Base modification analysis using electrical signals

Abstract

Systems and methods for determining base modifications using electrical signals and other data are described herein. Embodiments can utilize features derived from sequencing-related electrical signals, such as those obtained from using a nanopore (116), that are affected by various base modifications, as well as the identity of nucleotides in a window around a target position for which methylation status is determined. Other features can include vectors of statistics of segments of electrical signals corresponding to nucleotides, and statistics of electrical signals in windows within regions of a nucleic acid molecule. The detected base modifications can be used for further analysis of the biological sample.

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims the benefit of priority to U.S. Provisional Patent Application No. 63 / 173,728, filed April 12, 2021, which is incorporated by reference herein in its entirety for all purposes. [Background technology]

[0002] The occurrence of base modifications in nucleic acids varies across different organisms, including viruses, bacteria, plants, fungi, nematodes, insects, and vertebrates (e.g., humans), etc. The most common base modification is the addition of methyl groups to different DNA bases at different positions, so-called methylation. Methylation has been found at cytosine, adenine, thymine, and guanine, such as 5mC (5-methylcytosine), 4mC (N4-methylcytosine), 5hmC (5-hydroxymethylcytosine), 5fC (5-formylcytosine), 5caC (5-carboxylcytosine), 1mA (N1-methyladenine), 3mA (N3-methyladenine), N6-methyladenine (6mA), 7mA (N7-methyladenine), 3mC (N3-methylcytosine), 2mG (N2-methylguanine), 6mG (O6-methylguanine), 7mG (N7-methylguanine), 3mT (N3-methylthymine), and 4mT (O4-methylthymine). In vertebrate genomes, 5mC is the most common type of base methylation, followed by guanine methylation (i.e., in a CpG context).

[0003] DNA methylation is essential for mammalian development and plays notable roles in gene expression and silencing, embryonic development, transcription, chromatin structure, X-chromosome inactivation, protecting repetitive elements against activity, maintaining genome stability during mitosis, and regulating parent-of-origin genomic imprinting.

[0004] DNA methylation plays many important roles in silencing promoters and enhancers in a coordinated manner (Robertson, 2005; Smith and Meissner, 2013). Many human diseases have been found to be associated with abnormalities in DNA methylation, including but not limited to imprinting disorders (e.g., Beckwith-Wiedemann syndrome and Prader-Willi syndrome), repeat instability disorders (e.g., Fragile X syndrome), autoimmune disorders (e.g., systemic lupus erythematosus), metabolic disorders (e.g., diabetes mellitus type I and type II), neurological disorders, aging, etc.

[0005] Accurately measuring the methylomic modifications of DNA molecules has many clinical implications. One widely used method to measure DNA methylation is to use bisulfite sequencing (BS-seq) (Lister et al., 2009; Frommer et al., 1992). In this approach, DNA samples are first treated with bisulfite to convert unmethylated cytosines (i.e., C) to uracil. In contrast, methylated cytosines remain unchanged. The bisulfite-modified DNA is then analyzed by DNA sequencing. In another approach, following bisulfite conversion, the modified DNA is then subjected to polymerase chain reaction (PCR) amplification using primers that can distinguish bisulfite-converted DNA of different methylation profiles (Herman et al., 1996). This latter approach is called methylation-specific PCR.

[0006] One drawback of such bisulfite-based approaches is that the bisulfite conversion step has been reported to significantly degrade a large portion of the treated DNA (Grunau, 2001). Another drawback is that the bisulfite conversion step generates a strong CG bias (Olova et al., 2018), typically resulting in a reduced signal-to-noise ratio for DNA mixtures with heterogeneous methylation states. Furthermore, bisulfite sequencing is not an ideal method for sequencing long DNA molecules due to the degradation of DNA during bisulfite treatment.

[0007] There are many ongoing efforts to achieve bisulfite-free determination of nucleic acid base modifications. However, commercially viable tools that achieve sensitivity and specificity levels comparable to bisulfite sequencing are lacking. Nanopore sequencing is a type of sequencing that is attractive because it does not require chemical labeling of the sample. Detection of base modifications by nanopore sequencing can be relatively low-cost and efficient.

[0008] Therefore, there is a need to determine base modifications by nanopore sequencing. In this disclosure, we describe new methods and systems for processing current signals generated by nanopore sequencing that have high sensitivity and specificity for base modification determination. Summary of the Invention [Means for solving the problem]

[0009] The described embodiments allow for the determination of base modifications such as 5mC in nucleic acids without pretreatment of the template DNA such as enzymatic and / or chemical conversion, or protein and / or antibody binding. The embodiments present in the present disclosure can be used to detect different types of base modifications, including but not limited to, for example, 4mC, 5hmC, 5fC, 5caC, 1mA, 3mA, 6mA, 7mA, 3mC, 2mG, 6mG, 7mG, 3mT, 4mT, etc. Such embodiments can utilize features derived from electrical signals related to sequencing, such as those obtained from using nanopores, that are affected by various base modifications, as well as the identity of the nucleotide in a window around the target position where the methylation status is determined. The raw electrical signal for a nucleotide can also be related to the nucleotide upstream or downstream of the nucleotide. The raw electrical signals can be assigned to different nucleotides using suitable techniques.

[0010] Embodiments of the present invention can be used with nanopore sequencing. One example of a nanopore sequencing system is commercialized by Oxford Nanopore Technologies. The method can use electrical signals measured using a nanopore. The method can use a vector containing the identity of the nucleotide, the position of the nucleotide relative to the target position, statistics of segments of the electrical signal corresponding to the nucleotide, and statistics of the electrical signal in a window within a region of the nucleic acid molecule.

[0011] The method developed by the present inventors can serve as a tool to detect base modifications in biological samples and evaluate the methylation profile of the sample for various purposes, including but not limited to research and diagnostic purposes. The detected methylation profile can be used for different analyses. The methylation profile can be used to detect the origin of DNA (e.g., maternal or fetal, tissue, bacteria). The detection of abnormal methylation profiles in tissues helps to identify developmental and other disorders in individuals.

[0012] A better understanding of the nature and advantages of embodiments of the present invention may be obtained by reference to the following detailed description and the accompanying drawings. [Brief description of the drawings]

[0013] [Figure 1] Illustrates nanopore sequencing. [Diagram 2] 4 illustrates different signal characteristics according to an embodiment of the present invention. [Diagram 3] 1 illustrates current signal segmentation and construction of a signal feature vector according to an embodiment of the present invention. [Figure 4] 1 is a graph of the distribution of event lengths (i.e., durations) for each nucleotide passing through a nanopore, in accordance with an embodiment of the invention. [Diagram 5] 1 illustrates the principle for 5mC detection using an integrated display matrix including current pattern, sequencing position, and sequencing context according to an embodiment of the present invention. [Figure 6] 1 illustrates the principle for base modification detection using an integrated display matrix including current pattern, sequencing position, and sequencing context based on both stands of double-stranded DNA, according to an embodiment of the present invention. [Figure 7] 13 shows the effect of kernel size on the performance of base modification analysis according to an embodiment of the present invention. [Figure 8] 1 shows the number of sequencing molecules used for training and testing in terms of methylation detection according to an embodiment of the present invention. [Figure 9A] FIG. 13 is a box plot of methylated probability for CpGs between the WGA DNA dataset and the M.SssI-treated DNA dataset using the IPM-CNN and IPM-RNN approaches according to an embodiment of the present invention. [Figure 9B] FIG. 13 is a box plot of methylated probability for CpGs between the WGA DNA dataset and the M.SssI-treated DNA dataset using the IPM-CNN and IPM-RNN approaches according to an embodiment of the present invention. [Figure 9C] FIG. 13 is a box plot of methylated probability for CpGs between the WGA DNA dataset and the M.SssI-treated DNA dataset using the IPM-CNN and IPM-RNN approaches according to an embodiment of the present invention. [Figure 9D] FIG. 13 is a box plot of methylated probability for CpGs between the WGA DNA dataset and the M.SssI-treated DNA dataset using the IPM-CNN and IPM-RNN approaches according to an embodiment of the present invention. [Figure 10A] 4 illustrates receiver operator characteristic (ROC) curves for training and testing data sets according to an embodiment of the present invention. [Figure 10B] 4 illustrates receiver operator characteristic (ROC) curves for training and testing data sets according to an embodiment of the present invention. [Figure 11] 1 is a table of the performance of different tools for methylation analysis according to an embodiment of the present invention. [Figure 12] 1 is a flow chart of a process for detecting a nucleotide modification in a nucleic acid molecule according to an embodiment of the present invention. [Figure 13] 1 is a flow chart of a process for detecting a nucleotide modification in a nucleic acid molecule according to an embodiment of the present invention. [Figure 14] 1 illustrates a measurement system in accordance with an embodiment of the present invention. [Figure 15] FIG. 1 illustrates a block diagram of an exemplary computer system usable with systems and methods according to embodiments of the present invention. [Figure 16] 1 shows a graph of the effect of different combinations of parameters on the area under the ROC curve (AUC), in accordance with an embodiment of the present invention. [Figure 17] 13 shows a graph of the effect of window size on AUC in accordance with an embodiment of the present invention. [Figure 18] 1 illustrates the principle for 6mA detection using an integrated display matrix including current pattern, sequencing position, and sequencing context, according to an embodiment of the present invention. [Figure 19] 13 shows a graph of AUC for 6 mA detection in accordance with an embodiment of the present invention. [Figure 20] 1 is a comparison of single molecule methylation levels determined by the IPM-RNN model for DNA from buffy coat and NPC tumor samples, according to an embodiment of the present invention. [Figure 21] 1 shows examples of single molecule methylation patterns according to embodiments of the present invention. [Figure 22] 1 is a graph of single molecule methylation levels of maternal-specific and fetal-specific cell-free DNA molecules, in accordance with an embodiment of the invention. [Figure 23] 1 is a ROC curve for determining fetal and maternal origin of cell-free DNA molecules using methylation patterns determined by an IPM-CNN model, in accordance with an embodiment of the invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] term A "tissue" corresponds to a group of cells that group together as a functional unit. More than one type of cell may be found within a single tissue. Different types of tissue may be composed of different types of cells (e.g., liver cells, alveolar cells, or blood cells) and may correspond to tissues from different organisms (maternal vs. fetal, tissues of a transplanted subject, tissues of organisms infected with microorganisms or viruses) or healthy vs. tumor cells. A "reference tissue" may correspond to the tissue used to determine tissue-specific methylation levels. Multiple samples of the same tissue type from different individuals can be used to determine tissue-specific methylation levels of that tissue type.

[0015] "Biological sample" refers to any cell sample taken from a human subject. The biological sample may be a tissue biopsy, a fine needle aspirate, or blood cells. The sample may also be an acellular sample taken from a pregnant woman, such as plasma or serum or urine. In various embodiments, the majority of the DNA in a biological sample from a pregnant woman enriched for cell-free DNA (e.g., a plasma sample obtained via a centrifugation protocol) may be acellular, e.g., more than 50%, more than 60%, more than 70%, more than 80%, more than 90%, more than 95%, or more than 99% of the DNA may be acellular. The centrifugation protocol may include, for example, obtaining a fluid portion at 3,000g x 10 minutes and recentrifuging at 30,000g for another 10 minutes to remove residual cells. In certain embodiments, the 3,000g centrifugation step may be followed by filtration of the fluid portion (e.g., using a filter with a pore size of 5 μm or less in diameter).

[0016] A "sequence read" refers to a strand of nucleotides sequenced from any part or all of a nucleic acid molecule. For example, a sequence read can be a short strand (e.g., 20-150) of nucleotides sequenced from a nucleic acid fragment, a short strand of nucleotides at one or both ends of a nucleic acid fragment, or the sequencing of an entire nucleic acid fragment present in a biological sample. Sequence reads can be obtained, for example, using sequencing techniques, or in a variety of ways using probes, for example, hybridization arrays or capture probes, or amplification techniques such as polymerase chain reaction (PCR) or linear amplification, using single primers or isothermal amplification.

[0017] A "site" (also called a "genomic site") corresponds to a single site, which may be a single base position, or a group of correlated base positions, e.g., a CpG site, or a larger group of correlated base positions. A "locus" may correspond to a region that includes multiple sites. A locus may contain only one site, which would make the locus equivalent to the site in that context.

[0018] "Methylation state" refers to the state of methylation at a given site, e.g., a site is either methylated, unmethylated, or optionally undetermined.

[0019] A "methylation index" for each genomic site (e.g., a CpG site) may refer to the proportion of DNA fragments (e.g., as determined from sequence reads or probes) that exhibit methylation at that site over the total number of reads that cover that site. A "read" may correspond to information obtained from a DNA fragment (e.g., the methylation state at a site). A read may be obtained using a reagent (e.g., a primer or a probe) that preferentially hybridizes to DNA fragments of a particular methylation state at one or more sites. Typically, such a reagent is applied after processing with a process that differentially modifies or differentially recognizes DNA molecules according to their methylation state, such as bisulfite conversion, or methylation-sensitive restriction enzymes, or methylation-binding proteins, or anti-methylcytosine antibodies, or single molecule sequencing technologies that recognize methylcytosine and hydroxymethylcytosine (e.g., single molecule real-time sequencing (e.g., from Pacific Biosciences) and nanopore sequencing (e.g., from Oxford Nanopore Technologies)).

[0020] The "methylation density" of a region may refer to the number of reads at a site in the region that exhibits methylation divided by the total number of reads that cover the site in this region. The site may have a specific feature, e.g., a CpG site. 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 that cover the CpG sites in this region (e.g., a particular CpG site, a CpG site within a CpG island, or a larger region). For example, the methylation density for each 100 kb bin in the human genome may be determined from the total number of unconverted cytosines (corresponding to methylated cytosines) at the CpG sites after bisulfite treatment as a percentage of all CpG sites covered by sequence reads mapped to the 100 kb region. This analysis can also be performed for other bin sizes, such as, for example, 500 bp, 5 kb, 10 kb, 50 kb, or 1 Mb. The region may be the entire genome, or a chromosome, or a portion of a chromosome (e.g., a chromosome arm). Alternatively, methylation density can be determined without bisulfite conversion using nanopore sequencing using embodiments described in this disclosure. The methylation index of a CpG site is the same as the methylation density for that region if the region contains only that CpG site. "Percentage of methylated cytosines" may refer to the number of cytosine sites, "C", that are shown to be methylated (e.g., unconverted after bisulfite conversion), i.e., containing cytosines outside of a CpG context, over the total number of cytosine residues analyzed in the region. Methylation index, methylation density, count of molecules methylated at one or more sites, and proportion of molecules (e.g., cytosines) methylated at one or more sites are examples of "methylation levels."Apart from bisulfite conversion, other processes known to those skilled in the art can be used to examine the methylation state of DNA molecules, including but not limited to, single molecule sequencing using methylation state sensitive enzymes (e.g., methylation sensitive restriction enzymes), methylation binding proteins, methylation state sensitive platforms (e.g., by nanopore sequencing (Schreiber et al. Proc Natl Acad Sci 2013;110:18910-18915) and single molecule real-time sequencing (e.g., from Pacific Biosciences) (Flusberg et al. Nat Methods 2010;7:461-465)).

