Non-transitory computer-readable storage medium and anomaly fragment and cancer detection

Abstract

A non-transitory computer readable storage medium and anomalous fragment and cancer detection is performed by accessing a model obtained by a training process using a set of cancer fragments from a cancer patient and a set of non-cancer fragments from a healthy subject, wherein each of the cancer fragment set and the non-cancer fragment set comprises a plurality of training fragments; determining a low methylation score and a high methylation score of a CpG site in the training fragment, generating a feature vector based on the sorting of the training fragment, and applying the model to a test feature vector corresponding to a test object, to predict whether a subject suffers from cancer and to predict whether a test fragment from a test subject suspected to suffer from cancer has an abnormal methylation pattern.

Description

[0001] Related Applications

[0002] This application is a divisional application of application number 201980023723.4, application date March 13, 2019, and invention name “Abnormal fragment detection and classification”. Background Art Technical field:

[0003] DNA methylation plays an important role in regulating gene expression. Abnormal DNA methylation is associated with many disease processes, including cancer. DNA methylation spectra using methylation sequencing (e.g., whole genome bisulfite sequencing (WGBS)) are increasingly being considered as a valuable diagnostic tool for detecting, diagnosing and / or monitoring cancer. For example, specific patterns of differentially methylated regions and / or allele-specific methylation patterns can be used as molecular markers for non-invasive diagnosis using circulating cell-free DNA. However, there is still a need in the art for improved methods for analyzing methylation sequencing data from cell-free DNA to detect, diagnose and / or monitor diseases such as cancer. Summary of the invention

[0004] Early detection of cancer in subjects is important because it allows for earlier treatment and therefore a greater chance of survival. Sequencing of cell-free DNA (cfDNA) fragments and analysis of the methylation status of various dinucleotides of cytosine and guanine (called CpG sites) in the fragments can help understand whether the subject suffers from cancer. To this end, this description includes a method for analyzing the methylation status of CpG sites of cfDNA fragments. Specifically, the present disclosure provides a method for identifying a cfDNA fragment that has or may have an abnormal methylation pattern. In individuals who do not suffer from cancer, fragments that occur at high frequencies may not necessarily produce highly recognizable features for distinguishing cancer states. Therefore, relative to cfDNA fragments from a healthy sample (e.g., a subject who does not suffer from cancer), the identification of an abnormal methylation pattern is important for the selection of cfDNA fragments, and the cfDNA fragments can indicate the detection of cancer-specific methylation patterns with low interference. In low-interference regions, cfDNA fragments derived from genomic regions can be selected, which are most useful in identifying a cancer patient and a healthy subject, or subjects with other health conditions. A classifier trained on methylation sequencing data obtained from a subject with cancer and / or methylation sequencing data from a subject without cancer can be used to perform a comparison between a cancer patient and a healthy subject. Validation data are further provided, which demonstrate that analysis of aberrantly methylated cfDNA fragments identified using the methods described herein can be used to detect cancer with high sensitivity and specificity.

[0005] In one embodiment, a test sample comprising a plurality of cfDNA fragments is obtained from a subject in a control group. The plurality of cfDNA fragments in the test sample are processed to convert unmethylated cytosine to uracil, the cfDNA fragments are sequenced, and compared with a reference genome to identify the methylation status of each CpG site in a plurality of CpG sites. An analysis system establishes a data structure that counts the number of methylated and unmethylated control group fragments from a particular methylated string of some CpG sites for each identified CpG site in the reference genome, starting from the CpG site.

[0006] The analysis system creates a methylation state vector for each sequenced fragment, wherein the methylation state vector includes CpG sites in the fragment and their methylation states, e.g., methylated, unmethylated, or uncertain. For each fragment, the analysis system uses a probability analysis and a control group data structure to identify unexpected cases of a given fragment (or portion thereof) observed, wherein the given fragment (or portion thereof) has the observed methylation state at the CpG site in the fragment. In a specific embodiment, such a probability analysis lists alternative possibilities of methylation state vectors having a same length (at site) and position as a given fragment (or portion thereof) in the reference genome, and uses counts from the data structure to determine the probability of each of the possibilities. The analysis system can use a Markov chain probability analysis (along with a maximal order of a given Markov chain) to model the probabilities of various possibilities for each of the methylation state vectors. After calculating the probabilities of various possibilities of methylation state carriers, the analysis system generates a p-value score for the fragment by summing the probabilities for the possibilities of methylation state carriers whose probability is less than the probability of matching the possibility of the test methylation state carrier. The analysis system compares the generated p-value to a threshold value to identify abnormally methylated cfDNA fragments (also referred to herein as fragments with abnormal methylation patterns) relative to the control group.

[0007] In addition to the analysis system, a classifier also helps classify a subject as having cancer or not having cancer based on a probability. The classifier is trained on methylation sequencing data obtained from subjects with cancer and / or methylation sequencing data obtained from subjects without cancer. After sequencing each sequenced cfDNA fragment and generating a methylation status vector, the classifier is trained using cfDNA fragments identified as hypomethylated or hypermethylated compared to a healthy control group. As used herein, a "hypomethylated" cfDNA fragment can be defined as a fragment having at least 5 CpG sites and at least 80% of the CpG sites are not methylated. Similarly, a "hypermethylated" cfDNA fragment can be defined as a fragment having at least 5 CpG sites and at least 80% of the CpG sites are methylated. Afterwards, the classifier is run at each CpG site in the genome and calculates a low methylation score and a high methylation score. The calculation method of the two scores is similar. For the low methylation score, the classifier calculates the ratio of cancer fragments that are considered to be low methylated to all fragments, cancer and non-cancer, including the current CpG site. Similarly, the ratio of cancer fragments that are considered to be high methylated to all fragments that are considered to be high methylated is used to calculate the high methylation score of each CpG site.

[0008] Now, the classifier obtains a subject from the training group along with its multiple cfDNA fragments, and sequences the fragments to generate methylation state vectors. Using each methylation state vector for the subject, the classifier calculates a total high methylation score and a total low methylation score. Each total score is calculated based on the low methylation score and high methylation score of each CpG site. Afterwards, the classifier ranks the methylation state vectors of the subject by the subject's total low methylation score, and also ranks the subject by the subject's total high methylation score. Using the two rankings, the classifier generates a feature vector for the subject by selecting a score from the ranking. The classifier is then trained to distinguish the feature vector corresponding to the non-cancer training group and the feature vector region corresponding to the cancer training group. In one embodiment, the classifier uses L2 regularized kernel logistic regression with a Gaussian radial basis function kernel (RBF kernel).

[0009] Accordingly, in one aspect, the present disclosure provides a method for detecting an abnormal methylation pattern in a cell-free deoxyribonucleic acid (cfDNA) sample fragment, the method comprising: accessing a data structure, the data structure comprising: counts of strings of multiple CpG sites in a reference genome, and the individual methylation states of the multiple CpG sites from a training fragment set; generating a sample state vector for a sample fragment, the sample state vector comprising: a sample genomic position in the reference genome and a methylation state of each CpG site in the sample fragment, each methylation state being determined to be methylated or unmethylated; listing multiple possibilities of multiple methylation states having the same length as the sample state vector from the sample genomic position; for each possibility, calculating a probability by accessing multiple counts stored in the data structure; determining the possibility that matches the sample state vector, and correspondingly using the calculated probability as a sample probability; based on the sample probability, generating a score for the sample fragment of the sample state vector relative to the training fragment set; and determining whether the sample fragment has an abnormal methylation pattern based on the generated score.

[0010] In some embodiments, the string of each CpG site includes: the methylation state of each CpG site at a plurality of genomic positions within the reference genome, wherein each methylation state is determined to be methylated or unmethylated.

[0011] In some embodiments, the method further includes: constructing the data structure from the training segment set, and includes: generating a training state vector for each training segment in the training segment set, the training state vector including: a known genomic position within the reference genome and the methylation state of each CpG site in the training segment, each methylation state being determined as methylated or unmethylated; determining a plurality of strings, wherein each string is part of the training state vector, determining a count for each string from the plurality of training state vectors; and storing a plurality of counts for each string in the data structure.

[0012] In some embodiments, the step of determining whether the sample fragment has an abnormal methylation pattern based on the generated score further includes: determining whether the score generated for the sample fragment is below a threshold score, wherein the threshold score indicates a confidence level that the sample fragment has an abnormal methylation pattern. In some embodiments, the threshold score is 0.1 or less.

[0013] In some embodiments, the set of training fragments comprises: training fragments from one or more healthy subjects, wherein the one or more healthy subjects lack a specific medical disease, and wherein the sample fragment is determined to be abnormally methylated relative to the set of training fragments from the one or more healthy subjects.

[0014] In some embodiments, generating the score for the sample fragment includes: for a plurality of methylation state possibilities, identifying a calculated probability less than the sample probability; and generating the score for the sample fragment by adding all of the identified probabilities to the sample probability. In some embodiments, for each of the possibilities, the step of accessing a plurality of counts stored in the data structure to calculate a probability includes: for each of a plurality of conditional elements, wherein each conditional element is a conditional probability that considers a subset of CpG sites in the possibility, calculating a Markov chain probability of an order having a plurality of counts stored in the data structure by the following steps, the Markov chain probability of the order having a plurality of counts stored in the data structure, the steps including: identifying a first count of the number of strings that match the conditional element; identifying a second count of the number of strings that match the previous methylation state of the conditional element at an integer length; and calculating the Markov chain probability by dividing the first count by the second count. In some embodiments, the order is selected from the group consisting of 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, and 15. In some embodiments, calculating a Markov chain probability of an order, the Markov chain probability of the order having a plurality of counts stored in the data structure further comprises: implementing a smoothing algorithm.

[0015] In some embodiments, the sample carrier is divided into multiple windows, including a first window and a second window, wherein the first window and the second window are two different parts of the sample fragment; wherein the step of determining the possibility of matching the sample state carrier and correspondingly taking the calculated probability as the sample probability includes: identifying a first possibility with a first sample probability matching the first window, and identifying the second possibility with a second sample probability matching the second window; and wherein the generated score is based on one of the first sample probability and the second sample probability.

[0016] In some embodiments, the method further comprises: filtering the plurality of sample fragments based on the scores generated for the respective sample fragments, thereby resulting in a subset of the plurality of sample fragments having a plurality of abnormal methylation patterns.

[0017] In some embodiments, the method further comprises: when the sample fragment includes at least a threshold number of multiple CpG sites and the multiple CpG sites have more than a threshold percentage of being methylated, the sample fragment is identified as being hypermethylated. In some embodiments, the threshold number of multiple CpG sites is 5 or more CpG sites, and the threshold percentage of multiple CpG sites that are methylated is 80% or greater. In some embodiments, the method further comprises: when the sample fragment includes at least a threshold number of multiple CpG sites and the multiple CpG sites have more than a threshold percentage of being unmethylated, the sample fragment is identified as being hypomethylated. In some embodiments, the threshold number of multiple CpG sites is 5 or more CpG sites, and the threshold percentage of multiple CpG sites that are not methylated is 80% or greater.

[0018] In some embodiments, the method further comprises: applying the sample state carrier to a classifier, the classifier being trained using a set of cancer training segments from one or more subjects suffering from cancer and a set of non-cancer training segments from one or more subjects not suffering from cancer, wherein the classifier can be used to determine whether the sample segment is from a subject suffering from cancer. In some embodiments, applying the sample state carrier to the classifier generates at least one of a cancer probability and a non-cancer probability. In some embodiments, the method further comprises: generating a cancer state score based on at least one of the cancer probability and the non-cancer probability.

[0019] On the other hand, the present disclosure provides a method for determining whether a test subject has cancer, the method comprising: accessing a model obtained by a training process, the training process using a cancer fragment set from one or more training subjects suffering from cancer and a non-cancer fragment set from one or more training subjects not suffering from cancer, wherein both the cancer fragment set and the non-cancer fragment set include multiple training fragments, wherein the training process comprises: for each training fragment, determining whether the training fragment is hypomethylated or hypermethylated, wherein each hypomethylated and hypermethylated training fragment respectively includes at least a threshold number of multiple CpG sites, and at least a threshold percentage of the multiple CpG sites are unmethylated or methylated, for each CpG site in a reference genome: quantifying a count of multiple hypomethylated training fragments overlapping with the CpG site and a count of multiple hypermethylated training fragments overlapping with the CpG site; and generating a prediction model based on the counts of the multiple hypomethylated training fragments and the multiple hypermethylated training fragments. Generate a low methylation score and a high methylation score; for each training segment, generate an aggregate low methylation score based on the low methylation scores of the multiple CpG sites in the training segment, and generate an aggregate high methylation score based on the high methylation scores of the multiple CpG sites in the training segment; for each training subject: rank the multiple training segments based on the aggregate low methylation scores, and rank the multiple training segments based on the aggregate high methylation scores; and generate a feature vector based on the ranking of the multiple training segments; obtain multiple training feature vectors for one or more training subjects not suffering from cancer, and obtain multiple training feature vectors for one or more training subjects suffering from cancer; and use the multiple feature vectors for the one or more training subjects not suffering from cancer and the multiple feature vectors for the one or more training subjects suffering from cancer to train the model; and apply the model to a test feature vector corresponding to the test subject to determine whether the test subject suffers from cancer.

[0020] In some embodiments, the threshold number is 5 or greater. In some embodiments, the threshold percentage is 80% or greater.

[0021] In some embodiments, for each CpG site in a reference genome, the step of quantifying a count of multiple hypomethylated training fragments overlapping with the CpG site and a count of multiple hypermethylated training fragments overlapping with the CpG site further includes: quantifying a cancer count of multiple hypomethylated training fragments from the one or more training subjects suffering from cancer overlapping with the CpG site and a non-cancer count of multiple hypomethylated training fragments from the one or more training subjects not suffering from cancer overlapping with the CpG site; and quantifying a cancer count of multiple hypermethylated training fragments from the one or more training subjects suffering from cancer overlapping with the CpG site and a non-cancer count of multiple hypermethylated training fragments from the one or more training subjects not suffering from cancer overlapping with the CpG site.

[0022] In some embodiments, for each CpG site in a reference genome, the step of generating a low methylation score and a high methylation score based on the counts of multiple low methylation training segments and multiple high methylation training segments further includes: for generating the low methylation score, calculating a low methylation ratio, the low methylation ratio is the ratio of the cancer counts of multiple low methylation training segments compared to the low methylation sum of the cancer counts of multiple low methylation training segments and the non-cancer counts of multiple low methylation training segments; and for generating the high methylation score, calculating a high methylation ratio, the high methylation ratio is the ratio of the cancer counts of multiple high methylation training segments compared to the high methylation sum of the cancer counts of multiple high methylation training segments and the non-cancer counts of multiple high methylation training segments. In some embodiments, the low methylation ratio and the high methylation ratio are further calculated using a smoothing algorithm.

[0023] In some embodiments, for each CpG site in a reference genome, the step of generating a low methylation score and a high methylation score based on the counts of multiple low methylation training segments and multiple high methylation training segments further comprises: for generating the low methylation score, calculating a low methylation log ratio, the low methylation log ratio is the log ratio of the cancer counts of multiple low methylation training segments to the non-cancer counts of multiple low methylation training segments; and for generating the high methylation score, calculating a high methylation log ratio, the high methylation log ratio is the log ratio of the cancer counts of multiple high methylation training segments to the non-cancer counts of multiple high methylation training segments. In some embodiments, the low methylation ratio and the high methylation ratio are further calculated using a smoothing algorithm. In some embodiments, for each training fragment, the step of generating an aggregate low methylation score based on the low methylation scores of the multiple CpG sites in the training fragment, and generating an aggregate high methylation score based on the high methylation scores of the multiple CpG sites in the training fragment further includes: identifying a maximum low methylation score of the multiple CpG sites in the training fragment as the aggregate low methylation score, and identifying a maximum high methylation score of the multiple CpG sites in the training fragment as the aggregate high methylation score.

