions including said first position; 139. The method of claim 138, further comprising, as part of determining the fractional concentration, comparing a first amount of the cell-free DNA molecules terminating at the CpG site to the amount of each of the cell-free DNA molecules terminating at the position.

142. For each position of the at least two positions within the window, determining a respective amount of cell-free DNA molecules terminating at said location, thereby determining a respective amount including said second amount, said at least two locations including said first location; generating a feature vector including the respective quantities and the first quantity; 139. The method of claim 138, further comprising inputting the feature vector into a machine learning model as part of determining the fractional concentration, the machine learning model being trained using cell-free DNA molecules from the one or more calibration samples having known fractional concentrations of DNA from the first tissue type.

143. the first set of one or more CpG sites are all hypomethylated for the first tissue type in the reference genome, the feature vector is a first feature vector, and the method comprises: generating a second feature vector using a second set of one or more CpG sites that are all hypermethylated; 143. The method of claim 142, further comprising inputting the second feature vector and the first feature vector into the machine learning model as part of determining the fractional concentration.

144. 143. The method of claim 142, wherein said at least two positions include all positions within said window, said window being at least +4 to -4 from said CpG site.

145. 143. The method of claim 142, wherein the machine learning model is support vector regression.

146. 130. The method of claim 129, wherein the first amount is normalized.

147. The method of claim 146, wherein the normalization uses the number of cell-free DNA molecules that terminate within a region containing a CpG site.

148. The method of claim 146, wherein the normalization uses the number of cell-free DNA molecules covering the CpG site.

149. The method of claim 146, wherein the normalization uses the average or median depth of cell-free DNA molecules in a region containing the CpG site.

150. 130. The method of claim 129, wherein the fractional concentration of DNA is in the range of 30% or less.

151. 130. The method of claim 129, wherein the first amount is determined for cell-free DNA molecules that terminate at position 0 of any one of the first set of CpG sites.

152. 1. A method for measuring the fractional concentration of DNA from a first tissue type in a biological sample from a subject, the biological sample comprising: analyzing a plurality of cell-free DNA molecules from the biological sample of the subject, wherein analyzing the cell-free DNA molecules comprises: determining a genomic position in a reference genome corresponding to at least one end of the cell-free DNA molecule; determining a terminal sequence motif of at least one end of the cell-free DNA molecule, the end of the cell-free DNA molecule having a first position at an outermost position, a second position adjacent to the first position, and a third position adjacent to the second position; identifying a first set of CpG sites in the reference genome that are all hypermethylated or hypomethylated for the first tissue type, the first set of CpG sites comprising one or more CpG sites; determining a first amount of a first set of one or more end sequence motifs, the first set of one or more terminal sequence motifs has a C at the first position and a G at the second position; or determining that a first set of one or more terminal sequence motifs has a C at the second position and a G at the third position; determining a fractional concentration of the DNA from the first tissue type in the biological sample by comparing the first amount to a calibration value, the calibration value being determined from one or more calibration samples having known fractional concentrations of DNA from the first tissue type.

153. 1. A method for measuring 5hmC methylation of a region in a genome of a subject using cell-free DNA molecules, comprising: analyzing a plurality of cell-free DNA molecules from a biological sample of the subject, the analyzing the cell-free DNA molecules comprising: determining a terminal sequence motif of at least one end of the cell-free DNA molecules, wherein the end of the cell-free DNA molecule has a first position at an outermost position, a second position adjacent to the first position, and a third position adjacent to the second position, and the plurality of cell-free DNA molecules are located in the region of a reference genome; determining a first amount of a first set of one or more end sequence motifs, the first set of one or more terminal sequence motifs has a C at the first position and a G at the second position; or determining that a first set of one or more terminal sequence motifs has a C at the second position and a G at the third position; determining a classification of 5hmC methylation of said region in the genome of said subject by comparing said first amount to a calibration value, said calibration value being determined using cell-free DNA molecules derived from one or more calibration samples and located at CpG sites having known classifications of levels of 5hmC methylation.

154. 154. The method of claim 153, wherein the first set of one or more terminal sequence motifs has a C at the first position and a G at the second position.

155. determining a second amount of a second set of one or more end sequence motifs having a C at the second position and a G at the third position; 155. The method of claim 154, wherein determining the classification comprises normalizing the first amount by the second amount to obtain a normalized first amount that is compared to the calibration value.

156. 154. The method of claim 153, wherein the first set of one or more terminal sequence motifs has a C at the second position and a G at the third position.

157. determining an expected amount of the first set of one or more end sequence motifs based on a reference sequence of the reference genome in the region; 154. The method of claim 153, wherein determining the classification comprises normalizing the first amount with the expected amount to obtain a normalized first amount that is compared to the calibration value.

158. determining the amount of each trimer end sequence motif having a C at the first position and a G at the second position, thereby determining the amount of each; generating a feature vector including each of the quantities, including the first quantity; 154. The method of claim 153, further comprising, as part of determining the classification of the region in the genome of the subject, inputting the feature vector into a machine learning model, wherein the machine learning model is trained using cell-free DNA molecules derived from the one or more calibration samples.

159. 2. The method of any one of the preceding claims, wherein analysing the plurality of cell-free DNA molecules uses sequencing.

160. 2. The method of any one of the preceding claims, wherein analyzing the plurality of cell-free DNA molecules uses PCR.

161. 161. The method of claim 160, wherein the PCR targets a sequence in the repeat region.

162. 2. The method of any one of the preceding claims, wherein analysing the plurality of cell-free DNA molecules comprises analysing at least 1,000 cell-free DNA molecules.

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

164. 1. A system comprising: A computer product according to claim 163; and one or more processors for executing instructions stored on a computer-readable medium.

165. A system comprising means for carrying out any of the above methods.

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

167. A system comprising modules for performing each of the steps of any of the above methods.

Images