[0021] A "methylome" provides a measure of the amount of DNA methylation at multiple sites or loci in a genome. The methylome may correspond to the entire genome, a substantial portion of the genome, or a relatively small portion of the genome.

[0022] A "pregnant plasma methylome" is a methylome determined from the plasma or serum of a pregnant animal (e.g., a human). A pregnant plasma methylome is an example of a cell-free methylome because plasma and serum contain cell-free DNA. A pregnant plasma methylome is also an example of a mixed methylome because it is a mixture of DNA from different organs or tissues or cells in the body. In one embodiment, such cells are hematopoietic cells, including but not limited to cells of the erythroid (i.e., red cell) lineage, myeloid lineage (e.g., neutrophils and their precursors), and megakaryocytic lineage. During pregnancy, the plasma methylome may contain methylomic information from the fetus and the mother. A "cellular methylome" corresponds to the methylome determined from the patient's cells (e.g., blood cells). The methylome of blood cells is referred to as the blood cell methylome.

[0023] A "methylation profile" includes information related to DNA or RNA methylation at multiple sites or regions. Information related to DNA methylation may include, but is not limited to, the methylation index of CpG sites, the methylation density (abbreviated MD) of CpG sites in a region, the distribution of CpG sites across a contiguous region, the pattern or level of methylation of each individual CpG site within a region containing two or more CpG sites, and non-CpG methylation. In one embodiment, a methylation profile may include the pattern of methylation or non-methylation of two or more types of bases (e.g., cytosine or adenine). The methylation profile of a substantial portion of a genome can be considered equivalent to a methylome. "DNA methylation" in mammalian genomes typically refers to the addition of a methyl group to the 5' carbon of a cytosine residue between CpG dinucleotides (i.e., 5-methylcytosine). DNA methylation may occur at cytosine in other contexts, e.g., CHG and CHH, where H is adenine, cytosine, or thymine. Cytosine methylation can also be in the form of 5-hydroxymethylcytosine. 6 Non-cytosine methylations such as -methyladenine have also been reported.

[0024] "Methylation pattern" refers to the order of methylated and unmethylated bases. For example, the methylation pattern can be the order of methylated bases on a single DNA strand, a single double-stranded DNA molecule, or another type of nucleic acid molecule. As an example, three consecutive CpG sites can have any of the following methylation patterns: UUU, MMM, UMM, UMU, UUM, MUM, MUU, or MMU, where "U" indicates an unmethylated site and "M" indicates a methylated site. When extending this concept to base modifications, including, but not limited to, methylation, we will use the term "modification pattern" which refers to the order of modified and unmodified bases. For example, the modification pattern can be the order of modified bases on a single DNA strand, a single double-stranded DNA molecule, or another type of nucleic acid molecule. As an example, three consecutive potentially modifiable sites may have any of the following modification patterns: UUU, MMM, UMM, UMU, UUM, MUM, MUU, or MMU, where "U" indicates an unmodified site and "M" indicates a modified site. One example of a non-methylation-based base modification is an oxidative change, such as 8-oxo-guanine.

[0025] The terms "hypermethylated" and "hypomethylated" may refer to the methylation density of a single DNA molecule measured by the methylation level of that single molecule, e.g., the number of methylated bases or nucleotides in that molecule divided by the total number of methylatable bases or nucleotides in that molecule. A hypermethylated molecule is one in which the methylation level of a single molecule is equal to or greater than a threshold value, which may be defined for each application. This threshold value may be 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 95%. A hypomethylated molecule is one in which the methylation level of a single molecule is equal to or less than a threshold value, which may be defined for each application and may vary for each application. This threshold value may be 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 95%.

[0026] The terms "hypermethylated" and "hypomethylated" may also refer to the methylation level of a population of DNA molecules as measured by the methylation level of a plurality of molecules of these molecules. A hypermethylated population of molecules is a population in which the methylation level of a plurality of molecules is equal to or greater than a threshold, which may be defined for each application and may vary from application to application. This threshold may be 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 95%. A hypomethylated population of molecules is a population in which the methylation level of a plurality of molecules is equal to or less than a threshold, which may be defined for each application. This threshold may be 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or 95%. In one embodiment, the population of molecules may be aligned to one or more selected genomic regions. In one embodiment, the selected genomic region may be associated with a disease such as a genetic disorder, an imprinting disorder, an epigenetic disorder, a metabolic disorder, or a neurological disorder. The selected genomic region may have a length of 50 nucleotides (nt), 100 nt, 200 nt, 300 nt, 500 nt, 1000 nt, 2 knt, 5 knt, 10 knt, 20 knt, 30 knt, 40 knt, 50 knt, 60 knt, 70 knt, 80 knt, 90 knt, 100 knt, 200 knt, 300 knt, 400 knt, 500 knt, or 1 Mnt.

[0027] The term "classification" as used herein refers to any number or other feature associated with a particular characteristic of a sample. For example, a "+" symbol (or the word "positive") can mean that the sample is classified as having a deletion or amplification. Classification can be binary (e.g., positive or negative) or can have more levels of classification (e.g., a scale of 1-10 or 0-1).

[0028] The terms "cutoff" and "threshold" refer to a predetermined number used in the operation. For example, a cutoff size may refer to a size above which fragments are excluded. A threshold may be a value above or below which a particular classification is applied. Either of these terms may be used in any of these contexts. A cutoff or threshold may be a "reference value" or may be derived from a reference value that represents a particular classification or distinguishes between two or more classifications. Such a reference value may be determined in a variety of ways, as will be understood by those skilled in the art. For example, a metric may be determined for two different cohorts of subjects with different known classifications, and a reference value may be selected as representative of one classification (e.g., the average) or as a value between two clusters of the metric (e.g., selected to obtain a desired sensitivity and specificity). As another example, a reference value may be determined based on statistical analysis or simulation of samples.

[0029] "Level of pathology" (or level of injury) may refer to the amount, extent, or severity of pathology associated with an organism, which can be measured through the analysis of its cells. Another example of pathology is the rejection of a transplanted organ. Other example pathologies may include genomic imprinting disorders, autoimmune attacks (e.g., lupus nephritis, which damages the kidney, or multiple sclerosis, which damages the 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 healthy state of a subject may be considered a pathology-free classification.

[0030] "Pregnancy-related disorder" includes any disorder characterized by abnormal relative expression levels of genes in maternal and / or fetal tissues. These disorders include, but are not limited to, preeclampsia, intrauterine growth restriction, invasive placentation, preterm birth, hemolytic disease of the newborn, placental insufficiency, hydrops fetalis, fetal malformations, HELLP (hemolysis, elevated liver enzymes, and low platelet count) syndrome, systemic lupus erythematosus (SLE), and other maternal immune disorders. In some embodiments, a pregnancy-related disorder is any condition associated with physiological or morphological abnormalities during pregnancy.

[0031] The abbreviation "bp" refers to base pairs. In some cases, "bp" may be used to indicate the length of a DNA fragment even if the DNA fragment is single-stranded and does not contain base pairs. In the context of single-stranded DNA, "bp" may be interpreted as providing the length of the strand in nucleotides.

[0032] The abbreviation "nt" refers to a nucleotide. In some cases, "nt" may be used to indicate the length of a single strand of DNA in bases. Also, "nt" may be used to indicate a relative location, such as upstream or downstream, of the analyzed locus. In some contexts related to technical conceptualization, data display, processing, and analysis, "nt" and "bp" may be used interchangeably.

[0033] The term "sequence context" may refer to the base composition (A, C, G, or T) and base order in a stretch of DNA. Such a stretch of DNA may surround a base that is subjected to base modification analysis or a target base. For example, sequence context may refer to bases upstream and / or downstream of a base that is subjected to base modification analysis.

[0034] The term "machine learning model" may include models based on predicting test data using sample data (e.g., training data), and thus may include supervised learning. Machine learning models are often developed using a computer or processor. Machine learning models may include statistical models.

[0035] The term "data analytics framework" may include algorithms and / or models that can receive data as input and then output a predictive outcome. Examples of "data analytics frameworks" include statistical models, mathematical models, machine learning models, other artificial intelligence models, and combinations thereof.

[0036] The term "real-time sequencing" can refer to techniques that involve data collection or monitoring during the process involved in sequencing. For example, real-time sequencing can involve monitoring the electrical signal of the ionic current through a nanopore as a nucleotide strand translocates through the nanopore.

[0037] The term "electrical signal" may refer to a voltage or current that conveys information. Electrical signals may be represented in a variety of regular and / or irregular signal waveform types and / or shapes, such as square waves, rectangular waves, triangle waves, sawtooth waves, or various pulses and spikes. Electrical signals may include visual representations of the voltage or current variations over time. Measurements of electrical signals may be sampled at specific times (e.g., milliseconds). For example, current may be sampled at frequencies of 1 kHz, 2 kHz, 3 kHz, 4 kHz, 5 kHz, 10 kHz, 20 kHz, 30 kHz, 40 kHz, 50 kHz, 100 kHz, etc.

[0038] The term "signal segment" or "segment" may refer to a portion of the trace of the electrical signal associated with the sequencing of a particular nucleotide. A segment may correspond to a nucleotide determined from a base call in nanopore sequencing. A segment may cover a particular duration of the trace. Different segments may have different durations. Segments may not overlap. In some embodiments, the electrical signal amplitude may have a particular variation in a segment. For example, the electrical signal amplitude may be within 5%, 10%, 20%, 30%, or 40% of the mean or median electrical signal amplitude in a segment.

[0039] The term "about" or "approximately" may mean within an acceptable error range of a particular value as determined by one of ordinary skill in the art, which depends in part on the limitations of the method of measuring or determining the value, i.e., the measurement system. For example, "about" may mean within 1 or more than 1 standard deviation, as is customary in the art. Alternatively, "about" may mean within a range of up to 20%, up to 10%, up to 5%, or up to 1% of a given value. Alternatively, particularly with respect to biological systems or processes, the term "about" or "approximately" may mean within an order of magnitude, within 5-fold, and more preferably within 2-fold of a value. When a particular value is described in the present application and claims, unless otherwise indicated, the term "about" within an acceptable error range of the particular value should be assumed. The term "about" may have the meaning commonly understood by one of ordinary skill in the art. The term "about" may refer to ±10%. The term "about" may refer to ±5%. [Mode for carrying out the invention]

[0040] Accurate and efficient methods to detect base modifications (e.g., methylation) using nanopore sequencing are desired. Research studies have investigated the feasibility of using electrical signals generated by nanopore sequencing to analyze DNA methylation (Simpson et al. Nat Methods. 2017;14:407-410, Liu et al. Nat Commun. 2019;10:2449, Ni et al. Bioinformatics. 2019;35:4586-4595). Reported performance for 5-methylcytosine (5mC) detection has been suboptimal in many validation studies. For example, the sensitivity of 5mC detection using a computational tool called DeepSignal was reported to be 79% with 88% specificity when analyzing H. sapiens R9.4 1D data based on sample NA12878 (Ni et al. Bioinformatics. 2019;35:4586-4595). If one aims to achieve a higher specificity (e.g., >95%), the sensitivity is expected to decrease even further. For another tool called nanopolish (Liu et al. Nat Commun. 2019;10:2449), analyzing the same data set, the sensitivity was only 0.61 with a specificity of 0.46. The nanopolish software was based on a hidden Markov model with the following assumptions: (1) the electrical signal of a 6-nucleotide oligomer (i.e., 6-mer) in a DNA sequence followed a Gaussian distribution, (2) the probability of a methylation state (methylated or unmethylated) for a particular base depended only on the methylation state of the previous base, and (3) the probability of outputting a particular current level depended only on the methylation state generating the current signal, and not on any other methylation state or any other current signal. Those assumptions may be inaccurate in the actual current signal generated during nanopore sequencing, thus leading to lower sensitivity and specificity.

[0041] A recent computational tool called DeepMod for DNA methylation analysis based on Oxford Nanopore sequencing attempted to use a bidirectional recurrent neural network (RNN). However, the design of such an approach aims to measure methylation levels at genomic locations by aggregating predicted results from sequencing reads with electrical signals, and therefore lacks the ability to analyze methylation patterns at the single molecule level. In addition, the median sequencing depth across a dataset including Escherichia coli, Chlamydomonas reinhardtii, and Homo sapiens was approximately 33x. In many commercial applications, a lower sequencing depth is desirable to save economic costs and analysis time. It is unclear whether the DeepMod software can analyze methylation patterns with practically meaningful accuracy at the single molecule level.

[0042] In one study, Yuen et al. systematically benchmarked tools for CpG methylation detection from nanopore sequencing and concluded that most tools showed high variance and low concordance with the expected methylation percentage per CpG site (Yuen et al. bioRxiv. 2020; doi:doi.org / 10.1101 / 2020.10.14.340315).

[0043] Tse et al. reported that using single molecule real-time sequencing (SMRT-seq) from Pacific Biosciences (PacBio), kinetic features of DNA polymerases, including optical signals such as interpulse duration (IPD) and pulse width (PW) generated by incorporating fluorophore-labeled nucleotides during DNA polymerization, can be used to distinguish methylated and unmethylated CpG sites based on analyzing measurement windows consisting of two or more bases by using a convolutional neural network (Tse et al. Proc Natl Acad Sci USA. 2021;118:e2019768118, U.S. Patent No. 11,091,794). Such measurement windows organized IPDs and PWs into different sequencing contexts and sequencing positions. However, nanopore sequencing used a completely different sequencing mechanism that responds to current signals caused by strands of double-stranded DNA passing through the nanopore. Such raw electrical signals will vary with different nucleotides passing through the nanopore, and the electrical signal of a particular nucleotide will be affected by the upstream and downstream nucleotides nearby that nucleotide. Thus, different nucleotides will have different lengths of electrical signal traces detected, and even identical nucleotides will have different lengths of electrical signal traces. When analyzing electrical signals associated with a particular nucleotide or more than one nucleotide passing through the nanopore, the length of the electrical signal trace detected on each base is not fixed over time. In contrast, previous work on 5mC detection using PacBio SMRT-seq was based on two fixed measurements associated with the optical signal for each nucleotide: IPD and PW (Tse et al. Proc Natl Acad Sci USA. 2021;118:e2019768118). Therefore, the trained model presented in Tse et al.'s study (Tse et al. Proc Natl Acad Sci USA. 2021;118:e2019768118) does not apply to such electrical signals generated by nanopore sequencing.