[0024] In some embodiments, for each training subject, the step of generating a training feature vector based on the ranking of the multiple training fragments further includes: identifying a plurality of aggregated low methylation scores from the ranking, and identifying a plurality of aggregated high methylation scores from the ranking, and generating a training feature vector, wherein the training feature vector includes the plurality of aggregated low methylation scores and the plurality of high methylation scores.

[0025] In some embodiments, the step of training the model using the multiple training feature vectors for the one or more training subjects not suffering from cancer and the multiple training feature vectors for the one or more training subjects suffering from cancer is performed by training through a non-linear classifier.

[0026] In some embodiments, for each training subject, the training feature vector is normalized by an average length of multiple training fragments of the training subject. In some embodiments, the method further includes the step of: obtaining the test feature vector corresponding to the test subject, wherein the step of obtaining the test feature vector includes: obtaining multiple sequence reads of a test fragment set from the test subject; for each test fragment, determining whether the test fragment is hypomethylated or hypermethylated, wherein each of the multiple hypomethylated and hypermethylated test fragments respectively includes at least a threshold number of multiple CpG sites, and at least a threshold percentage of the multiple CpG sites are unmethylated or methylated, and for each CpG site in a reference genome: counting a number of hypomethylated test fragments overlapping the CpG site and a number of hypermethylated test fragments overlapping the CpG site; quantifying a count of multiple highly methylated test fragments; and generating a low methylation score and a high methylation score based on the counts of multiple low methylation test fragments and multiple high methylation test fragments; for each test fragment, generating an aggregate low methylation score based on the low methylation scores of the multiple CpG sites in the test fragment, and generating an aggregate high methylation score based on the high methylation scores of the multiple CpG sites in the test fragment; for the test object, ranking the multiple test fragments based on the aggregate low methylation scores, and ranking the multiple test fragments based on the aggregate high methylation scores; and generating the test feature vector based on the ranking of the multiple test fragments.

[0027] In some embodiments, the step of applying the model to the test feature carrier of the test subject to determine whether the test subject has cancer includes: generating a cancer probability for the test subject based on the model; and comparing the cancer probability with a threshold probability to determine whether the test subject has cancer.

[0028] In some embodiments, the diagnostic model comprises a kernel logistic regression classifier.

[0029] In another aspect, the present disclosure provides a method for determining whether a test subject suffers from cancer, the method comprising the steps of: accessing a model obtained through a training process, the training process using a cancer training segment set from one or more training subjects suffering from cancer and a non-cancer training segment set from one or more training subjects not suffering from cancer, wherein both the cancer training segment set and the non-cancer training segment set include a plurality of training segments, wherein the training process comprises: for each training segment, determining whether the training segment is hypomethylated or hypermethylated, wherein each of the hypomethylated and hypermethylated training segments The invention relates to a method for preparing a cancer-free model comprising: respectively including at least a threshold number of multiple CpG sites, and at least a threshold percentage of the multiple CpG sites being unmethylated or methylated, generating a training feature vector for each training subject based on the multiple low-methylated training fragments and the multiple high-methylated training fragments; and using the multiple training feature vectors from the one or more training subjects not suffering from cancer and the multiple feature vectors from the one or more training subjects suffering from cancer to train the model; and applying the model to a test feature vector corresponding to the test subject to determine whether the test subject suffers from cancer.

[0030] In some embodiments, for each training subject, the step of generating the training feature vector includes: for each CpG site in a reference genome: quantifying a count of multiple hypomethylated training fragments overlapping the CpG site and a count of multiple hypermethylated training fragments overlapping the CpG site; and generating a hypomethylation score and a hypermethylation score based on the counts of multiple hypomethylated training fragments and multiple hypermethylated training fragments; for each training fragment of the training subject, generating an aggregate hypomethylation score based on the hypomethylation scores of the multiple CpG sites in the training fragment, and generating an aggregate hypermethylation score based on the hypermethylation scores of the multiple CpG sites in the training fragment; and ranking the multiple training fragments of the training subject based on the aggregate hypomethylation scores, and ranking the multiple training fragments of the training subject based on the aggregate hypermethylation scores, wherein the training feature vector for the training subject is the ranking based on the aggregate hypomethylation scores and the ranking based on the aggregate hypermethylation scores.

[0031] In some embodiments, the method further comprises the step of: obtaining the test feature carrier corresponding to the test object, wherein the step of obtaining the test feature carrier comprises: obtaining a plurality of sequence reads of a test fragment set from the test object; for each test fragment, determining whether the test fragment is hypomethylated or hypermethylated, wherein each of the plurality of hypomethylated and hypermethylated test fragments respectively comprises at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites are unmethylated or methylated, and for each CpG site in a reference genome: calculating a count of a plurality of hypomethylated test fragments overlapping with the CpG site and a count of a plurality of hypermethylated test fragments overlapping with the CpG site; quantify a count of multiple hypermethylated test fragments; and generate a low methylation score and a high methylation score based on the counts of multiple low methylation test fragments and multiple high methylation test fragments; for each test fragment, generate a total low methylation score based on the low methylation scores of the multiple CpG sites in the test fragment, and generate a total high methylation score based on the high methylation scores of the multiple CpG sites in the test fragment; for the test object, rank multiple test fragments based on the total low methylation score, and rank multiple test fragments based on the total high methylation score; and generate the test feature vector based on the ranking of the multiple test fragments. In some embodiments, the diagnostic model includes a kernel logistic regression classifier.

[0032] In one aspect, the present disclosure provides a method for predicting whether a test fragment from a test subject suspected of having cancer has an abnormal methylation pattern, the method comprising the steps of: accessing a data structure, the data structure comprising: counts of strings of multiple CpG sites in a reference genome, and individual methylation states of the multiple CpG sites from a training fragment set; generating a test state vector for a test fragment, wherein the test state vector comprises: a test genome position in the reference genome and a methylation state of each CpG site in the test fragment, wherein each methylation state is determined to be one of methylated, unmethylated, and uncertain; based on the counts stored in the data structure, The invention relates to a method for preparing a methylation fragment comprising: calculating a test probability for the test state vector based on the multiple counts of the test state vector; sampling a subset of possible methylation state vectors having the same length as the test state vector from the test genomic position; for each of the multiple sampled possible methylation state vectors, calculating a probability corresponding to the sampled possible methylation state vector based at least in part on the multiple counts stored in the data structure; calculating a ratio of the multiple sampled possible methylation state vectors corresponding to a calculated probability less than or equal to the test probability; generating an estimated score for the test fragment based on the calculated ratio; and determining whether the test fragment is likely to have an abnormal methylation pattern based on the estimated score.

[0033] In some embodiments, the method further includes: filtering the test fragments by comparing the estimated score to a threshold score, selecting the threshold score so that multiple test fragments associated with an estimated score below the threshold score are more likely to include an abnormal methylation pattern. In some embodiments, the method further includes: in response to determining that the test fragment is likely to have an abnormal methylation pattern, calculating an exhaustive score for the test fragment of the test state carrier relative to the multiple training fragment sets, wherein the exhaustive score is based on the test probability and the multiple probabilities of the multiple possible methylation state carriers; and determining whether the test fragment has an abnormal methylation pattern based on the exhaustive score.

[0034] In some embodiments, the method further includes: applying a classifier to the test state carrier, using a first training segment set from one or more training subjects suffering from cancer and a second training segment set from one or more training subjects not suffering from cancer to train the classifier, wherein the classifier can be used to determine whether the test subject has cancer.

[0035] In another aspect, the present disclosure provides a non-transitory computer-readable storage medium storing a plurality of executable instructions for detecting an abnormal methylation pattern in a cell-free deoxyribonucleic acid (cfDNA) sample fragment. When executed by a hardware processor, the hardware processor is caused to perform the following steps, including: accessing a data structure, the data structure including: counts of strings of multiple CpG sites in a reference genome, and individual methylation states of the multiple CpG sites from a training fragment set; generating a sample state vector for a sample fragment, the sample state vector including: a sample genome position in the reference genome and A methylation state of each CpG site in the sample fragment, each methylation state is determined to be methylated or unmethylated; multiple possibilities of multiple methylation states having the same length as the sample state vector are listed from the sample genomic position; for each possibility, a probability is calculated by accessing multiple counts stored in the data structure; the possibility that matches the sample state vector is determined, and the calculated probability is used as a sample probability accordingly; based on the sample probability, a score is generated for the sample fragment of the sample state vector relative to the training fragment set; and based on the generated score, it is determined whether the sample fragment has an abnormal methylation pattern.

[0036] In some embodiments, in the non-transitory computer-readable storage medium of claim 44, the string of each CpG site includes: the methylation state of each of the CpG sites at multiple genomic positions within the reference genome, wherein each of the methylation states is determined to be methylated or unmethylated.

[0037] In some embodiments, the steps further include: constructing the data structure from the training segment set, and including: generating a training state vector for each training segment in the training segment set, the training state vector including: a known genomic position within the reference genome and the methylation state of each CpG site in the training segment, each methylation state being determined as methylated or unmethylated; determining a plurality of strings, wherein each string is part of the training state vector; determining a count for each string from the plurality of training state vectors; and storing a plurality of counts for each string in the data structure.

[0038] In some embodiments, the step of determining whether the sample fragment has an abnormal methylation pattern based on the generated score further includes: determining whether the score generated for the sample fragment is below a threshold score, wherein the threshold score indicates a confidence level that the sample fragment has an abnormal methylation pattern. In some embodiments, the threshold score is 0.1 or less.

[0039] In some embodiments, the set of training fragments comprises: training fragments from one or more healthy subjects, wherein the one or more healthy subjects lack a specific medical disease, and wherein the sample fragment is determined to be abnormally methylated relative to the set of training fragments from the one or more healthy subjects.

[0040] In some embodiments, generating the score for the sample fragment comprises: identifying, for a plurality of likelihoods of methylation states, a calculated probability that is less than the sample probability; and generating the score for the sample fragment by adding all of the identified probabilities to the sample probability.

[0041] In some embodiments, the step of calculating a probability for each of the possibilities by accessing a plurality of counts stored in the data structure comprises: for each of the plurality of conditional elements, wherein each conditional element is a conditional probability, the conditional probability taking into account a subset of CpG sites in the possibility, calculating a Markov chain probability of an order, the Markov chain probability of the order having a plurality of counts stored in the data structure by the following steps, the steps comprising: identifying a first count of the number of strings that match the conditional element; identifying a second count of the number of strings that match the previous methylation state of the conditional element at an integer length; and calculating the Markov chain probability by dividing the first count by the second count. In some embodiments, the order is selected from the group consisting of 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, and 15. In some embodiments, the step of calculating a Markov chain probability for an order, the Markov chain probability of the order having a plurality of counts stored in the data structure further comprises: implementing a smoothing algorithm.

[0042] In some embodiments, the sample state carrier is divided into multiple windows, the multiple windows including a first window and a second window, wherein the first window and the second window are two different parts of the sample fragment; wherein the step of determining the possibility of matching the sample state carrier and correspondingly taking the calculated probability as the sample probability includes: identifying a first possibility with a first sample probability matching the first window, and identifying the second possibility with a second sample probability matching the second window; and wherein the generated score is based on one of the first sample probability and the second sample probability.

[0043] In some embodiments, the step further comprises: filtering the plurality of sample fragments based on the scores generated for the respective sample fragments, thereby resulting in a subset of the plurality of sample fragments having a plurality of abnormal methylation patterns.

[0044] In some embodiments, the counting step further comprises identifying the sample fragments as hypermethylated when the sample fragments include at least a threshold number of CpG sites and the plurality of CpG sites have more than a threshold percentage methylated. In some embodiments, the threshold number of CpG sites is 5 or more CpG sites, and wherein the threshold percentage of the plurality of CpG sites methylated is 80% or greater.

[0045] In some embodiments, the steps further include: when the sample fragments include at least a threshold number of multiple CpG sites and the multiple CpG sites have more than a threshold percentage that are not methylated, the sample fragments are identified as being hypomethylated. In some embodiments, the threshold number of multiple CpG sites is 5 or more CpG sites, and the threshold percentage of multiple CpG sites that are not methylated is 80% or greater.

[0046] In some embodiments, the step further comprises: applying the sample state vector to a classifier, the classifier being trained using a set of cancer training segments from one or more subjects suffering from cancer and a set of non-cancer training segments from one or more subjects not suffering from cancer, wherein the classifier can be used to determine whether the sample segment is from a subject suffering from cancer. In some embodiments, the step of applying the sample state vector to the classifier generates at least one of a cancer probability and a non-cancer probability. In some embodiments, the step further comprises: generating a cancer state score based on at least one of the cancer probability and the non-cancer probability.

[0047] In another aspect, the present disclosure provides a non-transitory computer-readable storage medium storing a plurality of executable instructions for determining whether a test subject suffers from cancer, wherein the plurality of executable instructions, when executed by a hardware processor, causes the hardware processor to perform the following steps: accessing a model obtained by a training process, wherein the training process uses a cancer fragment set from one or more training subjects suffering from cancer and a non-cancer fragment set from one or more training subjects not suffering from cancer, wherein both the cancer fragment set and the non-cancer fragment set include a plurality of training fragments, wherein the training process includes: for each training fragment, determining whether the training fragment is hypomethylated or hypermethylated, wherein each hypomethylated and hypermethylated training fragment respectively includes at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites is unmethylated or methylated, and for each CpG site in a reference genome: quantifying a count of a plurality of hypomethylated training fragments overlapping with the CpG site and a count of a plurality of hypermethylated training fragments overlapping with the CpG site ; and generating a low methylation score and a high methylation score based on the counts of multiple low methylation training segments and multiple high methylation training segments; for each training segment, generating a total low methylation score based on the low methylation scores of the multiple CpG sites in the training segment, and generating a total high methylation score based on the high methylation scores of the multiple CpG sites in the training segment; for each training subject: ranking multiple training segments based on the total low methylation scores, and ranking multiple training segments based on the total high methylation scores; and generating a feature vector based on the ranking of the multiple training segments; obtaining multiple training feature vectors for the one or more training subjects not suffering from cancer, and obtaining multiple training feature vectors for the one or more training subjects suffering from cancer; using the multiple feature vectors for the one or more training subjects not suffering from cancer and the multiple feature vectors for the one or more training subjects suffering from cancer to train the model; and applying the model to a test feature vector corresponding to the test subject to determine whether the test subject suffers from cancer. In some embodiments, the threshold number is 5 or more. In some embodiments, the threshold percentage is 80% or greater.

[0048] In some embodiments, for each CpG site in a reference genome, the step of quantifying a count of multiple hypomethylated training fragments overlapping with the CpG site and a count of multiple hypermethylated training fragments overlapping with the CpG site further includes: quantifying a cancer count of multiple hypomethylated training fragments from the one or more training subjects suffering from cancer overlapping with the CpG site and a non-cancer count of multiple hypomethylated training fragments from the one or more training subjects not suffering from cancer overlapping with the CpG site; and quantifying a cancer count of multiple hypermethylated training fragments from the one or more training subjects suffering from cancer overlapping with the CpG site and a non-cancer count of multiple hypermethylated training fragments from the one or more training subjects not suffering from cancer overlapping with the CpG site. In some embodiments, for each CpG site in a reference genome, the step of generating a low methylation score and a high methylation score based on the counts of multiple low methylation training fragments and multiple high methylation training fragments further includes: for generating the low methylation score, calculating a low methylation ratio, the low methylation ratio is the ratio of the cancer counts of multiple low methylation training fragments to the low methylation sum of the cancer counts of multiple low methylation training fragments and the non-cancer counts of multiple low methylation training fragments; and for generating the high methylation score, calculating a high methylation ratio, the high methylation ratio is the ratio of the cancer counts of multiple high methylation training fragments to the high methylation sum of the cancer counts of multiple high methylation training fragments and the non-cancer counts of multiple high methylation training fragments.