[0044] The embodiments described herein use electrical signals obtained from nanopore sequencing to detect nucleotide modifications. Nucleotide methylation can include any of the methylations described herein. Information obtained from nanopore sequencing can include the identity of the nucleotide, the position of the nucleotide relative to the target position, a vector containing statistics of segments of the electrical signal corresponding to the nucleotide, and statistics of the electrical signal in a window within a region of the nucleic acid molecule.

[0045] The embodiments present in this disclosure can be used for DNA obtained from cell samples obtained from an organism (e.g., cell lines, solid organs, solid tissues, samples obtained via endoscopy, chorionic villus samples). The embodiments in this disclosure can also be used for cell samples obtained from the environment (e.g., bacteria, cellular contaminants), food (e.g., meat). The embodiments present in this disclosure can also be used for plasma or serum obtained from pregnant women. In some embodiments, the methods present in this disclosure can also be applied after a step in which a portion of the genome is first enriched, for example, using hybridization probes (Albert et al., 2007; Okou et al., 2007; Lee et al., 2011), or approaches based on physical separation (based on size, etc.) or following restriction enzyme digestion (e.g., MspI), or Cas9-based enrichment (Watson et al., 2019). Enzymatic or chemical conversion is not necessary for the invention to function, but in certain embodiments, such a conversion step may be included to further enhance the performance of the invention.

[0046] The embodiments of the present disclosure improve nanopore sequencing so that modified bases can be detected accurately and efficiently. Base modifications can be detected directly. The embodiments can avoid enzymatic or chemical conversions that may not preserve all modification information for detection. Furthermore, certain enzymatic or chemical conversions may not be compatible with certain types of modifications. The embodiments of the present disclosure can also avoid amplification by PCR, which may not transfer base modification information to the PCR product. Furthermore, both strands of DNA can be sequenced together, thereby allowing pairing of sequences from one strand with sequences complementary to the other strand. In contrast, PCR amplification splits the two strands of double-stranded DNA, making such combined analysis of sequences from two consecutive strands difficult.

[0047] Furthermore, nanopore sequencing is more cost-effective and portable than other sequencing technologies. For example, a nanopore sequencing system, Oxford Nanopore Technologies MinION™, costs approximately 5,000 USD, whereas an optical signal-based sequencing system, the PacBio SMRT™ Sequel II system, is on the order of about 500,000 to 700,000 USD. Nanopore sequencing speed is about 450 nucleotides per second, whereas PacBio SMRT™ sequencing is about 5 nucleotides per second. Thus, in the same period of time, nanopore sequencing can obtain more data than using an optical signal-based sequencing system.

[0048] The determined methylation profile, with or without enzymatic or chemical conversion, can be used to analyze biological samples. In one embodiment, the methylation profile can be used to detect the origin of cellular DNA (e.g., maternal or fetal, tissue, or viral). Detection of abnormal methylation profiles in tissues helps to identify developmental disorders in individuals. Methylation patterns of single molecules can identify chimeric DNA (e.g., between viral and human) and hybrid DNA (e.g., between two genes that are not normally fused in the native genome) or between two species (e.g., due to genetic or genomic engineering).

[0049] I. Principles of Nanopore Sequencing An example of a single molecule sequencing technology is nanopore sequencing (Oxford Nanopore Technologies). Figure 1 shows the principle for nanopore sequencing of DNA molecules (e.g., DNA molecule 104). When a single DNA molecule passes through a nanometer-sized pore, the electrical signal pattern caused by the flow of ionic current across the membrane was used to determine the sequence of the nucleic acid. Such pores can be constructed, for example, by proteins (e.g., alpha hemolysin, erolysin, and Mycobacterium smegmatis porin A (MspA)), or synthetic materials such as silicon or graphene, but are not limited to these (Magi et al, Brief Bioinform. 2018;19:1256-1272).

[0050] In one embodiment, the double-stranded DNA molecule was subjected to an end-repair process. Such a process converts the DNA into blunt-ended DNA and subsequently adds an A-tail that facilitates sequencing adaptor ligation. Sequencing adaptors (i.e., motor adaptors), each carrying a motor protein (e.g., motor protein 108), are ligated to both ends of the DNA molecule. The sequencing process begins when the motor protein (e.g., motor protein 112) unwinds the double-stranded DNA, allowing the first strand to pass through the nanopore. As the DNA strand passes through the nanopore 116, a sensor (e.g., an electrode) measures the change in ionic current over time (milliseconds, ms) in picoamps (pA) depending on the sequence context and associated base modifications (referred to as a one-dimensional (1D) read). Graph 120 shows an exemplary current signal versus time. In another embodiment, a hairpin sequence adaptor would be used to covalently link the first strand and its complementary strand together for a double-stranded DNA molecule. Thus, during sequencing, a strand of a double-stranded DNA molecule is sequenced, followed by the complementary strand (1D 2 In yet another embodiment, one end of a double-stranded DNA molecule linked by a protein increases the likelihood of sequencing the complementary strand following completion of sequencing of the first strand of the same molecule, resulting in 1D reads (also called two-dimensional (2D) reads) being sequenced, which may improve the accuracy of sequencing. 2 Generate leads.

[0051] The raw signals (e.g., the currents in graph 120) are used for base calling and base modification analysis. In some embodiments, the base calling and base modification analysis is performed by a machine learning approach, such as, but not limited to, a recurrent neural network (RNN), a convolutional neural network (CNN), a hidden Markov model (HMM), or one or more combinations thereof.

[0052] In one embodiment, the inventors developed a new method to process the current signals generated by nanopore sequencing and analyzed the processed signals for the determination of DNA methylation at the single molecule level based on a convolutional neural network (CNN) or a recurrent neural network (RNN).

[0053] II.Current signal analysis The current signal from nanopore sequencing can be analyzed to identify base modifications. However, the machine learning approach described in FIG. 1 does not use only the input of the raw current obtained using the nanopore. The embodiments described herein use one or more statistics of a portion of the current. A vector of these one or more statistics may be combined with other information corresponding to a window of nucleotides, including the identity of the nucleotide and the position of the nucleotide. The position of the nucleotide may be relative to a target position within the window, where the target position is the position where the modification or lack thereof is detected. The information about the window of nucleotides can be included with the statistics of the electrical signal within the region of the nucleic acid molecule to form an input data structure. Models trained on these input data structures can be used to detect base modifications.

[0054] A. Current Vector Parameters For a nucleotide strand passing through the nanopore, N events (i.e., signal segments associated with different identified nucleotides) are detected. In one embodiment, an event corresponds to one nucleotide identified during a base call with a series of electrical signals sampled at a specific time unit (e.g., milliseconds). In one example, the current was sampled at a frequency of 4 kHz (Rang et al. Genome Biol. 2018; 19:90). In another embodiment, an event corresponds to two or more nucleotides identified during a base call with a series of electrical signals sampled at a specific time rate.

[0055] FIG. 2 shows a graph of a current signal. Current amplitude in picoamps on the y-axis. Time in milliseconds is on the x-axis. Dots (e.g., dot 204) indicate individual signal measurements. Lines (e.g., line 208) passing through adjacent dots indicate signal segments of signal measurements associated with a nucleotide (e.g., A for line 208). For event i, m i Assuming that we have a current signal, let the amplitude of current signal j on event i be P ij In one embodiment, for a nucleotide, a signal feature vector including X1, X2, X3, X4, and X5 is used to characterize the pattern of electrical signals associated with that nucleotide. The definitions for X1, X2, and X3 are illustrated in FIG. 2. X1 is P ij X2 is the average of P ij X3 is the standard deviation of P ij X1 is the median of the mean of the current signal divided by the standard deviation. X2 is the median of the mean of the current signal divided by the standard deviation. X3 is the median of the absolute deviation of the current signal divided by the standard deviation. X4 is the median of the absolute deviation of the current signal divided by the standard deviation. X5 can be considered as the z-score of the segment's current signal.

[0056] In one embodiment, P ij may be a normalized signal. Normalization may involve rescaling the current signal from the original range such that the normalized signal value is within the range of 0 and 1, using minimum and maximum values ​​for a portion or the entire nucleotide chain. Normalization may involve rescaling the current signal such that the mean of the normalized signal values ​​is 0 and the standard deviation is 1. Normalization may involve rescaling the current signal by using the median and standard deviation for a portion or the entire nucleotide chain.

[0057] X1 and X2 are the P associated with event i. ij The mean and standard deviation are shown.

[0058] X1 is defined by:

number

[0059] X2 is defined as follows:

number

[0060] X3 is defined as follows: X3=Median(P ij ), where i ranges from l to r, including events surrounding the base under investigation for base modification analysis (e.g., methylation at a CpG site). The variables l and r represent the left and right window of the sequence of the event (corresponding to the nucleotide sequence). The nucleotide sequence between l and r should generally be longer than the integrated representation matrix of the current signal pattern (referred to as IPM) described below. For a given event i, j ranges from 1 to m i X3 may be the median of the current signals used in determining all of the segments. X3 may be the same value for all of the segments since X3 is determined using currents for more than just a single segment. In some embodiments, X3 may be for a particular window. In other embodiments, X3 may be the median across multiple windows.

[0061] X4 is defined as follows: X4 = Median (|P ij -X3|), where || represents the absolute value and i ranges from l to r, including events surrounding the base under investigation for base modification analysis (e.g., methylation at a CpG site). For a given i, j ranges from 1 to m i X4 may be in the range of 0.01 to 0.05. X4 may be the median of the absolute deviation of the current signals used in determining all segments. X4 may be calculated using currents for more than just a single segment (e.g., using all sampled current values) and therefore may be the same value for all segments.

[0062] X5 is defined as follows:

number

[0063] X1 and X2 can be used to reflect signal changes within event i, representing the local pattern of electrical signal for each nucleotide. X3, X4, and X5 can be used to reflect signal changes for event i relative to other surrounding events ranging from l to r. In some embodiments, the surrounding events can be X-nt upstream and Y-nt downstream of the interrogated base for base modification analysis. X can include, but is not limited to, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 100, 150, 200, 300, 400, 500, 1000, 2000, 4000, 5000, and 10000. Y can include, but is not limited to, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 100, 150, 200, 300, 400, 500, 1000, 2000, 4000, 5000, and 10000. In one embodiment, the surrounding event can be an entire nucleotide chain passing through the nanopore.

[0064] B. Single-strand analysis FIG. 3 shows a graph of a current signal. Current amplitude in picoamps on the y-axis. Time in milliseconds is on the x-axis. Trace 304 is the current amplitude over time. Signal segments (e.g., segment 308) are portions of trace 304 that are associated with nucleotides. The current change will vary in response to different nucleotides passing through the nanopore. Base calling in nanopore sequencing generally relies on converting the current signal into different locally stationary states (i.e., events). The process of converting the current signal into different events is called electrical signal segmentation. Ion current changes include, but are not limited to, the amplitude (e.g., measured in picoamps, pA) of events corresponding to one or more nucleotides in a signal segment, the direction of the ion current, the duration of current events corresponding to one or more nucleotides in a signal segment, the rate of change of the ion current, and the relative amplitude across different signal segments. Amplitude can refer to the strength or magnitude of the current and does not necessarily mean alternating current. These current events are assigned to different bases using software called Tombo (Stoiber et al. bioRxiv.2016;doi.org / 10.1101 / 094672). One nucleotide will be associated with a series of events with different amplitudes. With such a tool (Tombo), we tried to test the difference in the nanopore signal assigned to a genomic base between two samples to infer whether such a base was modified or not, based on the Mann-Whitney U test (Stoiber et al. bioRxiv.2016;doi.org / 10.1101 / 094672). This tool (Tombo) could not analyze the methylation pattern at the single molecule level, since it did not take into account the upstream and downstream signals as well as the sequence context, and all signals from different sequence reads were aggregated to the genomic base. Tombo's performance has been compared to that of other tools such as Nanopolish and DeepSginal (Yuen et al. bioRxiv. 2020; doi:doi.org / 10.1101 / 2020.10.14.340315).

[0065] In one embodiment, to characterize the current pattern in a signal segment associated with a nucleotide, the average (X1) and standard deviation (X2) of the current amplitudes of the events in the signal segment are calculated. The median (X3) of the current amplitudes of the events associated with the entire molecule and the median (X4) of the absolute deviation of the current amplitudes of the events associated with the entire molecule are determined. The normalized signal (X5) for the signal segment is determined by the following formula:

number

[0066] For a nucleotide, a signal feature vector including X1, X2, X3, X4, and X5 is used to reflect the pattern of electrical signals associated with that nucleotide. For example, segment 308 may have a signal feature vector of [X1, X2, X3, X4, X5].

[0067] X1 and X2 represent the mean and standard deviation of the current amplitudes of the events in signal segment i. X3 represents the median of the current amplitudes of the events associated with the entire molecule. X4 represents the median of the absolute deviation of the current amplitudes of the events associated with the entire molecule. X5 represents the normalized signal for signal segment i.

[0068] FIG. 4 is a plot of the frequency of signal segment lengths. The length of the current event associated with a nucleotide (i.e., duration in milliseconds) is on the x-axis. The frequency of the lengths is shown on the y-axis. FIG. 4 shows that the length of each signal segment associated with a nucleotide was variable with a median of 9 (range: 1-3540).

[0069] Base modification will affect the electrical signal associated with its upstream and downstream nucleotides. In this disclosure, we collectively used the current signal associated with the nucleotide for base modification analysis, the current signal associated with the nucleotide near the nucleotide of interest, and the sequencing context to improve performance. DNA methylation at CpG sites (i.e., methylation at the fifth carbon of cytosine) is the most common type of base methylation in vertebrate genomes. Analysis of DNA methylation at CpG sites was used as an illustrative example for this disclosure.

[0070] FIG. 5 shows a process for determining methylation using a current signal from one strand by nanopore sequencing. In block 504, a double-stranded DNA molecule is provided. In block 508, the double-stranded DNA molecule is ligated with a sequencing adaptor suitable for nanopore sequencing. In block 512, nanopore sequencing is performed. The strand of the single double-stranded molecule moves through a pore embedded in the membrane, changing the ionic current signal flowing through the nanopore. In block 516, the current signal is obtained. The ionic current signal can be measured, for example, by a transformer electrode.