[0049] In some embodiments, the hypomethylation ratio and the hypermethylation ratio are further calculated using a smoothing algorithm. In some embodiments, for each CpG site in a reference genome, the step of generating a hypomethylation score and a hypermethylation score based on the counts of multiple hypomethylated training fragments and multiple hypermethylated training fragments further includes: for generating the hypomethylation score, calculating a hypomethylation log ratio, the hypomethylation log ratio is the log ratio of the cancer counts of multiple hypomethylated training fragments to the non-cancer counts of multiple hypomethylated training fragments; and for generating the hypermethylation score, calculating a hypermethylation log ratio, the hypermethylation log ratio is the log ratio of the cancer counts of multiple hypermethylated training fragments to the non-cancer counts of multiple hypermethylated training fragments. In some embodiments, the hypomethylation ratio and the hypermethylation ratio are further calculated using a smoothing algorithm. In some embodiments, for each training fragment, the step of generating an aggregate low methylation score based on the low methylation scores of the multiple CpG sites in the training fragment, and generating an aggregate high methylation score based on the high methylation scores of the multiple CpG sites in the training fragment, further includes: identifying a maximum low methylation score of the multiple CpG sites in the training fragment as the aggregate low methylation score, and identifying a maximum high methylation score of the multiple CpG sites in the training fragment as the aggregate high methylation score.

[0050] In some embodiments, for each training subject, the step of generating a training feature vector based on the ranking of the multiple training fragments further includes: identifying a plurality of aggregated low methylation scores from the ranking, and identifying a plurality of aggregated high methylation scores from the ranking, and generating a training feature vector, wherein the training feature vector includes the plurality of aggregated low methylation scores and the plurality of high methylation scores.

[0051] In some embodiments, the step of training the model using the multiple training feature vectors for the one or more training subjects not suffering from cancer and the multiple training feature vectors for the one or more training subjects suffering from cancer is performed by training through a non-linear classifier.

[0052] In some embodiments, the step further comprises: for each training subject, normalizing the training feature vector by an average length of a plurality of training segments of the training subject.

[0053] In some embodiments, the step further includes: obtaining the test feature carrier corresponding to the test object, wherein the step of obtaining the test feature carrier includes: obtaining a plurality of sequence reads of a test fragment set from the test object; for each test fragment, determining whether the test fragment is hypomethylated or hypermethylated, wherein each of the plurality of hypomethylated and hypermethylated test fragments respectively includes at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites is unmethylated or methylated, and for each CpG site in a reference genome: a count of a plurality of hypomethylated test fragments overlapping with the CpG site and a count of a plurality of hypermethylated test fragments overlapping with the CpG site; A count of multiple hypermethylated test fragments is quantified; and a low methylation score and a high methylation score are generated based on the counts of multiple low methylation test fragments and multiple high methylation test fragments; for each test fragment, an aggregate low methylation score is generated based on the low methylation scores of the multiple CpG sites in the test fragment, and an aggregate high methylation score is generated based on the high methylation scores of the multiple CpG sites in the test fragment; for the test object, multiple test fragments are ranked based on the aggregate low methylation scores, and multiple test fragments are ranked based on the aggregate high methylation scores; and the test feature vector is generated based on the ranking of the multiple test fragments.

[0054] In some embodiments, the step of applying the model to the test feature carrier of the test subject to determine whether the test subject has cancer includes: generating a cancer probability for the test subject based on the model; and comparing the cancer probability with a threshold probability to determine whether the test subject has cancer.

[0055] In another aspect, the present disclosure provides a non-transitory computer-readable storage medium storing executable instructions for determining whether a test subject suffers from cancer, wherein when the executable instructions are executed by a hardware processor, the hardware processor performs the following steps: accessing a model obtained by a training process, wherein the training process uses a cancer training segment set from one or more training subjects suffering from cancer and a non-cancer training segment set from one or more training subjects not suffering from cancer, wherein both the cancer training segment set and the non-cancer training segment set include multiple training segments, wherein the training process includes: for each training segment, determining whether the training segment is low-frequency methylated or hypermethylated, wherein each of the hypomethylated and hypermethylated training fragments respectively includes at least a threshold number of multiple CpG sites, and at least a threshold percentage of the multiple CpG sites are unmethylated or methylated, for each training subject, generating a training feature vector based on the multiple hypomethylated training fragments and the multiple hypermethylated training fragments; and using the multiple training feature vectors from the one or more training subjects without cancer and the multiple feature vectors from the one or more training subjects with cancer to train the model; and applying the model to a test feature vector corresponding to the test subject to determine whether the test subject has cancer.

[0056] In some embodiments, for each training subject, the step of generating the training feature vector includes: for each CpG site in a reference genome: quantifying a count of multiple hypomethylated training fragments overlapping the CpG site and a count of multiple hypermethylated training fragments overlapping the CpG site; and generating a hypomethylation score and a hypermethylation score based on the counts of multiple hypomethylated training fragments and multiple hypermethylated training fragments; for each training fragment of the training subject, generating an aggregate hypomethylation score based on the hypomethylation scores of the multiple CpG sites in the training fragment, and generating an aggregate hypermethylation score based on the hypermethylation scores of the multiple CpG sites in the training fragment; and ranking the multiple training fragments of the training subject based on the aggregate hypomethylation scores, and ranking the multiple training fragments of the training subject based on the aggregate hypermethylation scores, wherein the training feature vector for the training subject is the ranking based on the aggregate hypomethylation scores and the ranking based on the aggregate hypermethylation scores.

[0057] In some embodiments, the step further includes: obtaining the test feature vector corresponding to the test object, wherein the step of obtaining the test feature vector includes: obtaining a plurality of sequence reads of a test fragment set from the test object; for each test fragment, determining whether the test fragment is hypomethylated or hypermethylated, wherein each of the plurality of hypomethylated and hypermethylated test fragments respectively includes at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites are unmethylated or methylated, and for each CpG site in a reference genome: calculating a count of a plurality of hypomethylated test fragments overlapping with the CpG site and a count of a plurality of hypermethylated test fragments overlapping with the CpG site; quantify a count of multiple hypermethylated test fragments; and generate a low methylation score and a high methylation score based on the counts of multiple low methylation test fragments and multiple high methylation test fragments; for each test fragment, generate a total low methylation score based on the low methylation scores of the multiple CpG sites in the test fragment, and generate a total high methylation score based on the high methylation scores of the multiple CpG sites in the test fragment; for the test subject, rank the multiple test fragments based on the total low methylation scores, and rank the multiple test fragments based on the total high methylation scores; and generate the test feature carrier based on the ranking of the multiple test fragments. In some embodiments, the step of applying the model to the test feature carrier of the test subject to determine whether the test subject suffers from cancer includes: generating a cancer probability for the test subject based on the model; and comparing the cancer probability with a threshold probability to determine whether the test subject suffers from cancer.

[0058] In one aspect, the present disclosure provides a non-transitory computer-readable storage medium storing a plurality of executable instructions for determining whether a test fragment from a test subject suspected of having cancer has an abnormal methylation pattern, which, when executed by a hardware processor, causes the hardware processor to perform the following steps, including: accessing a data structure comprising: counts of strings of multiple CpG sites in a reference genome, and individual methylation states of the multiple CpG sites from a training fragment set; generating a test state vector for a test fragment, wherein the test state vector comprises: a test genome position in the reference genome and a methylation state of each CpG site in the test fragment, wherein each methylation state is determined as methylated, unmethylated, or methylated; one of the methylated and uncertain ones; calculating a test probability for the test state vector based on the multiple counts stored in the data structure; sampling a subset of possible methylation state vectors having the same length as the test state vector from the test genomic position; for each of the multiple sampled possible methylation state vectors, calculating a probability corresponding to the sampled possible methylation state vector based at least in part on the multiple counts stored in the data structure; calculating a ratio of the multiple sampled possible methylation state vectors corresponding to a calculated probability less than or equal to the test probability; generating an estimated score for the test fragment based on the calculated ratio; and determining whether the test fragment is likely to have an abnormal methylation pattern based on the estimated score.

[0059] In some embodiments, the steps further include: filtering the test fragments by comparing the estimated score to a threshold score, the threshold score being selected so that a plurality of test fragments associated with an estimated score below the threshold score are more likely to include an abnormal methylation pattern. In some embodiments, the steps further include: in response to determining that the test fragment is likely to have an abnormal methylation pattern, calculating an exhaustive score of the test fragment for the test state vector relative to the plurality of training fragment sets, wherein the exhaustive score is based on the test probability and the plurality of probabilities of the plurality of possible methylation state vectors; and determining whether the test fragment has an abnormal methylation pattern based on the exhaustive score. In some embodiments, the steps further include: applying a classifier to the test state vector, training the classifier using a first training fragment set from one or more training subjects suffering from cancer and a second training fragment set from one or more training subjects not suffering from cancer, wherein the classifier can be used to determine whether the test subject suffers from cancer.

[0060] In another aspect, the present disclosure provides a non-transitory computer-readable storage medium storing a plurality of executable instructions, which, when executed by a hardware processor, cause the processor to execute a classifier to diagnose cancer, wherein the classifier is generated by the following process, the process comprising: (a) obtaining a plurality of sequence reads of a cancer fragment set from one or more subjects suffering from cancer, and a plurality of sequence reads of a non-cancer fragment set from one or more subjects not suffering from cancer, wherein both the cancer fragment set and the non-cancer fragment set include a plurality of sample fragments; (b) for each fragment, determining whether the fragment is hypomethylated or hypermethylated, wherein the hypomethylated and hypermethylated fragments include at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites are unmethylated or methylated; (c) for each CpG site in a reference genome: (i) separating a plurality of CpG sites overlapping with the CpG site from the CpG site; and (ii) determining whether the fragment is hypomethylated or hypermethylated. (i) quantifying a count of hypomethylated fragments and a count of multiple hypermethylated fragments overlapping the CpG site; and (ii) generating a hypomethylation score and a hypermethylation score based on the counts of the multiple hypomethylated fragments and the multiple hypermethylated fragments; (d) for each subject: (i) ranking the multiple fragments based on the total hypomethylation score, and ranking the multiple fragments based on the total hypermethylation score; and (ii) generating a feature vector based on the ranking of the multiple fragments; (e) training a diagnostic model based on the multiple feature vectors generated from the one or more subjects suffering from cancer and from the one or more subjects not suffering from cancer, the diagnostic model being configured to receive a test feature vector set from a test subject and output a possibility of suffering from cancer based on the test feature vector set from the test subject; and (f) storing a set of parameters representing the diagnostic model on the non-transitory computer-readable storage medium.

[0061] In some embodiments, the diagnostic model comprises a convolutional neural network having multiple layers of neural networks, the multiple layers of neural networks comprising: an input layer for receiving the multiple feature carriers from the one or more subjects suffering from cancer; and an output layer for indicating the possibility of a cancer based on the multiple feature carriers. In some embodiments, the diagnostic model further comprises: updating the neural network by repeatedly back-propagating one or more error terms, the one or more error terms being obtained by applying a training instance from multiple training instances to the diagnostic model, and calculating a loss function, wherein the multiple layers are updated based on the calculated loss function. In some embodiments, the diagnostic model comprises a kernel logistic regression classifier. In some embodiments, the step of determining whether a fragment is hypomethylated or hypermethylated comprises: (a) accessing a data structure comprising: counts of strings of multiple CpG sites in a reference genome, and the individual methylation states of the multiple CpG sites from a training fragment set; (b) generating a state vector for the fragment, the state vector comprising: a genomic position in the reference genome and a methylation state for each CpG site in the fragment, each methylation state being determined to be methylated or unmethylated; (c) listing from the genomic position the methylation states associated with the state (d) for each possibility, calculating a corresponding probability based on multiple counts of multiple strings stored in the data structure; (e) determining the possible methylation state that matches the state vector and the calculated probability corresponding to the determined possible methylation state; (f) generating a score for the segment of the state vector relative to the training segment set based on the determined calculated probability; and (g) determining whether the segment is one of hypomethylated and hypermethylated based on the generated score. In some embodiments, the diagnostic model is a test feature vector applied to a test subject, and the diagnostic model is configured to output a cancer probability for the test subject, and compare the output cancer probability with a threshold probability to determine whether the test subject suffers from cancer.

[0062] In another aspect, the present disclosure provides a non-transitory computer-readable storage medium storing a plurality of executable instructions. When the plurality of executable instructions are executed by a hardware processor, the processor is caused to execute a classifier to diagnose cancer, wherein the classifier is generated by the following process, wherein the process comprises: (a) obtaining a plurality of sequence reads of a cancer fragment set from one or more subjects suffering from cancer, and obtaining a plurality of sequence reads of a non-cancer fragment set from one or more subjects not suffering from cancer, wherein both the cancer fragment set and the non-cancer fragment set comprise a plurality of sample fragments; (b) for each fragment, determining whether the fragment has an abnormal methylation pattern, thereby obtaining a set of abnormally methylated fragments; (c) for each abnormally methylated fragment, determining whether the abnormally methylated fragment is hypomethylated or hypermethylated, wherein the hypomethylated and hypermethylated fragments respectively comprise at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites are unmethylated or methylated; (d) for each abnormally methylated fragment, determining whether the abnormally methylated fragment is hypomethylated or hypermethylated, wherein the hypomethylated and hypermethylated fragments respectively comprise at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites are unmethylated or methylated; (e) for each abnormally methylated fragment, determining whether the abnormally methylated fragment is hypomethylated or hypermethylated, wherein the hypomethylated and hypermethylated fragments respectively comprise at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites are unmethylated or methylated; (f) for each abnormally methylated fragment, determining whether the abnormally methylated fragment is hypomethylated or hypermethylated, wherein the hypomethylated and hypermethylated fragments respectively comprise at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites are unmethylated or methylated; (g) for each abnormally methylated fragment, determining whether the abnormal For each CpG site in the group: (i) quantify a count of multiple hypomethylated fragments overlapping the CpG site and a count of multiple hypermethylated fragments overlapping the CpG site; and (ii) generate a hypomethylation score and a hypermethylation score based on the counts of the multiple hypomethylated fragments and the multiple hypermethylated fragments; (e) for each subject: (i) rank the multiple fragments based on the total hypomethylation score, and rank the multiple fragments based on the total hypermethylation score; and (ii) generate a feature vector based on the ranking of the multiple fragments; (f) train a diagnostic model based on the multiple feature vectors generated from the multiple feature vectors from the one or more subjects suffering from cancer and from the one or more subjects not suffering from cancer, the diagnostic model being configured to receive a test feature vector set from a test subject and output a possibility of suffering from cancer based on the test feature vector set from the test subject; and (g) store a set of parameters representing the diagnostic model on the non-transitory computer-readable storage medium.

[0063] In some embodiments, the diagnostic model comprises a convolutional neural network having multiple layers of neural networks, the multiple layers of neural networks comprising: an input layer for receiving the multiple feature carriers from the one or more subjects suffering from cancer; and an output layer for indicating the possibility of a cancer based on the multiple feature carriers. In some embodiments, the diagnostic model further comprises: updating the neural network by repeatedly back-propagating one or more error terms, the one or more error terms being obtained by applying a training instance from multiple training instances to the diagnostic model, and calculating a loss function, wherein the multiple layers are updated based on the calculated loss function. In some embodiments, the diagnostic model comprises a kernel logistic regression classifier. In some embodiments, the step of determining whether a fragment is abnormally methylated comprises: (a) accessing a data structure, the data structure comprising: counts of strings of multiple CpG sites in a reference genome, and the methylation states of each of the multiple CpG sites from a training fragment set; (b) generating a state vector for the fragment, the state vector comprising: a genomic position in the reference genome and a methylation state of each CpG site in the fragment, each methylation state being determined to be methylated or unmethylated; (c) listing multiple possibilities of multiple methylation states from the genomic position having the same length as the state vector; (d) for each possibility, calculating a corresponding probability based on multiple counts of multiple strings stored in the data structure; (e) determining the possible methylation state that matches the state vector, and the calculated probability corresponding to the determined possible methylation state; (f) generating a score for the fragment of the state vector relative to the training fragment set based on the determined calculated probability; and (g) determining whether the fragment is abnormally methylated based on the generated score. In some embodiments, the diagnostic model is a test feature carrier applied to a test subject, and the diagnostic model is configured to output a cancer probability for the test subject and compare the output cancer probability with a threshold probability to determine whether the test subject has cancer. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1A A flowchart is provided for describing a process of sequencing a cell-free (cf) DNA fragment to obtain a methylation status vector according to an embodiment.