[0071] The current signals are processed by a segmentation step, for example using Tombo (Stoiber et al bioRxiv. 2016; doi.org / 10.1101 / 094672). These segmented electrical events are assigned to different nucleotides. In block 520, an integrated representation matrix (IPM) is constructed. The IPM is a matrix of current signal patterns, containing the current signal for each base, the sequencing context, and sequencing position information spanning a set of nucleotides near or surrounding the locus for base modification analysis. In one embodiment, the segmented electrical events associated with a nucleotide were described by a signal feature vector, i.e., [X1, X2, X3, X4, X5]. A cytosine in a CpG site, with several signal feature vectors, and e.g., 10 nt upstream and downstream of that cytosine (i.e., e.g., 21 nt in total), were used to form an IPM of the current signal patterns. For illustrative purposes,

number

[0072] Block 528 indicates CNN analysis. For CNN analysis, the IPM was fed to the input layer, followed by convolutional and output layer processes. The probability of methylation for CpGs (i.e., output methylation score, ranging from 0 to 1) was determined based on a sigmoid function in the output layer. This approach is referred to as IPM-CNN. In one embodiment, IPM for methylated CpG sites (M.SssI-treated DNA) and unmethylated CpG sites (Whole Genome Amplified (WGA) DNA) were used to train a CNN model. The target value of methylation for CpG sites in the dataset derived from M.Sss-treated DNA was defined as "1", while the target value of methylation for CpG sites in the dataset derived from WGA DNA was defined as "0". The optimal parameters of IPM-CNN were obtained by minimizing the overall prediction error (binary value: 0 or 1) between the output score calculated by the sigmoid function and the desired target output by iteratively updating the model parameters. The overall prediction error was determined by a sigmoid cross-entropy loss function in a deep learning algorithm (keras.io / ). The model parameters learned from the training dataset were used to analyze the methylation status in the test dataset and output a probability score suggesting the likelihood that the CpG site is methylated (i.e., the probability of methylation). In one embodiment, the CNN model used four two-dimensional (2D) convolutional layers with 32, 64, 128, and 256 filters, each with a kernel size of 25. The rectified linear unit (ReLU) activation function was used in the convolutional layers. Then, a batch normalization layer was applied. A flattening layer was further added, followed by a dropout layer with a dropout rate of 0.5, and then a fully connected layer containing 200 neurons with the use of a ReLU activation function. An output layer with one neuron was finally applied with a sigmoid activation function to obtain a probability score for the CpG site being methylated (i.e., the probability of methylation). The program for the CNN model was implemented based on the Keras deep learning framework (https: / / keras.io / ).

[0073] Block 532 indicates the RNN analysis. For the RNN analysis, the IPM was fed into the input layer, followed by the process of the long short-term memory (LSTM) layer and the output layer. The probability of methylation (ranging from 0 to 1) for the CpGs was determined based on a sigmoid function in the output layer. This approach is referred to as IPM-RNN. Using a training procedure similar to that used in IPM-RNN, optimal parameters of the IPM-RNN were obtained by minimizing the overall prediction error (binary value: 0 or 1) between the output score calculated by the sigmoid function and the desired target output by iteratively updating the model parameters. The model parameters learned from the training dataset were used to analyze the methylation status in the test dataset and output a probability score (i.e., probability of methylation) suggesting the likelihood that the CpG site is methylated. In one embodiment, an RNN model with LSTM units was used with two fully connected hidden layers, each with 256 hidden nodes. The last layer was followed by a dropout layer with a dropout rate of 0.2. An output layer with one neuron was finally applied with a sigmoid activation function to obtain a probability score for the methylated CpG site (i.e., the probability of methylation).The program for the CNN model was implemented based on the Keras deep learning framework (keras.io / ).

[0074] C. Double-stranded analysis FIG. 6 illustrates a process for determining methylation using current signals from both DNA strands by nanopore sequencing. In one embodiment, current signals from both nucleotide strands of a double-stranded DNA molecule can be obtained when such a double-stranded DNA molecule is sequenced such that a second nucleotide strand (called the complementary or Crick strand) immediately follows the completion of a first nucleotide strand (called the Watson strand) passing through the same nanopore. This technique for sequentially sequencing both nucleotide strands of a double-stranded DNA in the same nanopore can be performed in 1D. 2Or referred to as 2D sequencing. In block 604, a double-stranded DNA molecule is provided. In block 608, the double-stranded DNA molecule is ligated with a sequencing adaptor suitable for nanopore sequencing. In block 612, a strand of a single double-stranded molecule is translocated through a pore embedded in a membrane, followed by the complementary strand. In block 616, a current signal is obtained for both strands of each double-stranded DNA molecule. The ionic current signal can be measured by a transformer electrode. The obtained current signal is used to estimate (i.e., base call) the nucleotide information of the sequenced DNA molecule using Guppy (Oxford Nanopore Technologies Ltd). In some embodiments, other base calling tools may be used, including, but not limited to, Albacore (nanoporetech.com / ), WaveNano (Wang et al. Quantitative Biology. 2018;6:359-368), Chiron (Teng et al. GigaScience. 2018;7:giy037), Flappie (github.com / nanoporetech / flappie), Scrappie (github.com / nanoporetech / scrappie), and the like.

[0075] The current signals sampled at a certain time rate (e.g., milliseconds) are assigned to different detection nucleotides for base modification analysis. The current signals are processed by a segmentation step, for example, using Tombo (Stoiber et al bioRxiv.2016;doi.org / 10.1101 / 094672). These segmented electrical events are assigned to different nucleotides. In block 620, an integrated representation matrix (IPM) is constructed to include both strands from each double-stranded DNA molecule. In one embodiment, the segmented electrical events associated with a nucleotide were described by a signal feature vector, i.e., [X1, X2, X3, X4, X5]. The signal feature vector from the corresponding base of the complementary strand, i.e., [X1', X2', X3', X4', X5'], was obtained. A cytosine in a CpG site, with several signal feature vectors, and, for example, 10 nt upstream and downstream of that cytosine (i.e., for example, a total of 21 nt) were used to form an IPM of the current signal pattern. IPMs from the corresponding bases in the complementary strands of the same double-stranded DNA molecule were obtained. The IPMs from the Watson and Crick strands were combined to form a new IPM matrix with higher dimensions for base modification analysis.

[0076] In some embodiments, other computational tools can be used to assign current signals to different nucleotides, including NanoMod (Liu et al. BMC Genomics. 2019; 20: 78), Albacore (nanoporetech.com / ), Chiron (Teng et al. GigaScience. 2018; 7: giy037), Nanopolish (Simpson et al. Nat Methods. 2017; 13: 407-410), Scrappie (https: / / github.com / nanoporetech / scrappie), UNCALLED (Kovaka et al. Nat Biotechnol. 2020; doi: 10.1038 / s41587-020-0731-9), etc. These computational tools and other techniques described for double-stranded analysis may also be used for single-stranded analysis.

[0077] For illustrative purposes,

number

[0078] The remaining signal feature vectors for each nucleotide associated with the 21nt sequence context were filled using the same rules, thus forming a 21nt IPM. Thus, such an IPM would simultaneously encode the current signal pattern, the sequencing context, the sequencing position, and the changing pattern over time. Several IPMs derived from the methylated and unmethylated DNA datasets were used to train CNN or RNN models, which were subsequently used to determine the methylation status at CpG sites in the test samples.

[0079] Block 628 indicates CNN analysis. In an embodiment, the CNN model used four two-dimensional (2D) convolution layers with 32, 64, 128, and 256 filters, each with a kernel size of 1×25. A rectified linear unit (ReLU) activation function was used in the convolution layers. Then, a batch normalization layer was applied. A flattening layer was further added, followed by a dropout layer with a dropout rate of 0.5, and then a fully connected layer containing 200 neurons with the use of a ReLU activation function. An output layer with one neuron was finally applied with a sigmoid activation function to obtain a probability score for the methylated CpG sites (i.e., the probability of methylation). The program for the CNN model was implemented based on the Keras deep learning framework (keras.io / ). In some embodiments, the kernel size n×m may be varied, where "n" may include but is not limited to 1, 2, 3, 4, 5, 10, 15, 20, 30, 35, 40, 45, 50, 100, etc., and "m" may include but is not limited to 1, 2, 3, 4, 5, 10, 15, 20, 30, 35, 40, 45, 50, 100, etc.

[0080] Figure 7 is a table of the effect of kernel size on the performance of base modification analysis. Column 1 shows different kernel sizes. Column 2 shows the AUC (area under the ROC [receiver operator characteristic] curve) from the training data set. Column 3 shows the AUC from the test data set. Figure 7 shows that a range of kernel sizes such as 1x5, 1x10, 1x15, 1x20, and 1x25 give comparable performance in distinguishing between methylated and unmethylated CpG sites, as shown by the AUCs of 0.96, 0.96, 0.97, 0.96, and 0.96, respectively.

[0081] Block 632 indicates the RNN analysis. In an embodiment, an RNN model with LSTM units was used with two fully connected hidden layers, each with 256 hidden nodes. The current output of the LSTM hidden unit is determined by the current input and the previous information stored in the LSTM cell. As an example, the signal feature vector [X1, X2, X3, X4, X5] associated with the position shown in the first row of the 21nt IPM represents the input X for the LSTM unit at a particular time step. t The forward LSTM RNN recursively calculates the hidden layer H according to the time steps based on the following operations (Gers et al. IEEE Transactions on Neural Networks. 2001;12:1333-1340).

number

[0082] The backward LSTM RNN recursively computes the hidden layer H according to the time steps based on the following operations (Gers et al. IEEE Transactions on Neural Networks. 2001;12:1333-1340):

number

[0083] The outputs of the forward and backward LSTM RNN units are combined.

number

[0084] The last layer of the LSTM RNN output was followed by a dropout layer with a dropout rate of 0.2. An output layer with one neuron was finally applied with a sigmoid activation function to obtain a probability score for the methylated CpG sites (i.e., the probability of methylation). The program for the CNN model was implemented based on the Keras deep learning framework (keras.io / ).

[0085] D. Parameter Analysis The influence of different current vector parameters and different window sizes on AUC (area under the ROC [receiver operator characteristic] curve) is analyzed. Based on the IPM-CNN model, we analyzed the discrimination power of using different parameters in IPM according to the embodiment present in this disclosure. For this purpose, 8,282 molecules (38,238 CpG sites) and 8,247 molecules (39,708 CpG sites) were analyzed from the WGA DNA and M.SssI-treated DNA datasets, respectively.

[0086] Figure 16 shows a graph of the effect of different combinations of parameters on AUC. Different combinations of current vector parameters are on the x-axis and AUC is on the y-axis. Figure 16 shows that, without being limited thereto, the use of different combinations of parameters X1, X2, X3, X4 and X5 in IPM led to different performances of CpG methylation analysis. For example, the use of X1 in IPM resulted in an AUC of 0.954, while the combination of X1 and X2 in IPM produced an AUC of 0.893. The combination of X1, X2 and X3 in IPM raised the AUC to 0.963. The combination of X1, X2, X3 and X4 in IPM further raised the AUC to 0.978, followed by a plateau in performance at an AUC of 0.977 by the use of X1, X2, X3, X4 and X5 in this example. Thus, in some embodiments, different combinations of parameters in the IPM will allow determining the desired performance in distinguishing between methylated and unmethylated CpG sites.

[0087] The use of X1, X2, X3, X4, and X5 was tested individually, but not in combination. Results using X1, X2, X3, X4, and X5 individually resulted in AUCs of 0.95, 0.92, 0.98, 0.88, and 0.95, respectively. X3 (i.e., P within the region ij The median of the methylation ratio (M.Sssl) yielded a high AUC of 0.98. The high AUC may be at least partially a result of methylation differences at the complete fragment level. The dataset used involved WGA (fully unmethylated) and M.Sssl (fully methylated). However, in reality, fragments are either fully unmethylated or fully unmethylated. Using X3 alone on fully unmethylated or fully unmethylated samples may not yield a very high AUC.

[0088] FIG. 17 shows a graph of the effect of window size on AUC. The x-axis shows the window size in nucleotides. The y-axis shows the AUC. The number of nucleotides used in the IPM (also referred to as the window size) captures different information content of the current signal generated during nanopore sequencing and may affect the performance of the methylation analysis. FIG. 17 shows that the performance in distinguishing between methylated and unmethylated CpG sites using the IPM-CNN model appeared to gradually increase from an AUC of 0.715 to 0.969 as the number of nucleotides used in the IPM increased from 1 nt to 10 nt. In this example, a plateau in performance was reached at a window size of 7 nt. Thus, in some embodiments, adjusting the window size of the IPM will allow one to determine the desired performance in distinguishing between methylated and unmethylated CpG sites.

[0089] The embodiments may not require using the combination of current vector parameters or window sizes that leads to the highest AUC. A lower AUC may be sufficient for a particular application, or a higher AUC may not be worth the additional computational and storage costs associated with the additional parameters. Furthermore, different parameters may be adjusted to achieve a desired AUC, specificity, and / or sensitivity. For example, a larger window size may be used to compensate for using fewer parameters among X1, X2, X3, X4, and X5.

[0090] E. Detection of 6mA Modification To determine the applicability of current signal analysis to modifications other than 5mC, current signal analysis was used to detect N6-methyladenine (6 mA).

[0091] FIG. 18 shows a process for determining 6mA methylation using a current signal from one strand by nanopore sequencing. FIG. 18 is similar to FIG. 5 showing a process for determining 5mC methylation. In block 1804, a double-stranded DNA molecule is provided. In block 1808, the double-stranded DNA molecule is ligated with a sequencing adaptor suitable for nanopore sequencing. In block 1812, nanopore sequencing is performed. In block 1816, a current signal is obtained. In block 1820, an integrated representation matrix (IPM) is constructed. Blocks 1804-1820 can be the same as blocks 504-520.

[0092] For illustrative purposes, to determine methylation of 6mA,

number

[0093] The remaining signal feature vectors for each nucleotide associated with the 21nt sequence context were filled using the same rules, thus forming the 21nt IPM. Thus, such an IPM would simultaneously code the current signal pattern, the sequencing context, the sequencing position, and the pattern that changes over time. Several IPMs from the methylated and unmethylated DNA datasets associated with nucleotide A were used to train a CNN or RNN model, which was subsequently used to determine the methylation state at the A site in the test sample. Block 1828 shows the CNN analysis, and block 1832 shows the RNN analysis. These blocks can be the same as blocks 528 and 532.

[0094] To test whether our approach (IPM-CNN or IPM-RNN) illustrated above was able to determine adenine methylation (6mA), we downloaded two public datasets containing nanopore sequencing results of pUC19 plasmid DNA from a previous study (Rand et al. Nat Methods 2017;14:411-413). The first dataset (6mA dataset) was generated from pUC19 plasmid DNA grown in E. coli containing both dam and dcm methyltransferases, where all GATC motifs were assumed to be methylated at the A site. The second dataset (uA dataset) was generated from DNA subjected to PCR amplification with unmodified nucleotides, where all A sites were assumed to be unmethylated. In the training process, we used the IPM-CNN model to analyze 2052 molecules containing GATC motifs from the 6mA dataset and 2081 molecules from the uA dataset.

[0095] FIG. 19 shows the AUC obtained using the IPM-CNN model. The x-axis shows the specificity. The y-axis shows the sensitivity. Line 1904 shows the results from the training dataset. The AUC using the training dataset is 0.94. In the training process, we applied the trained IPM-CNN model to 522 molecules containing the GATC motif from the 6mA dataset and 481 molecules from the uA dataset. The AUC using the test dataset is 0.92. In addition, when using the IPM-RNN model, an AUC of 0.89 was achieved for both the training and test datasets. These data suggested that IPM-CNN and IPM-RNN may enable the differentiation of 6mA sites from unmethylated A sites.

[0096] In an embodiment, a training dataset for the determination of 6mA for human or non-human DNA may be constructed based on PCR amplification using 6mA nucleotides and unmethylated A nucleotides, respectively. After several cycles of PCR, the majority of DNA molecules will carry 6mA nucleotides for the dataset generated from DNA amplified with 6mA nucleotides, while the majority of DNA molecules will carry unmethylated A nucleotides for the dataset generated from DNA amplified with unmethylated A nucleotides. These two types of datasets can be used to train a CNN model and / or an RNN model for determining the methylation state of A nucleotides in a test sample.

[0097] The use of current signal analysis to detect 6mA in addition to 5mC demonstrates the applicability of such analysis to other methylation types. Thus, these methods should accurately detect other methylations described herein.

[0098] F. CpG methylation analysis between non-tumor and tumor tissues of human subjects Methylation of sites determined by using embodiments described herein can be used to distinguish different types of tissue. Using an IPM-RNN model according to embodiments of the present disclosure, we analyzed methylation patterns on cellular DNA molecules derived from nasopharyngeal carcinoma (NPC) tumors and buffy coat samples. For this, we used 147 molecules from NPC tumors with a median size of 4,406 bp (interquartile range (IQR): 1,962-8,128 bp) and a median of 32 CpGs per molecule (IQR: 13-61). We analyzed another 147 molecules from buffy coats with a median size of 6,823 bp (interquartile range (IQR): 2,515-9,304 bp) and a median of 49 CpGs per molecule (IQR: 23-118).

[0099] FIG. 20 shows a graph of the comparison of DNA molecules from buffy coat samples and NPC tumor tissue samples. The x-axis indicates tissue type. The y-axis indicates methylation levels as a percentage. Single molecule methylation levels (i.e., the percentage of CpG sites in the molecule determined to be methylated) in buffy coat (median: 74.8%, IQR: 71.1%-80.1%) were found to be significantly higher (P value < 0.0001, Wilcoxon rank sum test) than those in NPC tumors (median: 50, IQR: 45.7-53.1). DNA molecules derived from tumor tissues appeared to be hypomethylated, which was consistent with previous conclusions based on short-lead bisulfite sequencing (Chan et al. Proc Natl Acad Sci USA 2013;110:18761-8). However, the new nanopore sequencing technology described herein makes it possible to sequence nearly full-length DNA molecules and analyze methylation patterns on the DNA molecules. For example, nanopore sequencing can analyze DNA molecules with sizes larger than 600 bp, which cannot be interrogated by short-read sequencing platforms (e.g., Illumina).

[0100] FIG. 21 illustrates the methylation patterns in tumor and buffy coat DNA molecules. Filled circles (e.g., circle 2104) indicate methylated CpG sites. Open circles (e.g., circle 2108) indicate unmethylated CpG sites. The circles indicate the relative position of the CpG sites to the 5' end of the DNA molecule being analyzed (i.e., the left side of the DNA molecule in the figure is closer to the 5' end). As shown in FIG. 21, DNA molecules derived from tumor tissue tended to carry more unmethylated CpG sites in the molecule compared to those derived from buffy coat samples. Only 5.4% of the molecules from buffy coat samples had a single molecule methylation level below 50% with a median length of 2,091 bp. In comparison, 39.5% of the molecules from NPC tumor tissue had a single molecule methylation level below 50% with a median length of 2,924 bp. The lengths of the DNA molecules ranged from 897 bp to 10,424 bp.

[0101] These data indicate that the nanopore sequencing technology for detecting methylation described herein can be used for single molecule methylation pattern analysis to distinguish the tissue of origin of each DNA molecule (e.g., non-tumor DNA vs. tumor DNA molecules) from tissue biopsy samples. Analysis of single molecule methylation pattern analysis from tissue biopsies will enable testing of tumor grade or subtype, monitoring of treatment of cancer or other diseases, evaluation of organ abnormalities (e.g., renal failure), etc.

[0102] G. Analysis between fetal and maternal DNA molecules The methylation of the sites determined by using the embodiments described herein can be used to distinguish between fetal and maternal DNA molecules. According to the IPM-CNN model, we determined single molecule methylation patterns with at least 5 CpG sites for 1,262 fetal-specific cell-free DNA molecules (median size: 530 bp, IQR: 361-779 bp) and 6,108 maternal-specific cell-free DNA molecules (median size: 668 bp, IQR: 448-1,089 bp) obtained from pregnant women in the third trimester by utilizing SNP information between maternal buffy coat and placental tissues. The fetal DNA fraction in the plasma DNA of such pregnant women was 26.0%.

[0103] Figure 22 shows the single molecule methylation levels between maternal-specific and fetal-specific DNA molecules. The x-axis shows the category of cell-free DNA molecules, i.e., maternal-specific cell-free DNA molecules or fetal-specific cell-free DNA molecules. The y-axis shows the single molecule methylation levels in percentage. The median methylation level of single plasma DNA molecules (i.e., the percentage of CpG sites in the molecule that were determined to be methylated) was 66.6% (IQR: 28.5-86.6%) for fetal-specific cell-free DNA molecules, which was significantly lower than that for maternal-specific cell-free DNA molecules (median: 78.5%, IQR: 50-93.7%) (P value: <0.0001, Mann-Whitney U test). This result suggested that the use of methylation information of cell-free DNA molecules made it possible to distinguish the maternal and fetal origin of each plasma DNA molecule.

[0104] In addition, by comparing the methylation pattern determined by the IPM-CNN model with the respective reference methylation patterns of buffy coat and placental tissues, as described in U.S. patent application Ser. No. 17 / 168,950, filed February 5, 2021, an AUC of 0.87 can be achieved for distinguishing between plasma DNA molecules of fetal origin and plasma DNA molecules of maternal origin in pregnant women.

[0105] Figure 23 shows the ROC curve for fetal and maternal origin analysis of cell-free DNA molecules in pregnant women based on the methylation patterns determined by the IPM-CNN model. The x-axis is specificity and the y-axis is sensitivity.

[0106] III. Datasets for evaluation of IPM-based methylation decisions The unmethylated dataset included sequencing results from amplified DNA prepared via whole genome amplification (WGA) (shown as the WGA DNA dataset). The use of unmodified nucleotides in WGA resulted in amplified DNA that contained few base modifications (except for small amounts of input genomic DNA). The methylated dataset included sequencing results from DNA that was treated with M.SssI (a CpG methyltransferase isolated from a strain of Escherichia coli that contains the methyltransferase gene from Spiroplasma sp. strain MQ1, which methylates all CpG sites in double-stranded DNA) prior to sequencing (shown as the M.SssI-treated DNA dataset). The M.SssI methyltransferase methylated the CpG sites.

[0107] For preparation of the WGA DNA data set, exonuclease-resistant random primers are pre-annealed to 1 ng of DNA template by incubating the reaction mixture (containing phi29 reaction buffer and dNTPs) in a heating block at 95°C for 5 min, followed by cooling to 4°C. Phi29 polymerase was then added to the reaction mixture and incubated at 30°C for 4 h. DNA was purified with Ampure XP beads and quantified on a Qubit fluorometer. Typically, 200 ng of DNA can be obtained from a 20 μl reaction.

[0108] For preparation of M.SssI-treated DNA datasets, after WGA, half of the DNA was treated with M.SssI enzyme. Methyltransferase reaction buffer, S-adenosylmethionine (SAM), and M.SssI were mixed with the DNA and incubated at 37°C for 2 hours. The reaction was stopped by heating at 65°C for 20 minutes. A ligation sequencing kit (SQK-LSK109) (Oxford Nanopore) was used for library preparation. DNA was treated with NEBNext FFPE DNA Repair Mix together with NEBNext Ultra II End Repair / dA-tailing Module. After Ampure XP bead cleanup, sequencing adapters were ligated to the repaired DNA by adding Adapter Mix, Ligation Buffer, and NEBNext Quick T4 DNA Ligase. The ligated DNA was cleaned with Ampure XP beads and washed with Short Fragment Buffer. The libraries were resuspended in Elution Buffer. An R9.4.1 flow cell was used for sequencing each of the WGA (sample_01) and M.SssI treated (sample_02) libraries. The flow cell was first primed with a flow cell priming mix containing Flush Tether and Flush Buffer. The library loading mix was then prepared by mixing Sequencing Buffer, Loading Beads, and DNA library. The library loading mix was added dropwise to the flow cell sample port. The loaded flow cell was inserted into a slot in the PromethION and sequenced for 64 hours using default parameters.

[0109] We obtained 15.6 million and 15.3 million nanopore sequencing reads for sample_01 and sample_02, respectively, of which 13.8 million (88.7%) and 13.8 million (90.7%) reads could be aligned to the human reference genome (UCSC hg19) by using Minimap2 (Li H, Bioinformatics. 2018;34(18):3094-3100). The median read length was 510 nt (interquartile range (IQR): 333-778 nt) and 606 nt (IQR: 382-911 nt) for sample_01 and sample_02, respectively. In some embodiments, BLASR (Mark J Chaisson et al, BMC Bioinformatics. 2012; 13:238), BLAST (Altschul SF et al, J Mol Biol. 1990; 215(3): 403-410), BLAT (Kent WJ, Genome Res. 2002; 12(4): 656-664), BWA (Li H et al, Bioinformatics. 2010; 26(5): 589-595), NGMLR (Sedlazeck FJ et al, Nat Methods. 2018; 15(6): 461-468), and LAST (Kielbasa SM et al, Genome Res. 2011; 21(3): 487-493) may be used to align sequenced reads to a reference genome.

[0110] FIG. 8 is a table showing the number of sequencing molecules used to train and test CNN and RNN models based on IPM. Column 1 is the dataset. M.SssI-treated DNA is a methylated DNA dataset, and WGA DNA is an unmethylated DNA dataset. Column 2 is the number of molecules and the number of CpG sites used for training. Column 3 is the number of molecules and the number of CpG sites used for testing. For the training dataset, we randomly used 7,989 and 8,052 sequencing molecules from M.SssI-treated DNA (methylated DNA) and WGA DNA (unmethylated DNA), respectively. Such a training dataset contained 38,470 methylated CpG sites and 37,150 unmethylated CpG sites. For the test dataset, we randomly used 4,826 and 5,041 sequencing molecules from M.SssI-treated DNA (methylated DNA) and WGA DNA (unmethylated DNA), respectively. Such a training dataset contained 9,716 methylated CpG sites and 11,444 unmethylated CpG sites.

[0111] Figures 9A-9D are box plots of the probability of being methylated for CpGs between the WGA DNA dataset and the M.SssI-treated DNA dataset using the IPM-CNN and IPM-RNN approaches. The graphs have the dataset on the x-axis. The probability of methylation is on the y-axis. Figures 9A and 9B show the results of using IPM-CNN analysis. Figure 9A shows the IPM-CNN analysis of the training dataset, where the probability of methylation for CpGs in the M.SssI-treated DNA dataset (median: 0.99, IQR: 0.987-0.999) was significantly higher than that in the WGA DNA dataset (median: 0.03, IQR: 0.001-0.15) (P value < 0.0001, Mann-Whitney U test). Figure 9B shows the IPM-CNN analysis of the test dataset and also showed a significant difference in the probability of being methylated for CpGs between the WGA DNA dataset (median: 0.4, IQR: 0.002-0.18) and the M.SssI-treated DNA dataset (median: 0.99, IQR: 0.980-0.999) (P value < 0.0001, Mann-Whitney U test).

[0112] Figures 9C and 9D show the results of using IPM-RNN analysis. Figure 9C shows the IPM-RNN analysis of the training dataset, and the probability of being methylated for CpGs in the M.SssI-treated DNA dataset (median: 0.994, IQR: 0.92-0.99) was significantly higher than that in the WGA DNA dataset (median: 0.079, IQR: 0.059-0.118) (P value < 0.0001, Mann-Whitney U test). FIG. 9D shows the IPM-RNN analysis of the test dataset, which also showed a significant difference in the probability of being methylated for CpGs between the WGA DNA dataset (median: 0.077, IQR: 0.057-0.115) and the M.SssI-treated DNA dataset (median: 0.994, IQR: 0.919-0.999) (P-value < 0.0001, Mann-Whitney U test). These results indicated that it is feasible to determine the methylation status at CpG sites using electrical signals generated by nanopore sequencing according to embodiments present in the present disclosure. In one embodiment, a probability of methylation cutoff of 0.5 can be used to determine the methylation status at CpG sites. With the use of this cutoff, the specificity and sensitivity for DNA methylation detection for the IPM-CNN analysis were 96% and 91%, respectively, for the training dataset, and 93% and 88%, respectively, for the test dataset. For the IPM-RNN analysis, the specificity and sensitivity for DNA methylation detection were 97% and 88%, respectively, for both the training and test datasets. In some embodiments, the cutoff for the probability of methylation can be adjusted for different applications.

[0113] 10A and 10B show receiver operator characteristic (ROC) curve analysis. Specificity is shown on the x-axis. Sensitivity is shown on the y-axis. FIG. 10A shows the results for the training dataset. FIG. 10B shows the results for the test dataset. IPM-CNN results are shown by lines 1004 and 1008. IPM-RNN results are shown by lines 1012 and 1016. DeepMod (Liu et al. Nat Commun. 2019; 10:2449) results are shown by lines 1020 and 1024. Nanopolish (Liu et al. Nat Commun. 2019; 10:2449) results are shown by lines 1028 and 1032. IPM-based CNN and RNN analysis yielded good performance on both the training and test datasets, with an area under the ROC curve (AUC) of 0.95 or greater. The IPM-based CNN and RNN models yielded better performance with areas under the ROC curve (AUC) of 0.95 and 0.97 in the test dataset compared to DeepMod (0.83) and nanopolish (0.91). The P-values ​​(DeLong test) for all comparisons of IPM-based RNN or CNN against other tools, including DeepMod and nanopolish, were found to be <0.0001. These results indicated that IPM-CNN and IPM-RNN outperformed other tools for DNA methylation analysis.