[0065] Figure 1B According to one embodiment, for Figure 1A An illustration of the process of sequencing a cell-free (cf) DNA fragment to obtain a methylation status vector.

[0066] Figure 1C and Figure 1D Three graphs of data are shown that demonstrate the consistency of sequencing from a control group.

[0067] Figure 2 A flowchart of a process for establishing a data structure for a control group is described according to one embodiment.

[0068] Figure 3 According to one embodiment, the verification Figure 2 A flowchart of an additional step of configuring the control group data structure.

[0069] Figure 4 A flow chart depicting a process for identifying a plurality of abnormally methylated segments from a subject is provided, according to one embodiment.

[0070] Figure 5 is an illustration of an example of p-value score calculation according to one embodiment.

[0071] Figure 6 A flowchart depicting a process of training a classifier based on methylation status of multiple fragments is provided according to one embodiment.

[0072] FIG. 7A to FIG. 7C is a graph showing the cancer log-odds ratio determined for various cancers across different stages of cancer.

[0073] Fig. 8A is a flow chart of an apparatus for sequencing nucleic acid samples according to one embodiment. Figure 8B According to one embodiment, an analysis system is provided, wherein the analysis system is used to analyze the methylation status of cfDNA.

[0074] Fig. 9 Provided are ROC curves for all cancers (left) and high signal cancer types (right) in the training and test sets obtained from an experiment described in VIII. A. The upper graph describes the full range of specificity; the lower graph focuses on 90-100% specificity as shown to more clearly describe the sensitivity at 95%, 98% and 99% specificity.

[0075] Fig.10 The consistency between training and test set performance is shown. The sensitivity at 98% specificity is reported for various tumor types in training (x-axis) and testing (y-axis) using WGBS analysis. High signal cancers and sample sizes are shown. The gray shading represents the 95% confidence interval of the fitted line.

[0076] Fig.11A and Fig. 11BIt is shown that when analyzed by the WGBS, Fig.11A ) and test set ( Fig. 11B ) at 98% specificity (y-axis) for each tumor type (x-axis). Error bars represent 95% confidence intervals.

[0077] The accompanying drawings depict various embodiments of the present invention for purposes of illustration only. Those skilled in the art will readily recognize from the following discussion that alternative embodiments of the structures and methods shown herein may be employed without departing from the principles described herein. Specific embodiments

[0078] I. Overview

[0079] According to the present invention, cfDNA fragments from a test subject are processed to convert unmethylated cytosine into uracil, sequenced, and the sequence reads are compared with a reference genome to identify the methylation status of one or more CpG sites within a plurality of the fragments. Compared to healthy subjects, identifying abnormally methylated cfDNA fragments can help gain a deeper understanding of a subject's cancer status. As is well known in the art, abnormal DNA methylation (compared to a healthy control group) can cause different effects, which may lead to cancer. Various challenges arise in the identification of abnormally methylated cfDNA fragments. First, determining that one or more cfDNA fragments are abnormally methylated focuses only on a group of subjects compared to a group of pairs of subjects that are assumed to be normally methylated. In addition, in a group of control subjects, the methylation status may change, which may be difficult to interpret when determining that the subject's cfDNA is abnormally methylated. On the other hand, methylation of cytosine at a CpG site can affect methylation at subsequent CpG sites. Encapsulating this dependency is a challenge in itself.

[0080] Methylation usually occurs in deoxyribonucleic acid (DNA) when a hydrogen atom on the pyrimidine ring of a cytosine base is converted into a methyl group to form 5-methylcytosine. In particular, methylation tends to occur at dinucleotides of cytosine and guanine, referred to herein as "CpG sites". In other cases, methylation may occur on a cytosine outside a CpG site or on another nucleotide other than cytosine; however, such cases are rare. In the present disclosure, for the sake of clarity, methylation is discussed with reference to CpG sites. Abnormal methylation of cfDNA fragments can be further identified as hypermethylation or hypomethylation, both of which can be indicated as cancer states.

[0081] Those skilled in the art will appreciate that the principles described herein are equally applicable to detecting methylation in a non-CpG situation, including non-cytosine methylation. In such embodiments, the wet lab assays used to detect methylation may be different from those described herein. In addition, the methylation status vector may include elements of the vector that are typically sites where methylation has or has not occurred (even if they are not specific CpG sites). By such substitutions, the remaining methods described herein are the same, so the concepts of the invention described herein may be applied to those other forms of methylation.

[0082] The term "cell-free nucleic acid", "cell-free DNA" or "cfDNA" refers to a plurality of nucleic acid fragments or DNA fragments that circulate in a fluid from a body (e.g., blood) and originate from one or more healthy cells and / or from one or more cancer cells. In addition, cfDNA may come from other sources, such as viruses, embryos, etc.

[0083] The term "circulating tumor DNA" or "ctDNA" refers to nucleic acid fragments originating from tumor cells or other types of cancer cells that may be released from a subject into a fluid (e.g., blood circulating in the body) as a result of biological processes, such as apoptosis or necrosis of dying cells, or actively released by living tumor cells.

[0084] The term "individual" refers to a human individual. The term "healthy individual" refers to an individual who is assumed not to have cancer or disease. The term "subject" refers to an individual known to have or to be at risk of having cancer or disease.

[0085] The term "sequence read" refers to the nucleotide sequence read from a sample taken from an individual. Sequence reads can be obtained by various methods known in the art.

[0086] The term "read segment" or "read" refers to any nucleotide sequence, including sequence reads obtained from an individual and / or multiple nucleotide sequences, which are obtained from an initial sequence of a sample obtained from an individual.

[0087] II. Sample processing

[0088] Figure 1A1 is a flow chart 100 describing a process of sequencing a cell-free (cf) DNA fragment to obtain a methylation status vector according to one embodiment. To analyze DNA methylation, an analysis system first obtains a sample 110 from a subject, the sample comprising a plurality of cfDNA fragments. Typically, the sample may be from a healthy subject, a subject known to have or suspected of having cancer, or a subject with no prior disease data. The test sample may be a sample selected from the group consisting of blood, plasma, serum, urine, feces, and saliva samples. Alternatively, the test sample may include a sample selected from the group consisting of whole blood, a portion of blood, a tissue section, pleural fluid, pericardial fluid, cerebrospinal fluid, and peritoneal fluid.

[0089] From the sample, the cfDNA fragments are treated to convert unmethylated cytosine to uracil. In one embodiment, the method uses a bisulfite to treat the cfDNA fragments, which converts unmethylated cytosine to uracil without converting methylated cytosine. For example, the bisulfite conversion is performed using a commercially available kit, such as EZ DNA Methylation TM -Gold, EZ DNA Methylation TM -Direct or EZ DNA Methylation TM -Lightning kit (obtained from Uzmo Research, Inc. (Irvine, CA)). In another embodiment, the conversion of unmethylated cytosine to uracil is accomplished using an enzymatic reaction. For example, the conversion can be performed using a commercially available kit to convert unmethylated cytosine to uracil, such as APOBEC-Seq (New England Biolabs, Inc., Ipswich, MA).

[0090] From the converted cfDNA fragments, a sequencing library 130 is prepared. Optionally, the sequencing library can be enriched 135 for cfDNA fragments or genomic regions, which can provide information about the cancer state when multiple hybridization probes are used. The hybridization probes are short oligonucleotides that are capable of hybridizing with targeted cfDNA fragments or cfDNA fragments derived from one or more targeted regions, and enriching those fragments or regions for subsequent sequencing and analysis. Hybridization probes can be used for a targeted, high-depth analysis of a specific set of CpG sites of interest. Once prepared, the sequencing library or a portion thereof can be sequenced to obtain multiple sequence reads. The sequence reads can be in a computer-readable digital format for processing and interpretation by computer software.

[0091] From the sequence readings, the analysis system determines 150 a position and methylation state of each of the one or more CpG sites based on comparison with a reference genome. The analysis system generates 160 a methylation state vector for each fragment, which determines a position in the reference genome (e.g., as determined by the position of the first CpG site in each fragment, or another similar metric), multiple CpG sites in the fragment, and the methylation state of each CpG site in the fragment is methylated (e.g., represented as M), unmethylated (e.g., represented as U), or uncertain (e.g., represented as I). The observed state is the methylated and unmethylated state. Conversely, an unobserved state is uncertain. The methylation state vector can be stored in a temporary or permanent computer memory for later use and processing. In addition, the analysis system can remove duplicate readings or duplicate methylation state vectors from a single subject. In another embodiment, the analysis system can determine that a certain fragment has one or more CpG sites with uncertain methylation states. Uncertain methylation status may result from sequencing errors and / or disagreement between the methylation status of complementary strands of a DNA fragment. The analysis system may decide to exclude such fragments or selectively include such fragments, but establish a model to account for such uncertain methylation status. Figure 4 A model to describe this kind of.

[0092] Figure 1B According to one embodiment, for Figure 1A An illustration of a process 100 for sequencing a cell-free (cf) DNA fragment to obtain a methylation state vector. As an example, the analysis system uses a cfDNA fragment 112. In this example, the cfDNA fragment 112 includes three CpG sites. As shown, the first CpG site and the third CpG site of the cfDNA fragment 112 are methylated 114. During the processing step 120, the cfDNA fragment 112 is converted to produce a converted cfDNA fragment 122. During the processing 120, the cytosine of the unmethylated second CpG site is converted to uracil. However, the first CpG site and the third CpG site are not converted.

[0093] After conversion, a sequence library 130 is prepared and sequenced 140 to generate a sequence read 142. The analysis system compares the sequence read 142 to a reference genome 144 150. The reference genome 144 provides background information about where in the human genome the cfDNA fragments are derived. In this simplified example, the analysis system compares the sequence reads 150 so that three CpG sites are associated with CpG sites 23, 24, and 25 (for ease of description, arbitrary reference identifiers are used). Therefore, the analysis system generates information about the methylation status of all CpG sites on the cfDNA fragments 112 and where in the human genome the CpG sites are marked. As shown, the CpG sites methylated on the sequence reads 142 are read as cytosine. In this example, the cytosine only appears in the first CpG site and the third CpG site in the sequence read 142, which allows one to infer that the first CpG site and the third CpG site in the original cfDNA fragment are methylated. The second CpG site is read as a thymine (U is converted to T during the sequencing process), and therefore, the person can infer that the second CpG site is not methylated in the original cfDNA fragment. Using this information of both methylation status and position, the analysis system generates 160 a methylation state vector 152 for the cfDNA fragment 112. In this example, the resulting methylation state vector 152 is <M 23 , U 24 、M 25 >, wherein M corresponds to a methylated CpG site, U corresponds to an unmethylated CpG site, and the subscript numbers correspond to the positions of the respective CpG sites in the reference genome.

[0094] Figure 1C and Figure 1D Three data graphs are shown, which verify the consistency of sequencing from a control group. The first figure 170 shows the conversion accuracy of unmethylated cytosine to uracil (step 120) on cfDNA fragments obtained from a test sample, which spans subjects at different stages of cancer - stage I, stage II, stage III, stage IV, and non-cancer subjects. As shown in the figure, there is consistency in converting unmethylated cytosine on cfDNA fragments to uracil. The overall conversion accuracy is 99.47% with an accuracy of ±0.024%. The second figure 180 shows the average coverage at different stages of cancer. The average coverage of all groups is about 34 times the whole genome coverage of DNA fragments, and only fragments that are confidently marked to the genome are counted. The third figure 190 shows the cfDNA concentration of each sample across different stages of cancer.

[0095] III. Control Data Structure

[0096] III.A Establishment

[0097] Figure 2 1 is a flowchart of a process 200 for establishing a data structure for a control group, according to an embodiment. To establish a healthy control group data structure, the analysis system receives a plurality of DNA fragments (e.g., cfDNA) from a plurality of subjects. For example, through process 100, a methylation state vector is identified for each fragment.

[0098] Using the methylation state vectors of each fragment, the analysis system subdivides the methylation state vector into multiple strings of more than 210 CpG sites. In one embodiment, the analysis system subdivides 210 the methylation state vector so that the resulting strings are all less than a given length. For example, a methylation state vector of length 11 can be subdivided into strings of length less than or equal to 3, resulting in 9 strings of length 3, 10 strings of length 2, and 11 strings of length 1. In another example, a methylation state vector of length 7 can be subdivided into strings of length less than or equal to 4, resulting in 4 strings of length 4, 5 strings of length 3, and 7 strings of length 1. If the length of a methylation state vector is less than or equal to the length of a specific string, the methylation state vector can be converted into a single string, which includes all CpG sites of the vector.

[0099] The analysis system calculates 220 the number of strings present in the control group by counting the likelihood of each possible CpG site and methylation state in the vector, the control group having a specific CpG site as the first CpG site in the string and the likelihood of having a methylation state. For example, at a given CpG site, considering a string of length 3, there are 2^3 or 8 possible string configurations. At the given CpG site, for each of the 8 possible string configurations, the analysis system calculates 220 the possible number of occurrences of each methylation state vector in the control group. Continuing with this example, it may involve calculating the following quantities: for each starting CpG site x in the reference genome, <M x 、M x+1 、M x+2 >、 <M x 、M x+1 , U x+2 >、…、 x , U x+1 , U x+2 ​>. The analysis system establishes 230 the data structure, which stores counts of each starting CpG site and string likelihood.

[0100] There are several benefits to placing an upper limit on the length of a string. First, depending on the maximum length of a string, the size of the data structure created by the analysis system may increase dramatically. For example, a maximum string length of 4 means that each CpG site has at least 2^4 numbers to count a string of length 4. Increasing the maximum string length to 5 means that each CpG site has an additional 2^4 or 16 numbers to count, which doubles the number of calculations (and the required computer memory) compared to the previous string length. In terms of calculation and storage, reducing the size of the string helps maintain the establishment and performance of the data structure (for example, for subsequent access as described below). Second, a statistical consideration for limiting the maximum string length is to avoid overfitting downstream models that use the string counts. If long strings of CpG sites do not have a significant biological impact on the results (for example, abnormal predictions can predict the presence of cancer), then calculating probabilities based on a large number of CpG sites may cause problems because it requires a large amount of data that may not be available, and therefore is sparse and cannot perform correctly for a model. A large amount of data may not be available, so data that is too sparse for the model to perform properly will not work. For example, calculating the probability of abnormality / cancer conditioned on the previous 100 CpG sites will require counting strings in a data structure of length 100, and ideally some of the strings exactly match the previous 100 methylation states. If only counts of sparse strings of length 100 are available, there will not be enough data to determine whether a given string of length 100 is abnormal in a test sample.

[0101] III.A Data Structure Verification

[0102] Once the data structure has been established, the analysis system may seek to validate 240 the data structure and / or any downstream models utilizing the data structure. One type of validation is to check for consistency within the data structure of the control group. For example, if there are any abnormal subjects, samples, and / or fragments in the control group, the analysis system may perform various calculations to determine whether to exclude any fragments from one of those categories. In a representative example, the healthy control group may include samples that are not diagnosed but are cancerous, such that the samples include abnormal methylated fragments. The first type of validation can ensure that potential cancerous samples are removed from the healthy control group so as not to affect the purity of the control group.

[0103] The second validation is to use counts from the data structure itself (i.e., from the healthy control group) to check the probability model used to calculate the p-value. Figure 5 Describe the process for calculating the p-value. Once the analysis system generates a p-value for the methylation state carrier in the validation group, the analysis system will use the p-value to construct a cumulative density function (CDF). Using the CDF, the analysis system can perform various calculations on the CDF to verify the data structure of the control group. A test uses an argument that the CDF is ideally equal to or less than an identity function so that CDF(x)≤x. Conversely, being higher than the identity function shows that there are certain flaws in the probability model for the data structure of the control group. For example, if the p-value score of 1 / 100 of the fragments is 1 / 1000, which means that CDF(1 / 1000)=1 / 100>1 / 1000, then the second type of verification will fail, indicating that there is a problem with the probability model.

[0104] The third type of validation uses a set of healthy validation samples separate from the healthy validation samples used to build the data structure, which tests whether the data structure is built correctly and whether the model is valid. Figure 3 An example process for implementing this type of validation is described. The third type of validation can quantify how well the healthy control group generalizes the healthy sample distribution. If the third type of validation fails, the health of the healthy control group is not well distributed.

[0105] The fourth type of validation test uses samples from an unhealthy validation group. The analysis system calculates the p-value and constructs the CDF for the unhealthy validation group. For an unhealthy validation group, the analysis system expects to observe CDF(x)>x for at least one sample, or in other words, the opposite of the expected values ​​for the healthy control group and the healthy validation group in the second type of validation and the third type of validation. If the fourth type of validation fails, it indicates that the model did not properly identify the abnormality it was intended to identify.

[0106] Figure 3 According to one embodiment, the verification Figure 2 A flowchart of an additional step 240 of verifying the data structure of the control group. In this step 240 of validating the data structure, the system uses a validation group of subjects, samples and / or fragments having a composition assumed to be similar to that of the control group. For example, if the analysis system selects healthy subjects without cancer for the control group, the analysis system also uses healthy subjects without cancer in the validation group.

[0107] The analysis system uses the validation set and generates a set of 100 methylation state vectors as shown in FIG1. ​​The analysis system calculates a p-value for each methylation state vector from the validation set. Figure 4 and Figure 5 The calculation process of the p-value is further described. For various possibilities of methylation state carriers, the analysis system calculates 320 a probability from the data structure of the control group. Once the probability is calculated for the possibility of the methylation state carrier, the analysis system will calculate 330 a p-value score for the methylation state carrier based on the calculated probability. The p-value score represents an expectation, and the expectation finds that a specific methylation state carrier and other possible methylation state carriers have a lower probability in the control group. Therefore, compared with other methylation state carriers in the control group, a low p-value score generally corresponds to a relatively unexpected methylation state carrier, and compared with other methylation state carriers found in the control group, a high p-value score generally corresponds to a relatively more expected methylation state carrier. Once the analysis system generates a p-value score for the methylation state carrier in the validation group, the analysis system uses the p-value score from the validation group to establish 340 a cumulative density function (CDF). The CFD can be used in the validation test described in other parts of this section above.

[0108] IV. Identification of fragments with an aberrant methylation pattern

[0109] IV.A. General Process

[0110] Figure 4 is a flow chart of a process 400 for identifying a plurality of abnormally methylated segments from a subject, according to one embodiment. An example of the process 400 is intuitively illustrated in Figure 5 East, the following is a further description Figure 4 In process 400, the analysis system generates 100 methylation state vectors from cfDNA fragments of a subject. The analysis system processes each methylation state vector as follows.

[0111] In some embodiments, the analysis system filters 405 fragments with uncertain states at one or more CpG sites. In the embodiment, the analysis system implements a prediction model to identify fragments that are unlikely to have abnormal methylation patterns for filtering. For a sample fragment, the prediction model calculates the probability of a methylation state vector occurring when the sample fragment is compared with the healthy control group data structure. The prediction model randomly samples a possible subset of methylation state vectors of CpG sites included in the sample fragment methylation state vector. The prediction model calculates a probability corresponding to each sampled possible methylation state vector. The methylation state vector of the fragment and the probability of the sampled possible methylation state vector can be calculated according to a Markov chain model, which will be described in the example p-value score calculation in Section IV.B. below. The prediction model calculates the proportion of the sampled possible methylation state vectors corresponding to a probability less than or equal to the sample probability. The prediction model generates an estimated p-value score for the fragment based on the calculated proportion. The prediction model can filter fragments corresponding to p-value scores above a threshold and retain fragments corresponding to p-value scores below the threshold.

[0112] In additional embodiments, the prediction model may calculate a confidence probability that the prediction model uses to determine when to continue or when to terminate sampling. The confidence probability is based on the estimated p-value score and the probability of the sampled possible methylation state carriers to describe the likelihood that the true p-value score of the fragment is below a threshold (the calculation of the true p-value score will be described in the example p-value score calculation in Section IV.B. below). The prediction model may sample additional one or more possible methylation state carriers while repeatedly calculating the estimated p-value score and confidence probability. Thereafter, when the confidence probability is above a confidence threshold, the prediction model may terminate sampling.

[0113] For a given methylation state vector, the analysis system lists 410 all the possibilities of methylation state vectors having the same starting CpG site and the same length (i.e., a set of CpG sites) in the methylation state vector. Since each observed methylation state may be methylated or unmethylated, which are the only two possible states for each CpG site, the count of different possibilities of methylation state vectors depends on a power of 2, whereby a methylation state vector of length n is equal to 2 of the methylation state vectors. n For one or more CpG sites including methylation state vectors 2 of uncertain state, the analysis system may list 410 the possibilities of methylation state vectors considering only the conditions of the observed CpG sites.

[0114] The analysis system calculates 420 the probability of various possibilities of the observed methylation state vectors for the identified starting CpG site / methylation state vector length by accessing the healthy control group data structure. In one embodiment, the probability of observing a given possibility is calculated using a Markov chain probability to model the joint probability calculation, which will be referred to in detail in Figure 5 This is described in more detail below. In other embodiments, computational methods other than Markov chain probabilities are used to determine the probabilities of various possibilities for the observed methylation state vectors.

[0115] The analysis system uses the calculated probabilities of the various possibilities to calculate 430 a p-value score for the methylation state vector. In one embodiment, this includes identifying the calculated probability corresponding to the possibility that matches the methylation state vector in question. In particular, this is the possibility of having the same set of CpG sites as the methylation state vector or a similar starting CpG site and length as the methylation state vector. The analysis system adds the calculated probabilities of any possibilities having a probability less than or equal to the identified probability to generate the p-value score.

[0116] The p-value represents the probability of the observed methylation state carrier of the fragment or other methylation state carriers that are even lower in the healthy control group. Therefore, a low p-value score usually corresponds to a rare methylation state carrier in a healthy subject, and relative to the healthy control group, it causes the fragment to be abnormally methylated. A high p-value score is usually associated with the methylation state carrier expected to exist in a healthy subject. For example, if the healthy control group is a non-cancer group, a low p-value indicates that the fragment is abnormally methylated relative to the non-cancer group, and therefore may indicate the presence of cancer in the test subject.

[0117] As described above, the analysis system calculates a p-value score for each of a plurality of methylation state vectors, each representing a cfDNA fragment in the test sample. To identify which fragments are abnormally methylated, the analysis system may filter 440 the set of methylation state vectors based on their p-value scores. In one embodiment, filtering is performed by comparing the p-value scores to a threshold and retaining only those fragments below the threshold. This threshold p-value score may be on the order of 0.1, 0.01, 0.001, 0.0001, or the like.

[0118] According to the example results from process 400, the analysis system generated a median (range) of 2,800 (1,500-12,000) fragments with abnormal methylation patterns for participants who did not have cancer in training, and a median (range) of 3,000 (1,200-220,000) fragments with abnormal methylation patterns for participants who had cancer in training. Such filtered sets of fragments with abnormal methylation patterns can be used for downstream analysis, such as the example of a filtered set for abnormal fragments described in Section IV.D. below.

[0119] IV. Example of BP score calculation

[0120] Figure 5 5 is an illustration of an example of p-value score calculation according to one embodiment. To calculate the p-value score given a test methylation state vector 505, the analysis system uses the test methylation state vector 505 and enumerates 410 the possibilities of the methylation state vectors. In this illustrative example, the test methylation state vector 505 is<M23、M24、M25、U26> Since the length of the test methylation state vector 505 is 4, the possibility of a methylation state vector including CpG sites 23-26 is 2^4. In a general example, the number of possibilities of methylation state vectors is 2^n, where n is the length of the test methylation state vector, or the length of the sliding window (described further below).

[0121] The analysis system calculates 420 a probability 515 for the likelihood of the enumerated methylation state carriers. Since methylation is conditionally dependent on the methylation state of nearby CpG sites, one method of calculating the probability of an observed likelihood of a given methylation state carrier is to use a Markov chain model. In general, a methylation state carrier (e.g. 1 , S 2 ,...,S n >), where S represents the methylation status whether it is methylated (denoted as M), unmethylated (denoted as U), or uncertain (denoted as I), which has a joint probability, and it can be extended using the probability chain rule as:

[0122] P( 1 ,S 2 ,…,S n >)=P(S n |S 1 ,…,S n-1 )*P(S n-1 |S 1 ,…,S n-2 ​​)*…*P(S 2 |S 1 )*P(S 1 ) (1)

[0123] Markov chain models can be used to more efficiently calculate the conditional probabilities of various possibilities. In one embodiment, the analysis system selects a Markov chain order k that corresponds to how many previous CpG sites in the vector (or window) are considered in the conditional probability calculation, thereby modeling the conditional probability as P(S n |S 1 ,…,S n-1 )~P(S n |S n-k-2 ,…,S n-1 ).

[0124] To calculate the Markov modeling probability of each methylation state carrier, the analysis system accesses the control group data structure, specifically the counts of different strings of multiple CpG sites and states. n |S n-k-2 ,…,S n-1 ), the analysis system obtains a ratio of the number of stored character strings, the ratio being derived from n-k-2 ,…,S n-1 、M n >The data structure that matches is divided by n-k-2 ,…,S n-1 、M n > and n-k-2 ,…,S n-1 , U n > is the sum of the number of strings stored in the data structure. Therefore, P(M n |S n-k-2 ,…,S n-1 ) is calculated as a ratio with the following formula:

[0125]

[0126] The calculation may additionally be performed by applying a prior distribution to achieve a smoothing of the counts. In one embodiment, the prior distribution is a uniform prior, as in Laplace smoothing. As an example of this, a constant is added to the numerator and another constant (e.g., twice the constant in the numerator) is added to the denominator of the above equation. In other embodiments, an algorithmic technique such as Knesser-Ney smoothing is used.

[0127] ​​​In the illustration, the formula shown above is applied to the test methylation state vector 505 covering positions 23-26. Once the calculated probability 515 is completed, the analysis system calculates 430 a p-value score 525 that is less than or equal to the probability of the likelihood of the methylation state vector matching the test methylation state vector 505.

[0128] In embodiments with uncertain states, the analysis system may calculate a p-value score that sums the CpG sites with uncertain states in a methylation state vector for a fragment. The analysis system identifies all possibilities that are consistent with all methylation states of the methylation state vector (except the uncertain state). The analysis system may assign a probability to the methylation state vector as the sum of the probabilities of the identified possibilities. For example, since the methylation states of CpG sites 1 and 3 are observed and consistent with the methylation states of the fragment at CpG sites 1 and 3, the analysis system assigns a probability to the methylation state vector. <M 1 ,I 2 , U 3 The probability of a monomethylated state carrier of > is calculated as <M 1 ,I 2 , U 3 >With <M 1 , U 2 , U 3 > The sum of the probabilities of the likelihood of a methylation state vector with >. This method of summing CpG sites with uncertain states uses the calculation of probabilities up to 2^i, where i represents the number of uncertain states in the methylation state vector. In additional embodiments, a dynamic programming algorithm can be implemented to calculate the probability of a methylation state vector with one or more uncertain states. Advantageously, the dynamic programming algorithm operates in linear computation time.

[0129] In one embodiment, the computational burden of calculating probabilities and / or p-value scores can be further reduced by caching at least some of the calculations. For example, the analysis system may cache the probability of the likelihood of a methylation state carrier (or its window) in a temporary or persistent memory calculation. If other fragments have the same CpG site, the cached likelihood probability can effectively calculate the p-score without recalculating the potential likelihood probabilities. Equivalently, the analysis system can calculate the p-value score for each possibility of a methylation state carrier associated with a set of CpG sites from the carrier (or its window). The analysis system can cache the p-value score for determining the p-value score of other fragments including the same CpG site. Typically, the p-value score of the likelihood of a methylation state carrier with the same CpG site can be used to determine the p-value score of a different possibility from the same set of CpG sites.

[0130] IV.C. Sliding Window

[0131] In one embodiment, the analysis system uses 435 a sliding window to determine the likelihood of a methylation state vector and calculate a p-value. Rather than listing the likelihoods and calculating a p-value for the entire methylation state vector, the analysis system lists the likelihoods and calculates a p-value for only a window of consecutive CpG sites, where the length of the window (the length of the CpG site) is shorter than at least some of the fragments (otherwise, the window would be useless). The length of the window can be static, user-determined, dynamic, or otherwise selected.

[0132] When calculating the p-value of a methylation state carrier that is larger than the window, the window identifies a set of continuous CpG sites from the first CpG site in the carrier in a window starting from the first CpG site in the carrier. The analysis system calculates a p-value score for the window including the first CpG site. Afterwards, the analysis system "slides" the window to the second CpG site in the carrier and calculates another p-value score for the second window. Therefore, for a window size of l and a methylation carrier length of m, each methylation state carrier will generate m-1+1 p-value scores. After completing the p-value calculations for each part of the carrier, the lowest p-value score from all sliding windows is used as the overall p-value score for the methylation state carrier. In another embodiment, the analysis system sums the p-value scores of the methylation state carriers to generate an overall p-value score.

[0133] Using a sliding window helps reduce the number of possibilities enumerated by the methylation state vectors and their corresponding probability calculations that would otherwise need to be performed. The probability calculations for the examples are shown in Figure 5 However, typically the number of possibilities for methylation state vectors increases exponentially by a factor of 2 with the size of the methylation state vector. To give a real example, a fragment may have up to 54 CpG sites. The analysis system may use a window of size 5 (for example), but instead of calculating the probabilities of 2^54 (~1.8×10^16) possibilities to generate a single p-score, it performs 50 p-value calculations for each of the 50 windows of the methylation state vector of the fragment. Each of the 50 calculations lists the possibilities of 2^5 (32) methylation state vectors, with a total result of 50×2^5 (1.6×10^3) ​​probability calculations. This results in a significant reduction in the amount of calculations to be performed, but it does not make any sense for accurately identifying abnormal fragments. This additional step can also be applied when the methylation state vectors of the validation group are used to validate the 240 control group.

[0134] IV.D. Use Case Example of Filtering Anomalous Segment Sets

[0135] The analysis system may utilize the set of abnormal fragments to perform any type and / or possibility of additional analysis. An additional analysis identifies 450 hypomethylated fragments or hypermethylated fragments from the filter set. Hypomethylated or hypermethylated fragments can be defined as fragments having a specific length of CpG sites (e.g., more than 3, 4, 5, 6, 7, 8, 9, 10, etc.), wherein the fragments of the specific length of CpG sites have a high percentage of methylated CpG sites (e.g., more than 80%, 85%, 90% or 95%, or any other percentage in the range of 50%-100%) or a high percentage of unmethylated CpG sites (e.g., more than 80%, 85%, 90% or 95%, or any other percentage in the range of 50%-100%), respectively. Figure 6 , as described below, is an example process for identifying such hypomethylated or hypermethylated portions of a genome based on the set of aberrantly methylated fragments.