[0114] FIG. 11 is a table of sensitivity for a given specificity for different analyses. The first column indicates the type of analysis. The second column indicates the sensitivity. The third column indicates the specificity. FIG. 11 shows that at a given specificity, the IPM-CNN and IPM-RNN analyses achieved much higher sensitivity. For example, at a specificity of 90%, the IPM-CNN and IPM-RNN analyses achieved 90% and 93% sensitivity, respectively, while the DeepMod and nanopolish approaches achieved only 53% and 74%, respectively. At a specificity of 95%, the IPM-CNN and IPM-RNN analyses achieved 86% and 90% sensitivity, respectively, while the DeepMod and nanopolish approaches achieved only 38% and 55%, respectively. At 99% specificity, IPM-CNN and IPM-RNN analyses achieved sensitivities of 70% and 83%, respectively, while DeepMod and nanopolish achieved sensitivities of only 13% and 16%, respectively. These results further demonstrated that an integrated representation matrix of current signal patterns for sequence segments would significantly improve the accuracy of DNA methylation determination. In particular, IPM-RNN yielded the best performance among those approaches.

[0115] In some embodiments, for IPM, the length of the DNA stretch surrounding the base subjected to base modification analysis can be symmetric or asymmetric, for example, X-nt upstream and Y-nt downstream of the base can be used for base modification analysis. X can include, but is not limited to, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 100, 150, 200, 300, 400, 500, 1000, 2000, 4000, 5000, and 10000. Y can include, but is not limited to, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 100, 150, 200, 300, 400, 500, 1000, 2000, 4000, 5000, and 10000. X and Y can be the same or different.

[0116] In some embodiments, base modifications in nucleic acids will be analyzed according to embodiments in the present disclosure across different organisms, including viruses, bacteria, plants, fungi, nematodes, insects, and vertebrates (e.g., humans), etc. The most common base modification is the addition of methyl groups to different DNA bases at different positions, so-called methylation. Methylation has been found at cytosine, adenine, thymine, and guanine, such as 5mC (5-methylcytosine), 4mC (N4-methylcytosine), 5hmC (5-hydroxymethylcytosine), 5fC (5-formylcytosine), 5caC (5-carboxylcytosine), 1mA (N1-methyladenine), 3mA (N3-methyladenine), 6mA (N6-methyladenine), 7mA (N7-methyladenine), 3mC (N3-methylcytosine), 2mG (N2-methylguanine), 6mG (O6-methylguanine), 7mG (N7-methylguanine), 3mT (N3-methylthymine), and 4mT (O4-methylthymine).

[0117] In some embodiments, the integrated representation matrix of the current signal patterns may be analyzed by different statistical and / or mathematical models, including, but not limited to, linear regression, logistic regression, deep recurrent neural networks (e.g., long short-term memory, LSTM), Bayesian classifiers, hidden Markov models (HMM), linear discriminant analysis (LDA), k-means clustering, density-based spatial clustering for applications with noise (DBSCAN), random forest algorithm, and support vector machines (SVM). In yet another embodiment, natural language processing will be applied to electrical signal analysis for base modification analysis.

[0118] In some embodiments, different types of nanopores can be used, including, but not limited to, biological nanopores such as, for example, the protein alpha-hemolysin and its variants through protein engineering, programmed bacterially produced pore proteins, solid state nanopores fabricated from synthetic materials, graphene, etc.

[0119] In embodiments, these methods can be used to target multiple long DNA molecules that share homologous sequences by designing guide RNAs with reference to a reference genome, such as the human reference genome (hg19), e.g., long interspersed nuclear element (LINE) repeats. In one example, such analysis can be used to analyze circulating cell-free DNA in maternal plasma of pregnant women for detection of fetal aneuploidy (Kinde et al. PLOS One 2012;7(7):e41162. In embodiments, inactive or "dead" Cas9 (dCas9) and its associated single-stranded guide RNA (sgRNA) can be used to enrich for target long DNA without cleaving double-stranded DNA molecules. For example, the 3' end of the sgRNA can be designed with an extra universal short sequence. A biotinylated single-stranded oligonucleotide complementary to the universal short sequence can be used to capture those target long DNA molecules bound by dCas9. In another embodiment, biotinylated dCas9 protein or sgRNA, or both, can be used to facilitate enrichment.

[0120] In embodiments, size selection may be performed to enrich for long DNA fragments without limiting them to one or more specific genomic regions of interest, using approaches including, but not limited to, chemical, physical, enzymatic, gel-based, and magnetic bead-based methods, or a combination of more of such approaches.

[0121] IV. Exemplary Methods This section provides exemplary methods of using machine learning models to detect base modifications and exemplary methods of training machine learning models for the detection of base modifications.

[0122] A. Detection of Modifications FIG. 12 is a flow chart of an exemplary process 1200 associated with detecting a modification of a nucleotide in a nucleic acid molecule. The modification may include any methylation or any oxidation described herein. The oxidation may be 8-oxo-guanine. In some implementations, one or more process blocks of FIG. 12 may be performed by a system (e.g., measurement system 1400). In some implementations, one or more process blocks of FIG. 12 may be performed by another device or devices that are separate from or include the system. Additionally or alternatively, one or more process blocks of FIG. 12 may be performed by one or more components of the measurement system 1400, such as the detector 1420, logic system 1430, local memory 1435, external memory 1440, storage device 1445, and / or processor 1450.

[0123] At block 1210, an input data structure is received. The input data structure may correspond to a window of nucleotides to be sequenced in a sample nucleic acid molecule. The sample nucleic acid molecule is sequenced by measuring an electrical signal corresponding to the nucleotides. The electrical signal may be a current, a voltage, a resistance, an inductance, a capacitance, or an impedance. The sequencing may be by using a nanopore. Process 1200 may further include sequencing the sample nucleic acid using a nanopore. The nanopore may be any nanopore described herein.

[0124] The input data structure may include values ​​for several properties. The properties may include a vector including, for each nucleotide in the window, the identity of the nucleotide, the position of the nucleotide relative to the target position in the respective window, and a first segment statistic of a segment of the electrical signal corresponding to the nucleotide. The properties may include a first region statistic of the electrical signal in a region of the nucleic acid molecule equal to or larger than the window. For example, the input data structure may include an integrated representation matrix [IPM].

[0125] The identity of a nucleotide can be a base (e.g., A, T, C, or G). The base can be determined through base calling techniques using nanopore sequencing. The base calling techniques can associate segments of the electrical signal with the nucleotide. The position of the nucleotide can be the distance of the nucleotide relative to a target position. For example, if the nucleotide is one nucleotide away from the target position in one direction, the position can be +1, and if the nucleotide is one nucleotide away from the target position in the opposite direction, the position can be -1.

[0126] The first segment statistic may represent an average of a segment of the electrical signal corresponding to a nucleotide. In some embodiments, the first segment statistic may represent a variation (e.g., standard deviation) of the electrical signal of a segment of the electrical signal corresponding to a nucleotide. In embodiments, the first segment statistic may represent a normalized value of the average of the segment of the electrical signal corresponding to a nucleotide. Normalization may include rescaling the first segment statistic to be in a particular range (e.g., a range of 0 to 1). Normalization may include using a median, mean, and / or deviation for some or all of the nucleotide strands. Normalization may be any described herein, including z-score (e.g., X5).

[0127] The vector may include a second segment statistic representing the variation of a segment of the electrical signal corresponding to a nucleotide. The vector may include a third segment statistic representing a normalized value of the first segment statistic. The vector may include any combination of variables X1, X2, and X5 described herein.

[0128] The first regional statistic may represent the mean or median of the electrical signal in the region. For example, the first regional statistic may be X3. In embodiments, the first regional statistic may represent the median or average of the absolute value of the variation of the electrical signal from the mean or median of the electrical signal in the region. The variation may be a standard deviation. For example, the first regional statistic may be X4. In some embodiments, the first regional statistic may be optional.

[0129] The input data structure may further include a second region statistic representing the median or average of the absolute value of the variation of the electrical signal from the mean or median of the electrical signal within the region. For example, the second region statistic may be X4.

[0130] The first regional statistic may be the same value for different nucleotides within the window. The second regional statistic may be the same value for different nucleotides within the window. As a result, the first regional statistic and the second regional statistic may be considered as separate vectors with the first segment statistic and / or the second segment statistic. Alternatively, even if the values ​​are the same across nucleotides, the vector may also include the first regional statistic and / or the second regional statistic may be included in the vector for each nucleotide. Approaches to iterating regional statistics have been illustrated in IPM524 and IPM624.

[0131] The region may be on one strand of the sample nucleic acid molecule. In some embodiments, the region may be on two strands of the sample nucleic acid molecule. The window may include nucleotides on two strands of the sample nucleic acid molecule. The region may be the sample nucleic acid molecule. The region may include at least 5, 10, 15, 20 25, 30, 50, 100, 200, 300, 400, 500, 1k, 5k, 10k, 50k, or 1M nucleotides. In some embodiments, the region may be less than 50, 100, 200, 300, 400, 500, 1k, 5k, 10k, 50k, or 1M nucleotides. The region may be centered around the nucleotide at the target position.

[0132] The window of nucleotides may be centered around the nucleotide at the target position. In some embodiments, the window may not be centered around the nucleotide at the target position. The window may include X-nt upstream and Y-nt downstream from the nucleotide at the target position. X can include, but is not limited to, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 100, 150, 200, 300, 400, 500, 1000, 2000, 4000, 5000, and 10000. Y can include, but is not limited to, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 100, 150, 200, 300, 400, 500, 1000, 2000, 4000, 5000, and 10000. The minimum number of nucleotides in a window can be 2, 3, 4, 5, 6, 7, 8, 9, 10, 20, 30, 40, 50, 100, 200, or one more than the sum of any of the numbers of nucleotides upstream and downstream of the target position. The window can be similar to the window shown and described in FIG.

[0133] The window may comprise two strands of a nucleic acid molecule, similar to the technique illustrated in FIG.

[0134] In block 1220, an input data structure is input to the model. The model is trained by receiving a first plurality of first data structures. Each first data structure of the first plurality of data structures corresponds to a respective window of nucleotides sequenced in each nucleic acid molecule of the plurality of first nucleic acid molecules. Each of the first nucleic acid molecules is sequenced by measuring an electrical signal corresponding to the nucleotide. The modification has a known first state of the nucleotide at the target position in each window of each first nucleic acid molecule. Each first data structure includes values ​​for the same characteristics as the input data structure. The model can be any machine learning model described herein.

[0135] The model is further trained by storing a plurality of first training samples. Each first training sample includes one of the first plurality of first data structures and a first label indicating a first state of the nucleotide at the target position. In addition, the model is trained by using the plurality of first training samples to optimize parameters of the model based on the output of the model that matches or does not match the corresponding label of the first label when the first plurality of first data structures are input to the model. The output of the model specifies whether the nucleotide at the target position in each window has a modification. The training may be as described later in FIG. 13.

[0136] At block 1230, the model is used to determine whether a modification is present at a nucleotide at a target position within the window in the input data structure.

[0137] The modification state can be used for further analysis. In a sample obtained from a pregnant woman, embodiments in the present disclosure can be used to determine fetal or maternal origin of plasma DNA molecules based on methylation state. Maternal or fetal origin can be determined by genomic regions having methylation levels higher or lower than a reference value. In embodiments, the sample obtained from the pregnant woman can be acellular, for example, plasma or serum. In some embodiments, the sample nucleic acid molecule can be identified as aligning to a predetermined genomic region. The predetermined genomic region can be known to be hypermethylated or hypomethylated in the fetal or maternal genome. The method can include determining that the sample nucleic acid is of fetal or maternal origin using the modification state of the nucleotide at the target position and, optionally, the modification state of one or more other nucleotides of the sample nucleic acid molecule.

[0138] Determining whether the sample nucleic acid molecule is of fetal or maternal origin may include determining a methylation level of the sample nucleic acid molecule using the methylation state of one or more nucleotides. The methylation level of the sample nucleic acid molecule may be compared to a reference value. The reference value may be determined from the methylation levels of one or more maternal nucleic acid molecules. Comparing the methylation level of the sample nucleic acid molecule to the reference value may include determining that the methylation level of the sample nucleic acid molecule is lower than the reference value. Determining whether the sample nucleic acid molecule is of fetal or maternal origin may include determining that the sample nucleic acid molecule is of fetal origin using the comparison.

[0139] In some embodiments, the sample nucleic acid molecule can be one sample nucleic acid molecule of a plurality of sample nucleic acid molecules. The method can further include determining whether each of the plurality of sample nucleic acid molecules is of fetal or maternal origin using the methylation state. The fetal fraction can be determined using the determination of the fetal or maternal origin of the plurality of sample nucleic acid molecules.

[0140] In some embodiments, the modification state can be used to determine whether a copy number abnormality exists in the region. The modification can be methylation. The sample nucleic acid molecule can be acellular and can be obtained from a biological sample of a female subject carrying a fetus. The sample nucleic acid molecule can be a sample nucleic acid molecule of a plurality of sample nucleic acid molecules. The method can further include identifying that the plurality of sample nucleic acid molecules are aligned to a region of the fetal genome. The modification state of one or more nucleotides of each sample nucleic acid molecule of the plurality of sample nucleic acid molecules can be determined. The methylation level of the region can be determined using the methylation state of one or more nucleotides for each sample nucleic acid molecule of the plurality of sample nucleic acid molecules. The method can further include using the methylation level to determine whether a copy number abnormality exists in the region of the fetal genome. The region can be a chromosome, and the method can further include determining that a copy number abnormality exists and determining that the fetus has a chromosomal aneuploidy.

[0141] The modification may be determined to be present at one or more nucleotides. The classification of the disorder may be determined using the presence of the modification at one or more nucleotides. The classification of the disorder may include using the number of modifications. The number of modifications may be compared to a threshold value. Alternatively or additionally, the classification may include the location of the one or more modifications. The location of the one or more modifications may be determined by aligning the sequence reads of the nucleic acid molecule to a reference genome. If a modification is shown to be present at a particular location known to be correlated with the disorder, the disorder may be determined. For example, the pattern of methylation sites may be compared to a reference pattern for the disorder, and the disorder may be determined based on the comparison. A match with the reference pattern or a substantial match with the reference pattern (e.g., 80%, 90%, or 95% or more) may indicate a disorder or a high probability of the disorder. The disorder may be any pregnancy-related disorder (e.g., pre-eclampsia, intrauterine growth restriction, invasive placentation, and preterm birth).