[0136] An alternative analysis applies 460 a trained classification model to the abnormal set of fragments. The trained classification model can be trained to recognize any condition of interest that can be identified from the methylation state vector. In one embodiment, the trained classification model is based on the methylation state of cfDNA fragments obtained from a group of subjects suffering from cancer, and optionally based on the methylation state of cfDNA fragments obtained from a group of healthy subjects who do not suffer from cancer, a binary classifier is trained, and then based on the abnormal methylation state vector, it is used to classify the probability of a test subject suffering from cancer or not suffering from cancer. In a further embodiment, different classifiers can be trained using groups of subjects known to suffer from specific cancers (e.g., breast cancer, lung cancer, prostate cancer, etc.) to predict whether a test subject suffers from those specific cancers.

[0137] In one embodiment, a classifier is trained based on the information about the hyper / hypomethylated regions from process 450 and see Figure 6 Description is as follows.

[0138] Another additional analysis calculates the log-odds ratio of abnormal segments from a subject, which is generally a cancer or a specific type of cancer. The log-odds ratio can be calculated by the applied 460 classification model as the logarithm of the ratio of the determined probability of cancer to the probability of non-cancerous (i.e., 1 minus the probability of cancer).

[0139] FIG. 7A to FIG. 7C A graph showing various cancers from different subjects across different stages is drawn according to the process described above and see Figure 4The log-odds ratio of the identified abnormal fragments. This data was obtained by testing more than 1700 clinically evaluable subjects, including nearly 600 subjects without cancer and more than 800 subjects with cancer. Fig. 7A The first graph 700 in shows all cancer cases across three different grades - no cancer; stage I / II / II; and stage IV. The log odds ratio for stage IV cancer is significantly greater than the log odds ratios for stage I / II / II and no cancer. Fig. 7A The second graph 710 in FIG. 7 shows breast cancer cases across all stages of cancer and those without cancer, with similar log-odds ratio progression increasing at progressive stages of cancer. Figure 7B The third graph 720 in Figure 7 shows breast cancer subtypes. It is noteworthy that the subtypes HER2+ and TNBC are more widely distributed, while HR+ / HER2- is concentrated around ~1. Figure 7C The fourth graph 730 shows lung cancer cases across all stages of cancer and those without cancer, steadily progressing through the advanced stages of lung cancer. The fifth graph 740 shows colorectal cancer cases across all stages of cancer and those without cancer, again showing steady progression through the advanced stages of colorectal cancer. Figure 7C The sixth graph 750 in FIG. 1 shows prostate cancer cases across all stages of cancer and no cancer. This example is different from most of the previous examples in that only stage IV is significantly different compared to the other stages I / II / II and no cancer.

[0140] V. Hypermethylated / hypomethylated regions and a classifier

[0141] Figure 6 According to one embodiment, a flow chart is provided for describing a process 600 for training a classifier based on the methylation status of a plurality of fragments. An analysis system may be used to perform the process 600. The process accesses two training sets of samples - a non-cancer set and a cancer set - and obtains 400 a non-cancer methylation status vector set and a cancer methylation status vector set, wherein the cancer methylation status vector set includes abnormal fragments of the sample in each set. For example, the process may be performed based on Figure 4 The process 400 is to identify abnormal segments.

[0142] The analysis system determines 610 for each methylation state vector whether the methylation state vector is hypomethylated or hypermethylated. Here, a hypermethylated or hypomethylated label is assigned if at least some CpG sites have a particular state (methylated or unmethylated, respectively) and / or have a threshold percentage of sites of a particular state (again methylated or unmethylated, respectively). As defined above, a cfDNA fragment is identified as hypomethylated or hypermethylated if the fragment has at least five unmethylated or methylated CpG sites, and (logical AND) more than 80% of the fragment's CpG sites are unmethylated or methylated, respectively.

[0143] In an alternative embodiment, the analysis system considers a portion of the methylation state vector and determines whether the portion is hypomethylated or hypermethylated, and can distinguish whether the portion is hypomethylated or hypermethylated. This alternative embodiment addresses missing methylation state vectors whose vector size is large but includes at least one densely hypomethylated or hypermethylated region. This process of defining hypomethylated and hypermethylated can be applied to Figure 4 Step 450.

[0144] The analysis system generates 620 a hypomethylated score and a hypermethylated score for each CpG site in the genome. To generate a score at a given CpG site, the classifier performs four counts at the CpG site - (1) a count of (methylation status) carriers of the cancer set marked as hypomethylated and overlapping the CpG site; (2) a count of carriers of the cancer set marked as hypermethylated and overlapping the CpG site; (3) a count of carriers of the non-cancer set marked as hypomethylated and overlapping the CpG site; (4) a count of carriers of the non-cancer set marked as hypomethylated and overlapping the CpG site. In addition, the process can normalize these counts for each group to account for differences in group size between the non-cancer group and the cancer group.

[0145] In one embodiment, the hypomethylated fraction at a given CpG site is defined as the logarithm of the ratio of (1) to (3). Similarly, the hypermethylated fraction is calculated as the logarithm of the ratio of (2) to (4). In addition, an additional smoothing technique as described above can be used to calculate these ratios.

[0146] In another embodiment, the low methylation score is defined as the ratio of (1) to the sum of (1) and (3). The high methylation score is defined as the ratio of (2) to the sum of (2) and (4). Similar to the above embodiment, the smoothing technique can be implemented as a ratio.

[0147] The analysis system generates 630 a total hypomethylation score and a total hypermethylation score for each abnormal methylation state carrier. The total hypermethylation and hypomethylation scores are determined based on the hypermethylation and hypomethylation scores of the CpG sites in the methylation state carrier. In one embodiment, the total hypermethylation and hypomethylation scores are respectively designated as the maximum hypermethylation and hypomethylation scores of the sites in each state carrier. However, in alternative embodiments, the total score may be based on a mean, a median, or other calculations using the hyper / hypomethylation scores of the sites in each carrier. In one embodiment, the analysis system assigns the larger of the total hypomethylation score and the total hypermethylation score to the abnormal methylation state carrier.

[0148] Afterwards, the analysis system ranks 640 the methylation state carriers of all the subjects by their combined low methylation scores and their combined high methylation scores, thereby obtaining two rankings for each subject. The process selects the combined low methylation scores from the low methylation ranking, and selects the combined high methylation scores from the high methylation ranking. Using the selected scores, the classifier generates 650 single feature carriers for each subject. In one embodiment, the scores selected from any ranking are selected in a fixed order identical to the feature carriers generated for each subject in each training group. As an example, the classifier selects the first, second, fourth, eighth, sixteenth, thirty-second, and sixty-fourth combined high methylation scores from each ranking, as well as each combined low methylation score, and writes those scores into the feature carrier of the subject (a total of 14 features in the feature carrier). In an additional embodiment, in order to adjust the sample sequencing depth, the analysis system adjusts the ranking in linear proportion to the relative sample depth. For example, if the relative sample depth is x, the interpolated score is obtained by x*original rank (i.e., x=1.1, we obtain scores calculated with ranks 1.1, 2.2, ..., x*2i). The analysis system can then define feature vectors based on the adjusted ranks for further classification.

[0149] The analysis system trains 660 a binary classifier to identify feature carriers between cancer and non-cancer training groups. The analysis system can group training samples into one or more training sample sets for iterative batch training of the binary classifier. After inputting all training sample sets including its training feature carriers and adjusting classification parameters, the binary classifier is fully trained according to the test samples marked by its feature carriers within some error range. For example, in one embodiment, the classifier determines the possibility or probability score (e.g., from 0 to 100) that the sample feature carrier is from a subject suffering from cancer. In some embodiments, the probability score is compared with a threshold probability to determine whether the subject suffers from cancer. In other embodiments, a probability score greater than or equal to 60 indicates that the subject suffers from cancer. In other embodiments, a probability score greater than or equal to 65, greater than or equal to 70, greater than or equal to 75, greater than or equal to 80, greater than or equal to 85, greater than or equal to 90, greater than or equal to 95 indicates that the subject suffers from cancer. Generally, any of a variety of classification techniques can be used. These techniques are numerous and include potential use of kernel methods, machine learning algorithms (e.g., multi-layer neural networks), etc.

[0150] In one embodiment, the classifier is a nonlinear classifier. In a specific embodiment, the classifier is a nonlinear classifier that utilizes L2 regularized kernel logistic regression with a Gaussian radial basis function kernel (RBF kernel). Specifically, a regularized kernel logistic regression (KLR) is trained using isotropic radial basis functions (power exponent 2) as the kernel with a scaling parameter γ and an L2 regularization parameter λ. γ and λ are optimized against logarithmic loss using internal cross validation within the specified training data, and are optimized using gate search in the multiplication steps, starting from a maximum value and halving the parameter in each step. In other embodiments, the classifier may include other types of classifiers, such as a random forest classifier, a hybrid model, a convolutional neural network, or an autoencoder model.

[0151] VI. Example Sequencing and Analysis System

[0152] Fig. 8A FIG. 1 is a flow chart of an apparatus for sequencing a nucleic acid sample according to one embodiment. The illustrative flow chart includes devices such as a sequencer 820 and an analysis system 800. The sequencer 820 and the analysis system 800 can work in tandem to perform the process 100, Figure 2 200 Figure 3 240 Figure 4 400 Figure 6 One or more steps of 600, as well as other processes described herein.

[0153] In various embodiments, the sequencer 820 receives an enriched nucleic acid sample 810. Fig. 8A As shown, the sequencer 820 may include a graphical user interface 825 that allows the user to interact with specific tasks (e.g., start sequencing or terminate sequencing), and one or more loading stations 830 for loading a sequencing cartridge including the enriched fragment sample and / or for loading buffer required for sequencing analysis. Therefore, once the user of the sequencer 820 has provided the necessary reagents and sequencing cartridges to the loading station 830 of the sequencer 820, the user can start sequencing by interacting with the graphical user interface 825 of the sequencer 820. Once sequencing is started, the sequencer 820 performs sequencing and outputs sequence reads of the enriched fragments from the nucleic acid sample 810.

[0154] In some embodiments, the sequencer 820 is communicatively coupled to the analysis system 800. The analysis system 800 includes some computing devices that are used to process sequence reads for various applications, such as evaluating the methylation status at one or more CpG sites, variant calling, or quality control. The sequencer 820 can provide sequence reads in a BAM file format to the analysis system 800. The analysis system 800 can be communicatively coupled to the sequencer 820 via a wireless, wired, or a combination of wireless and wired communication technology. Generally, the analysis system 800 is configured with a processor and a non-transitory computer-readable storage medium storing computer instructions, which, when executed by a processor, causes the processor to process sequence reads or perform one or more steps of any method or step described herein.

[0155] In some embodiments, the sequence reads can be aligned to a reference genome using methods known in the art to determine alignment position information, for example, Figure 1A . The alignment position may generally describe a starting position and an ending position of a region in the reference genome, which corresponds to a starting nucleotide and an ending nucleotide of a given sequence read. Corresponding to methylation sequencing, the alignment position information may generally indicate a first CpG site and a last CpG site included in the sequence read based on the alignment with the reference genome. The alignment position information may further indicate the methylation status and position of all CpG sites in a given sequence read. A region in the reference genome may be associated with a gene or a fragment of a gene. Thus, the analysis system 800 may label a sequence read using one or more genes that are aligned with the sequence read sequence. In one embodiment, the length (or size) of the fragment is determined from the starting and ending positions.

[0156] In various embodiments, for example, when a paired end sequencing process is used, a sequence read includes a read pair represented by R_1 and R_2. For example, the first read R_1 can be sequenced from a first end of a double-stranded DNA (dsDNA) molecule, and the second read R_2 can be sequenced from a second end of a double-stranded DNA (dsDNA). Therefore, the nucleotide base pairs of the first read R_1 and the second read R_2 can be aligned consistently with the nucleotide bases of the reference genome (e.g., in opposite directions). The alignment position information obtained from the read pair R_1 and R_2 can include a starting position in the reference genome corresponding to an end of a first read (e.g., R_1), and an ending position in the reference genome corresponding to an end of a second read (e.g., R_2). In other words, the starting position and the ending position in the reference genome represent possible positions corresponding to the nucleic acid fragment in the reference genome. An output file having a SAM (sequence alignment map) formula or a BAM (binary) formula can be generated and output for further analysis.

[0157] See now Figure 8B , Figure 8B FIG. 8 is a block diagram of an analysis system 800 for processing DNA samples according to one embodiment. The analysis system implements one or more computing devices for analyzing DNA samples. The analysis system 800 includes a sequence processor 840, a sequence database 845, a model database 855, a model 850, a parameter database 865, and a scoring engine 860. In some embodiments, the analysis system 800 performs the process 100, Figure 2 200 Figure 3 240 Figure 4 400 Figure 6 One or more steps of 600, as well as other processes described herein.

[0158] The sequence processor 840 generates a methylation state vector for a fragment from a sample. At each CpG site on a fragment, the sequence processor 840 generates a methylation state vector for each fragment, thereby specifying the position of the fragment in the reference genome, the number of CpG sites in the fragment, and the methylation state of each CpG site in the fragment, by Figure 1A The process 100 of determining whether a fragment is methylated, unmethylated, or uncertain. The sequence processor 840 may store the methylation state vectors for the fragments in a sequence database 845. The data in the sequence database 845 may be organized so that the methylation state vectors from a sample are associated with each other.

[0159] In addition, multiple different models 850 can be stored in the model library 855, or can be retrieved for use with test samples. In one example, a model is a trained cancer classifier that is used to use a feature vector derived from abnormal fragments to determine a cancer prediction for a test sample. The training and use of the cancer classifier will be further discussed in Section V. "High / Low Methylation Regions and a Classifier". The analysis system 800 can train one or more models 850 and store various trained parameters in the parameter library 865. The analysis system 800 stores the models 850 and functions in the model library 855.

[0160] During inference, the scoring engine 860 uses one or more models 850 to return outputs. The scoring engine 860 accesses the models 850 in the model library 855 along with trained parameters from the parameter library 865. According to each model, the scoring engine receives an appropriate input for the model and computes an output based on the received input, parameters, and a function of each model related to the input and output. In some use cases, the scoring engine 860 further computes metrics related to a confidence in the output computed from the model. In other use cases, the scoring engine 860 computes other intermediate values ​​used in the model.

[0161] VII. Application

[0162] In some embodiments, the method, analysis system and / or classifier of the present invention can be used to detect the presence of cancer, monitor the progression or recurrence of cancer, monitor treatment response or effectiveness, determine the presence or monitor minimal residual disease (MRD), or any combination thereof. For example, as described herein, a classifier can be used to generate a possibility or probability score (e.g., from 0 to 100), and its sample feature carrier is from a subject suffering from cancer. In some embodiments, the probability score is compared with a threshold probability to determine whether the subject suffers from cancer. In other embodiments, the possibility or probability score can be evaluated at different time points (e.g., before or after treatment) to monitor the progression of the disease or monitor the therapeutic effect (e.g., therapeutic effect). In other embodiments, the possibility or probability score can be used to make or influence a clinical decision (e.g., diagnosis of cancer, treatment selection, evaluation of treatment effectiveness, etc.). For example, in one embodiment, if the possibility or probability score exceeds a threshold, the physician can prescribe an appropriate treatment method.

[0163] VII.A. Early Detection of Cancer

[0164] In some embodiments, the methods and / or classifiers of the invention are used to detect the presence or absence of cancer in a subject suspected of having cancer. For example, a classifier (as described herein) can be used to determine a sample feature carrier likelihood or probability score from a subject suffering from cancer.

[0165] In one embodiment, a probability score greater than or equal to 60 indicates that the subject suffers from cancer. In other embodiments, a probability score greater than or equal to 65, greater than or equal to 70, greater than or equal to 75, greater than or equal to 80, greater than or equal to 85, greater than or equal to 90, greater than or equal to 95 indicates that the subject suffers from cancer. In other embodiments, a probability score may indicate the severity of the disease. For example, a probability score of 80 may indicate a more severe form or advanced stage of cancer compared to a score less than 80 (e.g., a score of 70). Similarly, an increase in the probability score over time (e.g., at a time point after the second) may indicate disease progression, or a decrease in the probability score over time (e.g., at a time point after the second) may indicate successful treatment.

[0166] In another embodiment, as described herein, a cancer log-odd ratio of a test subject can be calculated by taking the logarithm of the ratio of the probability of cancer to the probability of non-cancer (i.e., minus the probability of cancer). According to this embodiment, a cancer log-odd ratio greater than 1 can indicate that the subject suffers from cancer. In other embodiments, a cancer log-odd ratio greater than 1.2, greater than 1.3, greater than 1.4, greater than 1.5, greater than 1.7, greater than 2, greater than 2.5, greater than 3, greater than 3.5, or greater than 4.5 indicates that the subject suffers from cancer. In other embodiments, a cancer log-odd ratio can indicate the severity of the disease. For example, a cancer log-odd ratio greater than 2 may indicate a more severe form or advanced stage of cancer compared to a score less than 2 (e.g., a score of 1). Similarly, an increase in the cancer log-odd ratio over time (e.g., at a time point after the second) can indicate disease progression, or a decrease in the cancer log-odd ratio over time (e.g., at a time point after the second) can indicate treatment success.

[0167] According to aspects of the invention, the methods and systems of the invention can be trained to detect or classify a variety of cancer indications. For example, the methods, systems and classifiers of the invention can be used to detect one or more, two or more, three or more, five or more, ten or more, fifteen or more, or twenty or more different types of cancer that exist.

[0168] Examples of cancers that can be detected using the methods, systems, and classifiers of the present invention include malignancies, lymphomas, blastomas, sarcomas, and leukemias or lymphoid malignancies. More specific examples of such cancers include, but are not limited to, squamous cell carcinomas (e.g., epithelial squamous cell carcinomas), skin cancers, melanomas, lung cancers, including small cell lung cancers, non-small cell lung cancers ("NSCLC"), lung adenocarcinomas and lung squamous cell carcinomas, peritoneal cancers, gastric cancers or gastric cancers including gastrointestinal cancers, pancreatic cancers (e.g., pancreatic ductal adenocarcinomas), cervical cancers, ovarian cancers (e.g., high-grade serous ovarian cancers), liver cancers (e.g., hepatocellular carcinoma (HCC)), liver cancers, liver cancers, bladder cancers (e.g., urothelial bladder cancers), testicular (germ cell tumor) cancers, breast cancers (e.g., HER2-positive, HER2-negative, and triple-negative breast cancers), brain cancers (e.g., astrocytomas, gliomas (e.g., glioblastomas), gliomas (e.g., gliomas), ovarian cancers (e.g., ovarian cancer ... Examples of cancers include, but are not limited to, retinoblastoma, theca cell tumor, androgenic tumor, hematologic malignancies, including, but not limited to, non-Hodgkin lymphoma (NHL), multiple myeloma and acute hematologic malignancies, endometriosis, fibrosarcoma, choriocarcinoma, laryngeal cancer, Kaposi's sarcoma, schwannoma, neuroblastoma, rhabdomyosarcoma, osteogenic sarcoma, leiomyosarcoma, and urinary tract cancer.

[0169] In some embodiments, the cancer is anorectal cancer, bladder cancer, breast cancer, cervical cancer, colorectal cancer, esophageal cancer, gastric cancer, head and neck cancer, hepatobiliary cancer, leukemia, lung cancer, lymphoma, melanoma, multiple myeloma, ovarian cancer, pancreatic cancer, prostate cancer, kidney cancer, thyroid cancer, cervical cancer, or any combination thereof.

[0170] In some embodiments, one or more cancers may be a "high signal" cancer (defined as a cancer with a 5-year cancer-specific mortality rate greater than 50%), such as anorectal, colorectal, esophageal, head and neck, hepatobiliary, lung, ovarian, and pancreatic cancers, as well as lymphoma and multiple myeloma. High signal cancers tend to be more aggressive and typically have a higher than average concentration of cell-free nucleic acid in a test sample obtained from a patient.

[0171] VII.B. Cancer and Treatment Monitoring

[0172] In some embodiments, the likelihood or probability score can be evaluated at different time points (e.g., before or after treatment) to monitor disease progression or monitor treatment efficacy (e.g., treatment efficacy). For example, the present invention includes the following method: involving obtaining a first sample (e.g., a first plasma cfDNA sample) from a cancer patient at a first time point, determining a first likelihood or probability score (as described herein) therefrom, obtaining a second test sample (e.g., a second plasma cfDNA sample) from the cancer patient at a second time point, and determining a second likelihood or probability score (as described herein) therefrom.

[0173] In a specific embodiment, the first time point is before a cancer treatment (e.g., before a resection or medical treatment intervention), and the second time point is after a cancer treatment (e.g., after a resection or medical treatment intervention), and the method for monitoring the effect of the treatment. For example, if the second possibility or probability score decreases compared to the first possibility or probability score, the treatment is considered to have been successfully treated. However, if the second possibility or probability score increases compared to the first possibility or probability score, the treatment is considered to have been unsuccessful. In other embodiments, the first time point and the second time point are both before the cancer treatment (e.g., before a resection or medical treatment intervention). In other other embodiments, the first time point and the second time point are both after the cancer treatment (e.g., before a resection or treatment intervention), and the method for monitoring the effect of the treatment or the loss of the effect of the treatment. In other other embodiments, cfDNA samples can be obtained from cancer patients at a first and second time point and analyzed. For example, to monitor the progression of a cancer, to determine whether a cancer is in remission (eg, after treatment), to monitor or detect residual disease or recurrence of disease, or to monitor the efficacy of a treatment (eg, therapy).

[0174] Those skilled in the art will readily appreciate that test samples can be obtained from a cancer patient at any desired set of time points and analyzed according to the methods of the present invention to monitor the cancer status of the patient. In some embodiments, the time interval between the first time point and the second time point is from about 15 minutes to about 30 years, such as about 30 minutes, such as about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23 or about 24 hours, such as about 1, 2, 3, 4, 5, 10, 15, 20, 25 or about 30 days, or such as about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 or 12 months, or such as about 1, 1.5, 2, 2.5, 3, 3. 5, 4, 4.5, 5, 5.5, 6, 6.5, 7, 7.5, 8, 8.5, 9, 9.5, 10, 10.5, 11, 11.5, 12, 12.5, 13, 13.5, 14, 14.5, 15, 15.5, 16, 16.5, 17, 17.5, 18, 18.5, 19, 19.5, 20, 20.5, 21, 21.5, 22, 22.5, 23, 23.5, 24, 24.5, 25, 25.5, 26, 26.5, 27, 27.5, 28, 28.5, 29, 29.5 or about 30 years. In other embodiments, test samples can be obtained from a patient at least once every three months, at least once every six months, at least once a year, at least once every two years, at least once every three years, at least once every three years, at least once every four years, or at least once every five years.

[0175] VII.C. Treatment

[0176] In yet another embodiment, the likelihood or probability score can be used to make or influence a clinical decision (e.g., diagnosis of cancer, treatment selection, evaluation of treatment effectiveness, etc.). For example, in one embodiment, if the likelihood or probability score exceeds a threshold, the physician can prescribe an appropriate treatment method (e.g., a resection, radiation therapy, chemotherapy, and / or immunotherapy).

[0177] A classifier (as described herein) can be used to determine the likelihood or probability score that a sample feature carrier is from a subject suffering from cancer. In one embodiment, when the likelihood or threshold exceeds a threshold, a prescription for an appropriate treatment (e.g., resection surgery or treatment) can be prescribed. For example, in one embodiment, if the likelihood or probability score is greater than or equal to 60, one or more appropriate treatment prescriptions can be prescribed. In another embodiment, if the likelihood or probability score is greater than or equal to 65, greater than or equal to 70, greater than or equal to 75, greater than or equal to 80, greater than or equal to 85, greater than or equal to 90, or greater than or equal to 95, one or more appropriate treatment methods can be prescribed. In other embodiments, a cancer log-odd ratio can indicate the effectiveness of a cancer treatment. For example, an increase in the cancer log-odd ratio over time (e.g., at the second second after treatment) can indicate that the treatment is ineffective. Similarly, a decrease in the cancer log-odd ratio over time (e.g., at the second second after treatment) can indicate successful treatment. In another embodiment, if the cancer log-odds ratio is greater than 1, greater than 1.5, greater than 2, greater than 2.5, greater than 3, greater than 3.5, or greater than 4, one or more appropriate treatments are prescribed.

[0178] In some embodiments, the treatment is one or more selected from the group consisting of the following cancer therapeutic agents: a chemotherapeutic agent, a targeted cancer therapeutic agent, a differentiation therapeutic agent, a hormone therapeutic agent, and an immunotherapeutic agent. For example, the treatment can be one or more selected from alkylating agents, antimetabolites, anthracyclines, antitumor antibiotics, cytoskeletal disruptors (taxanes), topoisomerase inhibitors, mitotic inhibitors, corticosteroids, kinase inhibitors, nucleotide analogs, platinum-based agents, and any combination thereof. In some embodiments, the treatment is one or more targeted cancer therapeutic agents, which are selected from the group consisting of signal transduction inhibitors (e.g., tyrosine kinase and growth factor receptor inhibitors), histone deacetylase (HDAC) inhibitors, retinoic acid receptor agonists, proteosome inhibitors, angiogenesis inhibitors, and monoclonal antibody conjugates. In some embodiments, the treatment is one or more differentiation therapeutic agents, including retinoids, such as tretinoin, alum retinoic acid, and bexarotene. In some embodiments, the treatment is one or more hormonal therapeutic agents selected from the group consisting of antiestrogens, aromatase inhibitors, progestins, estrogens, antiandrogens, and GnRH agonists or analogs. In one embodiment, the treatment is one or more immunotherapeutics selected from the group consisting of monoclonal antibody therapies such as rituximab (RITUXAN) and alemtuzumab (CAMPATH), nonspecific immunotherapies and adjuvants such as BCG, interleukin 2 (IL-2) and interferon-α, immunomodulatory drugs such as thalidomide and lenalidomide (REVLIMID). It is within the capabilities of a skilled physician or oncologist to select an appropriate cancer therapeutic agent based on characteristics such as the type of tumor, stage of cancer, previous exposure to cancer treatments or therapeutic agents, and other characteristics of the cancer.

[0179] VIII. Examples

[0180] VIII.A. Methods for detecting abnormal methylated fragments for cancer diagnosis

[0181] Study Design and Samples: CCGA (NCT02889978) is a prospective, multicenter, case-control, observational study with longitudinal follow-up. Deidentified biospecimens were collected from approximately 15,000 participants at 142 sites. Samples were divided into training (1,785) and testing (1,015) groups; samples were selected to ensure a prespecified distribution across cancer types and non-cancer in each group, and cancer and non-cancer samples were matched for age and sex.

[0182] Whole-genome bisulfite sequencing: cfDNA was isolated from plasma and analyzed using whole-genome bisulfite sequencing (WGBS; 30x depth). cfDNA was extracted from two plasma tubes from each patient (maximum combined volume of 10 ml) using a modified QIAamp circulating nucleic acid kit (Kagan; Germantown, MD). Up to 75 ng of plasma cfDNA was bisulfite converted using the EZ-96 DNA methylation kit (Zymo Research, D5003). The converted cfDNA was used to prepare dual-indexed sequencing libraries using the Accel-NGS Methyl-Seq DNA Library Preparation Kit (Accel Biosciences; Ann Arbor, Michigan), and the constructed libraries were quantified using the KAPA Library Quantification Kit for Illumina Platforms (Kapa Biosystems; Wilmington, MA). The four libraries and 10% PhiX v3 library (Illumina, FC-110-3001) were pooled and clustered on an Illumina NovaSeq 6000S2 flow cell followed by 150 bp paired-end sequencing (30x).

[0183] Analysis of cfDNA and classification of cancer and non-cancer: For each sample, the WGBS fragment set was reduced to a small subset of fragments with an abnormal methylation pattern. In addition, hypermethylated or hypomethylated cfDNA fragments were selected. In this article, the cfDNA fragments selected for having an abnormal methylation pattern and being hypermethylated or hypermethylated are called "abnormal fragments of extreme methylation status" or "UFXM". In individuals without cancer, fragments that occur at high frequencies or fragments with unstable methylation are unlikely to produce highly discriminatory features for distinguishing cancer status. Therefore, we used an independent reference set (i.e., a reference genome) of 108 non-smoking participants from the CCGA study who did not suffer from cancer (age: 58±14 years, 79 [73%] women) to establish a statistical model and a data structure for typical fragments. These samples were used to train a Markov chain model (3rd order) to estimate the probability of a CpG methylation state of a given sequence in a fragment, as further described in Section IV.B. above. This model has been shown to calibrate within the normal fragment range (p-value>0.001) and was used to reject fragments with p-values ​​>= 0.001 from the markov model as being highly anomalous.

[0184] As described above, a further data reduction step selected only fragments covering at least 5 CpGs and with an average methylation of >0.9 (hypermethylated) or <0.1 (hypomethylated). The result of this process was a median (range) of 2,800 (1,500-12,000) UFXM fragments for participants who were not trained on cancer, and a median (range) of 3,000 (1,200-220,000) UFXM fragments for participants who were trained on cancer. Since this data only used the reference set data for the reduction procedure, this classification only needed to be applied once to each sample.

[0185] At selected loci within the genome, an approximate log-ratio score of cancer status information is constructed for hypermethylated and hypomethylated UFXM, respectively. First, a binary feature is generated for each sample at the site: 0 if there is no UFXM fragment and the site within the sample overlaps; 1 if there is a UFXM fragment overlapping the site. Then, the number of positive values ​​(1s) in the samples is counted from participants with cancer (C_c) and those without cancer (C_nc). The log-ratio score is then constructed as: log(C_c+1)-log(C_nc+1), adding a normal term to the count and discarding the normalization term related to the total number of samples in each group because it is a constant (log[N_nc+2]-log[N_c+2]). Scores were constructed at the locations of all CpG sites in the genome, resulting in approximately 25M loci with assigned scores: one score for UFXM hypermethylated fragments and one score for UFXM hypomethylated fragments.

[0186] Given a locus-specific log odds score, the UFXM segments in a sample are scored by taking the maximum of all log odds scores for the loci within the segment and matching the methylation category of hypermethylated or hypomethylated. This results in a score for each UFXM segment in a sample.

[0187] By taking the score of a subset of extremely methylated fragments in each sample, segmented into hypermethylated and hypomethylated fragments, we reduce the score of the fragment level in a sample to a small set of features for each sample. In this way, information about the most informative fragments in each sample can be obtained using a small set of useful functional features. In a low cfDNA tumor fraction sample, only a few fragments are expected to provide abnormal information.

[0188] In each category of fragments, the highest ranking scores 1, 2, 4, ... 64 (2 i, i in 0:6), and obtained 14 features (7 and 7). To adjust the sample sequencing depth, the ranking process is regarded as a function that assigns ranks to scores, and then we interpolate between the observed scores to obtain the score corresponding to the adjusted ranking. The ranking is adjusted linearly with the relative sample depth: if the relative sample depth is x, the interpolated score is x*original rank (i.e. x=1.1, we use the score between ranks 1.1, 2.2, ..., x*2 i Each sample was then assigned a set of 14 adjusted extreme scores for further classification.

[0189] Given the feature vector, a kernel logistic regression classifier is used to capture the underlying nonlinearity to predict cancer / non-cancer status based on the feature. Specifically, a regularized kernel logistic regression (KLR) is trained (by dividing by m) using an isotropic radial basis function (power 2) as the kernel with a scaling parameter γ and an L2 regularization parameter λ. 2 , where m is the number of samples, so λ scales naturally with the amount of training data). Internal cross-validation within the specified training data is used to optimize γ and λ with log-retaining loss and to achieve a 1,1e -2 (γ), 1e 3 -1e 1 The optimization was performed using a grid search over the range (λ), starting from the maximum value and halving the parameter in each step. The best median parameter over the internal cross-validation folds was 0.125 for γ and 125 for λ.