[0142] A statistically significant number of nucleic acid molecules can be analyzed to provide accurate determination of the disorder, tissue origin, or clinically relevant DNA fraction in one or more pregnant subjects. In some embodiments, at least 1,000 nucleic acid molecules are analyzed. In other embodiments, at least 10,000 or 50,000 or 100,000 or 500,000 or 1,000,000 or 5,000,000 or more nucleic acid molecules can be analyzed. As a further example, at least 10,000 or 50,000 or 100,000 or 500,000 or 1,000,000 or 5,000,000 sequence reads can be generated.

[0143] The method can include classifying the disorder by determining that the subject has the disorder. The classification can include a level of disorder using the number of modifications and / or the site of the modifications.

[0144] The presence of modifications at one or more nucleotides can be used to determine the fetal DNA fraction, the fetal methylation profile, the maternal methylation profile, the presence of imprinted gene regions.

[0145] Process 1200 may include additional implementations, such as any single implementation or any combination of implementations, in conjunction with one or more other processes described below and / or elsewhere herein.

[0146] 12 illustrates example blocks of process 1200, in some implementations, process 1200 may include additional, fewer, different, or differently arranged blocks than those illustrated in FIG 12. Additionally or alternatively, two or more of the blocks of process 1200 may be performed in parallel.

[0147] B. Model Training FIG. 13 shows an exemplary method 1300 for detecting modifications of nucleotides in a nucleic acid molecule. The exemplary method 1300 can be a method for training a model to detect modifications. The modifications can include methylation. The methylation can include any methylation described herein. The modifications can have distinct states, such as methylated and unmethylated, and can specify the type of methylation. Thus, a nucleotide can have more than two states (classifications). The training in FIG. 13 can be used with the method 1200 in FIG. 12.

[0148] At block 1310, a plurality of first data structures are received. Various examples of data structures are described herein, e.g., in Figs. 5 and 6. Each first data structure of the first plurality of first data structures may correspond to a respective window of nucleotides sequenced in a respective nucleic acid molecule of the plurality of first nucleic acid molecules. Each window associated with the first plurality of data structures may include four or more consecutive nucleotides, including 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21 or more consecutive nucleotides. Each window may include the same number of consecutive nucleotides. The windows may overlap. Each window may include nucleotides on a first strand of the first nucleic acid molecule and nucleotides on a second strand of the first nucleic acid molecule. The first data structure may also include a value of a strand property for each nucleotide in the window. The strand property may indicate the nucleotide present or either the first strand or the second strand. A window may include the nucleotide of the second strand that is not complementary to the nucleotide at the corresponding position of the first strand.In some embodiments, all nucleotides on the second strand are complementary to the nucleotide on the first strand.In some embodiments, each window may include the nucleotide of only one strand of the first nucleic acid molecule.

[0149] The first plurality of first data structures may include 5,000-10,000, 10,000-50,000, 50,000-100,000, 100,000-200,000, 200,000-500,000, 500,000-1,000,000, or more than 1,000,000 first data structures. The plurality of first nucleic acid molecules may include at least 1,000, 10,000, 50,000, 100,000, 500,000, 1,000,000, 5,000,000, or more nucleic acid molecules. As further examples, at least 10,000 or 50,000 or 100,000 or 500,000 or 1,000,000 or 5,000,000 sequence reads can be generated.

[0150] Each of the first nucleic acid molecules is sequenced by measuring an electrical signal corresponding to a nucleotide. The electrical signal can be from nanopore sequencing.

[0151] The modification has a known first state of the nucleotide at the target position in each window of each first nucleic acid molecule. The first state can be that the modification is not present in the nucleotide or that the modification is present in the nucleotide. The modification can be known not to be present in the first nucleic acid molecule or the first nucleic acid molecule can be treated such that the modification is not present. The modification can be known to be present in the first nucleic acid molecule or the first nucleic acid molecule can be treated such that the modification is present. When the first state is a state in which the modification is not present, the modification can be absent in each window of each first nucleic acid molecule and present only at the target position. The known first state can include a methylation state of the first location of the first data structure and an unmethylated state of the second location of the first data structure. The known first state for methylation can be determined through a technique using bisulfite sequencing or using an optical signal from single molecule real-time sequencing.

[0152] The target position may be the center of each window. For windows that span an even number of nucleotides, the target position may be the position immediately upstream or downstream of the center of the window. In some embodiments, the target position may be at any other position of each window, including the first or last position. For example, if the window spans n nucleotides, from the 1st position to the nth position (either upstream or downstream) of one strand, the target position may be at any position from the 1st position to the nth position.

[0153] Each first data structure includes values ​​for a property within the window, which may be any of the properties described in block 1210.

[0154] At block 1320, a plurality of first training samples are stored, each of the first training samples including one of the first plurality of first data structures and a first label indicating a first state for a modification of a nucleotide at a target position.

[0155] In block 1330, a second plurality of second data structures is received. Block 1330 is optional. Each second data structure of the second plurality of second data structures corresponds to a respective window of nucleotides sequenced in a respective nucleic acid molecule of the plurality of second nucleic acid molecules. The second plurality of nucleic acid molecules may be the same or different from the plurality of first nucleic acid molecules. The modification has a known second state of the nucleotide at the target position in each window of each second nucleic acid molecule. The second state is a state different from the first state. For example, if the modification is present in the first state, the modification is absent in the second state, and vice versa. Each second data structure includes values ​​for the same property as the first plurality of first data structures.

[0156] At block 1340, a plurality of second training samples are stored. Block 1340 is optional. Each second training sample includes one of the second plurality of second data structures and a second label indicating a second state of the modification of the nucleotide at the target position.

[0157] In block 1350, the model is trained using a plurality of first training samples and optionally a plurality of second training samples. The training is performed by optimizing the parameters of the model based on the output of the model that matches or does not match the corresponding labels of the first label and optionally the second label when the first plurality of first data structures and optionally the second plurality of second data structures are input to the model. The output of the model specifies whether the nucleotide at the target position in each window has a modification. This method may include only a plurality of first training samples, since the model may identify outliers as a state different from the first state. The model may be a statistical model, also called a machine learning model.

[0158] In some embodiments, the output of the model may include a probability for each of a number of states. The state with the highest probability may be considered as the state.

[0159] The model may include a convolutional neural network (CNN). The CNN may include a set of convolutional filters configured to filter the first and optionally the second plurality of data structures. The filters may be any filters described herein. The number of filters in each layer may be 10-20, 20-30, 30-40, 40-50, 50-60, 60-70, 70-80, 80-90, 90-100, 100-150, 150-200, or more. The kernel size of the filters may be 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 15-20, 20-30, 30-40, or more. The CNN may include an input layer configured to receive the filtered first and optionally the filtered second plurality of data structures. The CNN may also include multiple hidden layers including multiple nodes. A first layer of the multiple hidden layers is coupled to the input layer. The CNN may further include an output layer coupled to the last layer of the multiple hidden layers and configured to output an output data structure. The output data structure may include the characteristics.

[0160] The model may include a recurrent neural network (RNN). The RNN model includes a number of long short-term memory (LSTM) units associated with a number of nucleotides in the measurement window. The number of LSTM units may be equal to the number of nucleotides in the measurement window. In some embodiments, the number of LSTM units may be less than the number of nucleotides in the measurement window. The number of LSTM units may be, but is not limited to, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 30, 40, 50, 100, 200, 300, 400, 500, 1,000, 2,000, 3,000, 4,000, 5,000, 10,000, 50,000, etc. One LSTM unit can transmit information related to the current signal features, which will be subjected to many rounds of linear or nonlinear transformations, to the next LSTM unit. Such information transmission across LSTM units is generally organized in a sequential manner (e.g., according to time steps). Such information transmission across LSTM units can be bidirectional (i.e., including temporal order and reserved temporal order). Each LSTM unit includes programmable operations such as forget gates, input gates, cell states, and output gates. Through those operations, one LSTM can decide whether the current signal information coming from the previous time step should be stored or whether it is irrelevant and can be forgotten (forget gate). One LSTM unit tries to learn new information from the input to such unit (input gate). The unit passes updated information from the current time step to the next time step (output gate). The cell state in this specification carries information with every time step. Several layers of LSTM units may be used. The number of LSTM layers can be 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 30, etc. Full connections between layers may be used. Sigmoid functions are commonly used as gate functions for the input gate, output gate, and forget gate.The output value of the sigmoid function can be between 0 and 1, determining whether there is no information flow through the gate or if it is completely flowing. A hyperbolic tangent activation function (also referred to as Tanh) can be used as an output activation function that processes the information values ​​from the output gate to form new information with values ​​between -1 and 1 that can be passed to the next LSTM unit. In some embodiments, other activation functions can be used, including but not limited to binary step functions, linear activation functions, sigmoid functions, rectified linear units, etc. The values ​​generated by the final layer of LSTM can be passed onto an output layer (i.e., a dense layer with a certain number of neurons) where each neuron is fully connected. The number of neurons in the dense layer can be, but is not limited to, 2, 3, 4, 5, 6, 7, 8, 9, 10, 20, 30, 40, 50, 100, 200, 300, 400, 500, 1000, 2000, etc. Several dense layers may be used, including but not limited to 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 20, 30, 40, 50, 100, 5000, 1000, etc. The output layer may output a methylation score, for example, based on a sigmoid activation function or a SoftMax activation function, which may be used to classify the methylation state. For example, if the methylation score is greater than 0.5, the base is determined to be methylated. Otherwise, the base is determined to be unmethylated. In some embodiments, the threshold used to classify the methylation state may be, but is not limited to, at least 0.1, 0.2, 0.3, 0.4, 0.6, 0.7, 0.8, 0.9, etc. In some embodiments, some of the neurons in the model may be dropped out to minimize the problem of overfitting. The percentage of dropped out neurons can be, but is not limited to, 1%, 5%, 10%, 15%, 20%, 25%, 30%, 40%, 50%, 60%, 70%, etc., and may vary for different layers.

[0161] The models may include supervised learning models, which 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 processing, kernel estimators, learning automata, learning classifier systems, minimum message length (decision trees, decision graphs, etc.), multiple linear subspace learning, naive Bayes classifiers, maximum entropy classifiers, conditional random fields, nearest neighbor algorithms, probabilistic approximately correct learning (PAC) learning, ripple down rules, knowledge acquisition methods, symbolic machine learning algorithms, sub-symbolic machine learning algorithms, support vector machines, minimal complexity machines (MCM), random forests, ensembles of classifiers, regular classification, data pre-processing, processing of imbalanced data sets, statistical relationship learning, or Proaftn, multi-criteria classification algorithms. The model can be a linear regression, a logistic regression, a deep recurrent neural network (e.g., long short-term memory, LSTM), a Bayesian classifier, a hidden Markov model (HMM), a linear discriminant analysis (LDA), k-means clustering, density-based spatial clustering for applications with noise (DBSCAN), a random forest algorithm, a support vector machine (SVM), or any model described herein.

[0162] As part of training the machine learning model, the parameters of the machine learning model (weights, thresholds, etc., which can be used, for example, in the activation function of a neural network) are optimized based on the training samples (training set) to provide optimized accuracy in classifying nucleotide modifications at the target positions. Various forms of optimization can be performed, such as backpropagation, empirical risk minimization, and structural risk minimization. A validation set of samples (data structure and labels) can be used to validate the accuracy of the model. Cross-validation can be performed using different parts of the training set for training and validation. The model can include multiple sub-models, thereby providing an ensemble model. The sub-models can be weaker models, but when combined, provide a more accurate final model.

[0163] V. Exemplary Systems FIG. 14 illustrates a measurement system 1400 according to an embodiment of the invention. The system shown includes a sample 1405, such as DNA molecules, in a sample holder 1410, which can be contacted with an assay 1408 to provide a signal of a physical characteristic 1415. An example of a sample holder can be a flow cell, which includes probes and / or primers of an assay, or a tube through which a droplet (with the droplet containing the assay) travels. The physical characteristic 1415 (e.g., fluorescence intensity, voltage, or current) from the sample is detected by a detector 1420. The detector 1420 can take measurements at intervals (e.g., periodic intervals) to obtain data points that constitute a data signal. In one embodiment, an analog-to-digital converter converts the analog signal from the detector to digital form multiple times. The sample holder 1410 and the detector 1420 can form an assay device, such as a sequencing device, that performs sequencing according to embodiments described herein. The data signal 1425 is transmitted from the detector 1420 to a logic system 1430. The data signal 1425 may be stored in a local memory 1435, an external memory 1440, or a storage device 1445.

[0164] The logic system 1430 may be or include a computer system, an ASIC, a microprocessor, etc. It may also include or be coupled to a display (e.g., a monitor, LED display, etc.), and a user input device (e.g., a mouse, keyboard, buttons, etc.). The logic system 1430 and other components may be part of a stand-alone or networked computer system, or may be directly attached to or incorporated into a device (e.g., a sequencing device) that includes the detector 1420 and / or the sample holder 1410. The logic system 1430 may also include software that executes on the processor 1450. The logic system 1430 may include a computer readable medium that stores instructions for controlling the system 1400 to perform any of the methods described herein. For example, the logic system 1430 may provide commands to a system that includes the sample holder 1410 such that sequencing or other physical operations are performed. Such physical operations may be performed in a particular order, e.g., reagents are added and removed in a particular order. Such physical actions may be performed by a robotic system, including, for example, a robotic arm, such that the system may be used to obtain samples and perform assays.

[0165] Any of the computer systems referred to herein may utilize any suitable number of subsystems. An example of such a subsystem is shown in FIG. 15 of computer system 10. In some embodiments, the computer system includes a single computing device and the subsystems may be components of the computing device. In other embodiments, the computer system may include multiple computing devices, each of which is a subsystem and has internal components. The computer systems may include desktop and laptop computers, tablets, mobile phones, and cloud-based systems.

[0166] The subsystems shown in FIG. 15 are interconnected via a system bus 75. Additional subsystems are shown, such as a printer 74, a keyboard 78, a storage device 79, a monitor 76 (e.g., a display screen such as LED) coupled to a display adapter 82, and others. Peripherals and input / output (I / O) devices coupled to the I / O controller 71 may be connected to the computer system by any number of means known in the art, such as an input / output (I / O) port 77 (e.g., USB, Lightning, Thunderbolt™). For example, the I / O port 77 or an external interface 81 (e.g., Ethernet, Wi-Fi, etc.) may be used to connect the computer system 10 to a wide area network such as the Internet, a mouse input device, or a scanner. The interconnection via the system bus 75 allows the central processor 73 to communicate with each subsystem and control the execution of instructions from the system memory 72 or storage device 79 (e.g., a fixed disk such as a hard drive or optical disk), and the exchange of information between the subsystems. The system memory 72 and / or storage device 79 may embody a computer readable medium. Another subsystem is a data collection device 85, such as a camera, microphone, and accelerometer. Any of the data mentioned herein may be output from one component to another and may be output to a user.