[0190] Validation of trained cancer classifiers: To evaluate the performance of this extreme score classifier procedure on the CCGA substudy dataset, a cross validation was performed on the training set, dividing the samples into 10 folds. Each fold was retained and an ERS classifier was trained on the remaining 9 / 10 of the data (internal cross validation was used within those folds to optimize γ and λ). The log-ratio scores used in the characterization only accessed the data of the trained folds. The scores output by each retained fold were summed and used to construct a receiver-operator characteristic (ROC) curve of performance. To evaluate the CCGA test set, the entire training dataset was used to construct a score and a single KLR classifier, which was then applied to the test dataset.

[0191] Sensitivity and specificity are estimated by classifiers; each classifier corrects or suppresses assay-specific interfering biological signals (e.g., CH, hematological conditions, age-related changes). After correction for interfering signals, cases without cancer are used to estimate specificity. The relationship between sensitivity and specificity is given by Fig. 9The receiver-operator characteristic (ROC) curve provided in is depicted. Fig. 9 The analysis was confirmed to have the potential for high specificity. The area under the curve (AUC) values ​​in the training data set and the test data set were similar. The AUC values ​​of a specific cancer group (0.88 and 0.87) were significantly higher than those of all cancers (0.73 and 0.71). The cancer group that showed high specificity in the analysis, which was called "high signal cancer" (defined as cancer with a 5-year specific cancer mortality rate of more than 50%), included several solid cancers (anorectal cancer, colorectal cancer, esophageal cancer, head cancer and neck, hepatobiliary, lung cancer, ovarian cancer and pancreatic cancer), as well as lymphoma and multiple myeloma.

[0192] Sensitivity was further investigated at a cutoff point with 98% specificity for: (1) an estimated 1.3% annual incidence of occult cancer in persons aged ≥50 years (SEER) and (2) continued follow-up of cancer-free participants. Sensitivity estimates were generally consistent between the training and test sets across cancer types ( Fig.10 ). The results are FIG. 11A to FIG. 11B Further described in.

[0193]

[0194]

[0195] In the training set of all cancer types, the overall sensitivity at 98% specificity was 39.5% (36-43%); it was consistent in the test set (34.2% [30-39%]). As expected, sensitivity increased with cancer stage. In the training set with 98% specificity, the sensitivity at 98% specificity in high signal cancers was 70.2% (65-75%), which was consistent with the test set (66.9% [59-74%]).

[0196] The results show that cfDNA sequencing and analysis of its methylation status can detect cancer with high specificity. This supports the feasibility of early detection, which may detect a larger proportion of cancers across stages, including some high-mortality cancers.

[0197] VIII. Other considerations

[0198] It should be understood that the drawings and descriptions of the present disclosure have been simplified to illustrate the elements relevant to a clear understanding of the present disclosure, while eliminating many other elements found in typical systems for the purpose of clarity. It will be appreciated by those of ordinary skill in the art that other elements and / or steps are desirable and / or required in implementing the present disclosure. However, since the elements and steps are well known in the art, and because they do not assist in providing a better understanding of the present disclosure, discussion of such elements and steps is not provided herein. The disclosure herein is intended for all such variations and modifications to the elements and methods known to those skilled in the art.

[0199] Some portions of the above description are based on algorithms and symbolic representations of operations on information to describe embodiments. Such algorithmic descriptions and representations are typically used by those skilled in the art of data processing to effectively convey the essence of their work to others skilled in the art. Although such operations are described functionally, computationally, or logically, it should be understood that they are implemented by computer programs or equivalent circuits, microcodes, etc. The described operations and their associated modules may be implemented in software, firmware, hardware, or any combination thereof.

[0200] As used herein, any reference to "one embodiment" or "an embodiment" represents that a particular element, feature, structure, or characteristic associated with the embodiment is included in at least one embodiment. The phrase "in an embodiment" appearing in various places in the specification is not necessarily the same embodiment, thereby providing a framework for the various possibilities of the described embodiments to work together.

[0201] As used herein, the terms "comprises," "including," "comprising," "including," "having," "having," or any other variation thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, article, or apparatus that comprises a list of elements is not necessarily limited to only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. In addition, unless expressly stated to the contrary, "or" refers to an inclusive or and not an exclusive or. For example, a condition A or B is satisfied by any of the following conditions: A is true (or exists) and B is false (or does not exist), A is false (or does not exist) and B is true (or exists), and both A and B are true (or exists).

[0202] In addition, "a" or "an" is used to describe elements and components of the embodiments herein. It is only for convenience and to give a general sense of the invention. The description should be understood to include one or at least one, and the singular also includes the plural, unless it is obvious that it is otherwise stated.

[0203] Although specific embodiments and applications have been shown and described, it should be understood that the disclosed embodiments are not limited to the precise configurations and components disclosed herein. Various modifications, changes and variations may be made to the arrangement, operation and details of the methods and apparatus disclosed herein without departing from the spirit and scope defined by the appended claims, which will be apparent to those skilled in the art.

Claims

1. A method for determining whether a test subject has cancer, characterized in that: The method comprises the steps of: accessing a model obtained by a training process using a set of cancer segments from one or more training subjects suffering from cancer and a set of non-cancer segments from one or more training subjects not suffering from cancer, wherein both the set of cancer segments and the set of non-cancer segments include a plurality of training segments, wherein the training process comprises: For each training segment, determining whether the training segment is hypomethylated or hypermethylated, wherein each of the hypomethylated and hypermethylated training segments respectively includes at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites is unmethylated or methylated, For each CpG site in a reference genome: quantifying a count of a plurality of low-methylated training segments overlapping the CpG site and a count of a plurality of high-methylated training segments overlapping the CpG site; and generating a low methylation score and a high methylation score based on the counts of the plurality of low methylation training segments and the plurality of high methylation training segments; For each training segment, generating an aggregate low methylation score based on the low methylation scores of the multiple CpG sites in the training segment, and generating an aggregate high methylation score based on the high methylation scores of the multiple CpG sites in the training segment; For each training subject: rank the plurality of training segments based on the aggregated low methylation scores, and rank the plurality of training segments based on the aggregated high methylation scores; and generating a feature vector based on the ranking of the plurality of training segments; Obtaining a plurality of training feature vectors for one or more training subjects not suffering from cancer, and obtaining a plurality of training feature vectors for the one or more training subjects suffering from cancer; and training the model using the plurality of feature vectors for the one or more training subjects not suffering from cancer and the plurality of feature vectors for the one or more training subjects suffering from cancer; and The model is applied to a test feature carrier corresponding to the test subject to determine whether the test subject suffers from cancer.

2. The method according to claim 1, characterized in that: The threshold number is 5 or more.

3. The method according to claim 1, characterized in that: The threshold percentage is 80% or greater.

4. A method for determining whether a test subject has cancer, characterized in that: The method comprises the steps of: accessing a model obtained by a training process using a set of cancer training segments from one or more training subjects suffering from cancer and a set of non-cancer training segments from one or more training subjects not suffering from cancer, wherein both the set of cancer training segments and the set of non-cancer training segments include a plurality of training segments, wherein the training process comprises: For each training segment, determining whether the training segment is hypomethylated or hypermethylated, wherein each of the hypomethylated and hypermethylated training segments respectively includes at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites is unmethylated or methylated, For each training subject, generating a training feature vector based on the plurality of hypomethylated training segments and the plurality of hypermethylated training segments; and training the model using the plurality of training feature vectors from the one or more training subjects not suffering from cancer and the plurality of feature vectors from the one or more training subjects suffering from cancer; and The model is applied to a test feature carrier corresponding to the test subject to determine whether the test subject suffers from cancer.

5. A method for predicting whether a test fragment from a test subject suspected of having cancer has an abnormal methylation pattern, characterized in that: The method comprises the steps of: Accessing a data structure comprising: counts of strings of a plurality of CpG sites in a reference genome, and respective methylation states of the plurality of CpG sites from a set of training fragments; Generating a test status vector for a test fragment, wherein the test status vector comprises: a test genomic position within the reference genome and a methylation state of each CpG site in the test fragment, wherein each methylation state is determined to be one of methylated, unmethylated, and indeterminate; calculating a test probability for the test state carrier based on the plurality of counts stored in the data structure; Sampling a subset of possible methylation state vectors from the test genomic position that is the same length as the test state vector; For each of the plurality of sampled possible methylation state carriers, calculating a probability corresponding to the sampled possible methylation state carrier based at least in part on the plurality of counts stored in the data structure; calculating a proportion of the plurality of sampled possible methylation state carriers that corresponds to a calculated probability that is less than or equal to the test probability; generating an estimated score for the test segment based on the calculated ratio; and Based on the estimated score, it is determined whether the test fragment is likely to have an abnormal methylation pattern.

6. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores a plurality of executable instructions for determining whether a test subject has cancer. When the plurality of executable instructions are executed by a hardware processor, the hardware processor performs the following steps: accessing a model obtained by a training process using a set of cancer segments from one or more training subjects suffering from cancer and a set of non-cancer segments from one or more training subjects not suffering from cancer, wherein both the set of cancer segments and the set of non-cancer segments include a plurality of training segments, wherein the training process comprises: For each training segment, determining whether the training segment is hypomethylated or hypermethylated, wherein each of the hypomethylated and hypermethylated training segments respectively includes at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites is unmethylated or methylated, For each CpG site in a reference genome: quantifying a count of a plurality of low-methylated training segments overlapping the CpG site and a count of a plurality of high-methylated training segments overlapping the CpG site; and generating a low methylation score and a high methylation score based on the counts of a plurality of low methylation training segments and a plurality of high methylation training segments; For each training segment, generating an aggregate low methylation score based on the low methylation scores of the multiple CpG sites in the training segment, and generating an aggregate high methylation score based on the high methylation scores of the multiple CpG sites in the training segment; For each training subject: Ranking the plurality of training segments based on the aggregated low methylation scores, and ranking the plurality of training segments based on the aggregated high methylation scores; and generating a feature vector based on the ranking of the plurality of training segments; Obtaining a plurality of training feature vectors for the one or more training subjects not suffering from cancer, and obtaining a plurality of training feature vectors for the one or more training subjects suffering from cancer; training the model using the plurality of feature vectors for the one or more training subjects not suffering from cancer and the plurality of feature vectors for the one or more training subjects suffering from cancer; and The model is applied to a test feature carrier corresponding to the test subject to determine whether the test subject suffers from cancer.

7. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores executable instructions for determining whether a test subject has cancer. When the executable instructions are executed by a hardware processor, the hardware processor performs the following steps: accessing a model obtained by a training process using a set of cancer training segments from one or more training subjects suffering from cancer and a set of non-cancer training segments from one or more training subjects not suffering from cancer, wherein both the set of cancer training segments and the set of non-cancer training segments include a plurality of training segments, wherein the training process comprises: For each training segment, determining whether the training segment is hypomethylated or hypermethylated, wherein each of the hypomethylated and hypermethylated training segments respectively includes at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites is unmethylated or methylated, For each training subject, generating a training feature vector based on the plurality of hypomethylated training segments and the plurality of hypermethylated training segments; and training the model using the plurality of training feature vectors from the one or more training subjects not suffering from cancer and the plurality of feature vectors from the one or more training subjects suffering from cancer; and The model is applied to a test feature carrier corresponding to the test subject to determine whether the test subject suffers from cancer.

8. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores a plurality of executable instructions for determining whether a test fragment from a test subject suspected of having cancer has an abnormal methylation pattern. When the plurality of executable instructions are executed by a hardware processor, the hardware processor performs the following steps, including: Accessing a data structure comprising: counts of strings of a plurality of CpG sites in a reference genome, and respective methylation states of the plurality of CpG sites from a set of training fragments; Generating a test status vector for a test fragment, wherein the test status vector comprises: a test genomic position within the reference genome and a methylation state of each CpG site in the test fragment, wherein each methylation state is determined to be one of methylated, unmethylated, and indeterminate; calculating a test probability for the test state carrier based on the plurality of counts stored in the data structure; Sampling a subset of possible methylation state vectors from the test genomic position that is the same length as the test state vector; For each of the plurality of sampled possible methylation state carriers, calculating a probability corresponding to the sampled possible methylation state carrier based at least in part on the plurality of counts stored in the data structure; calculating a proportion of the plurality of sampled possible methylation state carriers that corresponds to a calculated probability that is less than or equal to the test probability; generating an estimated score for the test segment based on the calculated ratio; and Based on the estimated score, it is determined whether the test fragment is likely to have an abnormal methylation pattern.

9. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores a plurality of executable instructions, which, when executed by a hardware processor, cause the processor to execute a classifier to diagnose cancer, wherein the classifier is generated by the following process, the process comprising: (a) obtaining a plurality of sequence reads of a set of cancer fragments from one or more subjects suffering from cancer, and obtaining a plurality of sequence reads of a set of non-cancer fragments from one or more subjects not suffering from cancer, wherein both the set of cancer fragments and the set of non-cancer fragments comprise a plurality of sample fragments; (b) for each fragment, determining whether the fragment is hypomethylated or hypermethylated, wherein the hypomethylated and hypermethylated fragments include at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites are unmethylated or methylated; (c) For each CpG site in a reference genome: (i) quantifying a count of a plurality of hypomethylated fragments overlapping the CpG site and a count of a plurality of hypermethylated fragments overlapping the CpG site; and (ii) generating a low methylation score and a high methylation score based on the counts of the plurality of low methylation fragments and the plurality of high methylation fragments; For each subject: (i) ranking the plurality of fragments based on the aggregated low methylation scores, and ranking the plurality of fragments based on the aggregated high methylation scores; and (ii) generating a feature vector based on the ranking of the plurality of fragments; (e) training a diagnostic model based on the plurality of feature vectors generated from the one or more subjects suffering from cancer and the plurality of feature vectors from the one or more subjects not suffering from cancer, the diagnostic model being configured to receive a test feature vector set from a test subject and output a likelihood of suffering from cancer based on the test feature vector set from the test subject; and (f) storing a set of parameters representing the diagnostic model on the non-transitory computer-readable storage medium.

10. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores a plurality of executable instructions, which, when executed by a hardware processor, cause the processor to execute a classifier to diagnose cancer, wherein the classifier is generated by the following process, the process comprising: (a) obtaining a plurality of sequence reads of a set of cancer fragments from one or more subjects suffering from cancer, and obtaining a plurality of sequence reads of a set of non-cancer fragments from one or more subjects not suffering from cancer, wherein both the set of cancer fragments and the set of non-cancer fragments comprise a plurality of sample fragments; (b) determining, for each fragment, whether the fragment has an abnormal methylation pattern, thereby obtaining a group of abnormally methylated fragments; (c) for each abnormally methylated fragment, determining whether the abnormally methylated fragment is hypomethylated or hypermethylated, wherein the hypomethylated and hypermethylated fragments respectively include at least a threshold number of a plurality of CpG sites, and at least a threshold percentage of the plurality of CpG sites are unmethylated or methylated; (d) For each CpG site in a reference genome: (i) quantifying a count of a plurality of hypomethylated fragments overlapping the CpG site and a count of a plurality of hypermethylated fragments overlapping the CpG site; and (ii) generating a low methylation score and a high methylation score based on the counts of the plurality of low methylation fragments and the plurality of high methylation fragments; (e) For each subject: (i) ranking the plurality of fragments based on the aggregated low methylation scores, and ranking the plurality of fragments based on the aggregated high methylation scores; and (ii) generating a feature vector based on the ranking of the plurality of fragments; (f) training a diagnostic model based on the plurality of feature vectors generated from the one or more subjects suffering from cancer and the plurality of feature vectors from the one or more subjects not suffering from cancer, the diagnostic model being configured to receive a test feature vector set from a test subject and output a likelihood of suffering from cancer based on the test feature vector set from the test subject; and (g) storing a set of parameters representing the diagnostic model on the non-transitory computer-readable storage medium.

Images