[0167] A computer system may include multiple identical components or subsystems connected together, for example, by an external interface 81, by an internal interface, or through a removable storage device that can connect one component to another. In some embodiments, computer systems, subsystems, or devices may communicate over a network. In such an example, one computer may be considered a client and another computer a server, each of which may be part of the same computer system. The client and server may each include multiple systems, subsystems, or components.

[0168] Aspects of the embodiments can be implemented using hardware circuitry in the form of control logic (e.g., application specific integrated circuits or field programmable gate arrays) and / or using computer software with general purpose programmable processors in a modular or integrated manner. As used herein, a processor may include a single core processor, a multi-core processor on the same integrated chip, or multiple processing units on a single circuit board or networked, as well as dedicated hardware. Based on this disclosure and the teachings provided herein, those skilled in the art will recognize and understand other ways and / or methods of implementing embodiments of the present invention using hardware and combinations of hardware and software.

[0169] Any of the software components or functions described in this application may be implemented as software code executed by a processor using any suitable computer language, such as, for example, Java, C, C++, C#, Objective-C, Swift, or a scripting language, such as, for example, Perl or Python, using conventional or object-oriented techniques. The software code may be stored as a series of instructions or commands on a computer-readable medium for storage and / or transmission. Suitable non-transitory computer-readable media may include random access memory (RAM), read only memory (ROM), magnetic media (such as a hard drive or a floppy disk), or optical media (such as a compact disk (CD) or a digital versatile disk (DVD), or a Blu-ray disk, a flash memory, or the like. The computer-readable medium may be any combination of such storage or transmission devices.

[0170] Such programs may also be coded and transmitted using carrier signals adapted for transmission over wired, optical, and / or wireless networks according to various protocols, including the Internet. Thus, computer-readable media may be created using data signals coded with such programs. Computer-readable media coded with program code may be packaged with a compatible device or provided separately from other devices (e.g., via Internet download). Any such computer-readable medium may be present on or within a single computer product (e.g., a hard drive, CD, or an entire computer system), or may be present on or within different computer products within a system or network. A computer system may include a monitor, printer, or other suitable display for providing a user with any of the results described herein.

[0171] Any of the methods described herein may be implemented in whole or in part using a computer system including one or more processors that can be configured to perform the steps. Thus, embodiments may be directed to a computer system configured to perform the steps of any of the methods described herein, with potentially different components performing each step or each group of steps. The steps of the methods herein, although presented as numbered steps, may be performed simultaneously or at different times or in different orders. Furthermore, portions of these steps may be used with portions of other steps from other methods. Also, all or portions of a step may be optional. Furthermore, any of the steps of any of the methods may be implemented in a module, unit, circuit, or other means of a system for performing these steps.

[0172] The specific details of the particular embodiments may be combined in any suitable manner without departing from the spirit and scope of the embodiments of the invention, however, other embodiments of the invention may be directed to particular embodiments relating to each individual aspect, or particular combinations of these individual aspects.

[0173] The above description of exemplary embodiments of the present disclosure has been presented for purposes of illustration and description. It is not intended to be exhaustive or to limit the disclosure to the precise form described, and many modifications and variations are possible in light of the above teachings.

[0174] The descriptions "a," "an," or "the" are intended to mean "one or more," unless specifically indicated to the contrary. The use of "or" is intended to mean "inclusive or" rather than "exclusive or," unless specifically indicated to the contrary. A reference to a "first" element does not necessarily require that a second element be provided. Furthermore, a reference to a "first" or "second" element does not limit the referenced element to a particular location unless expressly stated. The term "based on" is intended to mean "based at least in part on."

[0175] All patents, patent applications, publications, and specifications mentioned herein are incorporated by reference in their entirety for all purposes. None is admitted to be prior art.

Claims

1. 1. A method for detecting a nucleotide modification in a nucleic acid molecule, comprising: receiving an input data structure, the input data structure corresponding to a window of sequenced nucleotides in a sample nucleic acid molecule, the sample nucleic acid molecule being sequenced by measuring electrical signals corresponding to the nucleotides, the input data structure comprising values ​​for the following properties: For each nucleotide in the window the identity of said nucleotide, the position of the nucleotide relative to a target position within each window; and a vector including first segment statistics for segments of the electrical signal corresponding to the nucleotides; inputting the input data structure into a model, the model comprising: receiving a first plurality of first data structures, each first data structure of said first plurality of first data structures corresponding to a respective window of nucleotides sequenced in a respective nucleic acid molecule of a plurality of first nucleic acid molecules, each of said first nucleic acid molecules being sequenced by measuring said electrical signal corresponding to said nucleotide, said modification having a known first state of a nucleotide at a target position in each window of each first nucleic acid molecule, each first data structure including values ​​for the same property as said input data structure; storing a plurality of first training samples, each including one of the first plurality of first data structures and a first label indicating the first state of the nucleotide at the target position; and optimizing parameters of the model based on outputs of the model that match or do not match corresponding labels of the first labels using the first training samples when the first plurality of first data structures are input to the model, the outputs of the model specifying whether the nucleotide at the target position in the respective window has the modification; and using the model to determine whether the modification is present at a nucleotide at the target position within the window of the input data structure.

2. 2. The method of claim 1, wherein the first segment statistic represents an average of the segment of the electrical signal corresponding to the nucleotide.

3. 2. The method of claim 1, wherein the first segment statistic represents a variance of the electrical signal for the segment of the electrical signal corresponding to the nucleotide.

4. 2. The method of claim 1, wherein the first segment statistic represents an average normalized value of the segment of the electrical signal corresponding to the nucleotide.

5. 2. The method of claim 1, wherein the vector comprises a second segment statistic representing the variance of the segment of the electrical signal corresponding to the nucleotide.

6. 2. The method of claim 1, wherein the vector comprises a second segment statistic representing an average normalized value of the segment of the electrical signal corresponding to the nucleotide.

7. the vector includes a second segment statistic representative of the variance of the segment of the electrical signal corresponding to the nucleotide; The method of claim 2 , wherein the vector includes a third segment statistic representing a normalized value of the first segment statistic.

8. 2. The method of claim 1, wherein the input data structure comprises values ​​for a first regional statistic of the electrical signal within a region of the nucleic acid molecule that is equal to or greater than the window.

9. The method of claim 8 , wherein the first regional statistic represents a mean or median of the electrical signal within the region.

10. The method of claim 8 , wherein the first regional statistic represents a median or average of absolute values ​​of the variation of the electrical signal from the mean or median of the electrical signal within the region.

11. 10. The method of claim 9, wherein the input data structure further comprises a second regional statistic representing a median or mean of the absolute value of the variation of the electrical signal from the mean or median of the electrical signal within the region.

12. The method of claim 8 , wherein the region is on one strand of the sample nucleic acid molecule.

13. 9. The method of claim 8, wherein the region is or comprises at least 5, 10, 15, 20, 25, 30, 50, 100, 200, 300, 400, 500, or 1k, 5k, 10k, 50k, or 1M nucleotides of the sample nucleic acid molecule.

14. The method of claim 8, wherein the region is centered around the nucleotide.

15. The method of claim 1 , wherein the window comprises nucleotides on both strands of the sample nucleic acid molecule.

16. The method of claim 1 , wherein the modification is methylation or oxidation.

17. The method of claim 1 , wherein the electrical signal is a current, a voltage, a resistance, an inductance, a capacitance, or an impedance.

18. 10. The method of claim 1, further comprising sequencing the sample nucleic acid molecule using a nanopore.

19. the modification is methylation, the sample nucleic acid molecule is acellular and obtained from a biological sample of a female subject carrying a fetus; The method further comprising:

2. The method of claim 1, further comprising using the modification state of the nucleotide at the target position to determine whether the sample nucleic acid molecule is of fetal or maternal origin, wherein the modification state is whether the modification is present, and optionally the modification state of one or more other nucleotides of the sample nucleic acid molecule.

20. Determining whether the sample nucleic acid molecules are of fetal or maternal origin determining a methylation level of the sample nucleic acid molecule using the modification state of the one or more nucleotides; and comparing the methylation level of the sample nucleic acid molecule to a reference value.

21. 21. The method of claim 20, wherein the reference value is determined from the methylation level of one or more maternal nucleic acid molecules.

22. comparing the methylation level of the sample nucleic acid molecule to the reference value comprises determining that the methylation level of the sample nucleic acid molecule is lower than the reference value; 21. The method of claim 20, wherein determining whether the sample nucleic acid molecule is of fetal or maternal origin comprises determining that the sample nucleic acid molecule is of fetal origin using the comparison.

23. 20. The method of claim 19, further comprising identifying the sample nucleic acid molecule as aligning to a predetermined genomic region.

24. the sample nucleic acid molecule is a sample nucleic acid molecule of a plurality of sample nucleic acid molecules; The method further comprising: determining whether each of said plurality of sample nucleic acid molecules is of fetal or maternal origin using said modification state; 20. The method of claim 19, further comprising: using said determination of fetal or maternal origin of said plurality of sample nucleic acid molecules to determine a fetal fraction.

25. the modification is methylation, the sample nucleic acid molecule is acellular and obtained from a biological sample of a female subject carrying a fetus; the sample nucleic acid molecule is a sample nucleic acid molecule of a plurality of sample nucleic acid molecules; The method further comprising: identifying the plurality of sample nucleic acid molecules that align to a region of a fetal genome; determining the modification state of one or more nucleotides of each sample nucleic acid molecule of the plurality of sample nucleic acid molecules; determining a methylation level of the region using the modification state of the one or more nucleotides for each sample nucleic acid molecule of the plurality of sample nucleic acid molecules; 2. The method of claim 1, further comprising using the methylation level to determine whether a copy number abnormality is present in the region of the fetal genome.

26. 1. A method for detecting a nucleotide modification in a nucleic acid molecule, comprising: receiving a first plurality of first data structures, each first data structure of said first plurality of first data structures corresponding to a respective window of nucleotides sequenced in a respective nucleic acid molecule of a plurality of first nucleic acid molecules, each of said first nucleic acid molecules being sequenced by measuring an electrical signal corresponding to said nucleotides, said modification having a known first state of nucleotides at target positions in each window of each first nucleic acid molecule, each first data structure comprising values ​​for the following properties: For each nucleotide in the window the identity of said nucleotide, the position of the nucleotide relative to a target position within each window; and a vector including first segment statistics for segments of the electrical signal corresponding to the nucleotides; storing a plurality of first training samples, each including one of the first plurality of first data structures and a first label indicating the first state for the modification of the nucleotide at the target position; training the model using the first plurality of first training samples when the first plurality of first data structures are input to a model by optimizing parameters of the model based on an output of the model that matches or does not match a corresponding label of the first label, wherein the output of the model specifies whether the nucleotide at the target position in the respective window has the modification or not.

27. receiving a second plurality of second data structures, each second data structure of the second plurality of second data structures corresponding to a respective window of nucleotides sequenced in a respective nucleic acid molecule of a plurality of second nucleic acid molecules, the modification having a known second state of nucleotides at target positions within each window of each second nucleic acid molecule, each second data structure including values ​​for the same properties as the first plurality of first data structures; storing a plurality of second training samples, each comprising one of the second plurality of second data structures and a second label indicating the second state of the nucleotide at the target position; In training, the first state or the second state is that the modification is present, and the other state is that the modification is absent; 27. The method of claim 26, further comprising: optimizing parameters of the model based on outputs of the model that match or do not match corresponding labels of the second labels when the second plurality of second data structures are input to the model, using the second plurality of second training samples.

28. 28. The method of claim 27, wherein the plurality of first nucleic acid molecules is the same as the plurality of second nucleic acid molecules.

29. each window associated with the first plurality of first data structures includes a nucleotide on a first strand of the first nucleic acid molecule and a nucleotide on a second strand of the first nucleic acid molecule; 27. The method of claim 26, wherein each first data structure further comprises, for each nucleotide in the window, a value of a strand property, the strand property indicating that the nucleotide is present in either the first strand or the second strand.

30. 27. The method of claim 26, wherein the modification comprises methylation of the nucleotide at the target position.

31. 31. The method of claim 30, wherein the known first state comprises a methylation state for a first location of the first data structure and an unmethylation state for a second location of the first data structure.

32. 27. The method of claim 26, wherein the first segment statistic represents an average of the segment of the electrical signal corresponding to the nucleotide.

33. 27. The method of claim 26, wherein the first segment statistic represents a variance of the electrical signal for the segment of the electrical signal corresponding to the nucleotide.

34. 27. The method of claim 26, wherein the first segment statistic represents an average normalized value of the segment of the electrical signal corresponding to the nucleotide.

35. 27. The method of claim 26, wherein the vector comprises a second segment statistic representing the variance of the segment of the electrical signal corresponding to the nucleotide.

36. 27. The method of claim 26, wherein the vector comprises a second segment statistic representing an average normalized value of the segments of the electrical signal corresponding to the nucleotide.

37. the vector includes a second segment statistic representative of the variance of the segment of the electrical signal corresponding to the nucleotide; 33. The method of claim 32, wherein the vector includes a third segment statistic representing a normalized value of the first segment statistic.

38. 27. The method of claim 26, wherein each first data structure includes a value for a first regional statistic of the electrical signal within a region of the respective nucleic acid molecule that is equal to or greater than the window.

39. 40. The method of claim 38, wherein the first regional statistic represents a mean or median of the electrical signal within the region.

40. 40. The method of claim 38, wherein the first regional statistic represents a median or average of absolute values ​​of the variation of the electrical signal from the mean or median of the electrical signal within the region.

41. 40. The method of claim 39, wherein the first data structure further comprises a second regional statistic representing a median or mean of the absolute value of the variation of the electrical signal from the mean or median of the electrical signal within the region.

42. 39. The method of claim 38, wherein the region is on one strand of each of the nucleic acid molecules.

43. 39. The method of claim 38, wherein said regions comprise or are at least 5, 10, 15, 20, 25, 30, 50, 100, 200, 300, 400, 500, or 1k, 5k, 10k, 50k, or 1M nucleotides of said respective nucleic acid molecule.

44. 39. The method of claim 38, wherein the region is centered around the nucleotide.

45. 27. The method of claim 26, wherein the window comprises nucleotides on the two strands of the respective nucleic acid molecule.

46. A computer product comprising a non-transitory computer readable medium storing a number of instructions, which when executed controls a computer system to perform the method of any one of claims 1 to 45.

47. 1. A system comprising:

47. A computer product according to claim 46; and one or more processors for executing instructions stored on the computer-readable medium.

48. A system comprising means for carrying out a method according to any one of claims 1 to 45.

49. A system comprising one or more processors configured to implement a method according to any one of claims 1 to 45.

50. A system comprising modules each performing a method according to any one of claims 1 to 45.

Images