Component mixing model for tissue identification in DNA samples

Abstract

A method and system for component deconvolution through a hybrid model based on methylation information is disclosed. The hybrid model may be trained in a manner independent of marked or known component contributions. The system generates a methylation feature for each training sample of a plurality of training samples. The methylation signature may be based on counts or percentages of methylation variation (s) expressed in methylation sequence reads of the training sample at each of the plurality of genomic regions. The system may train the hybrid model using maximum likelihood estimates to deconvolve component contributions. The hybrid model may include a group sub-model and a deconvolution sub-model. The compositional molecular model predicts compositional likelihood based on methylation characteristics. The deconvolution sub-model predicts component contributions based on component possibilities.

Description

Cross - Reference to Related Applications

[0001] This application claims the benefit of U.S. Provisional Application No. 63 / 417,616, filed Oct. 19, 2022, which is hereby incorporated by reference in its entirety. Background Art Technical Field

[0002] DNA methylation plays an important role in regulating gene expression. Aberrant DNA methylation has been associated with many disease processes, including cancer. DNA methylation profiling using methylation sequencing (e.g., whole-genome bisulfite sequencing (WGBS)) is increasingly recognized as an important 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 cell-free (cf) DNA. When training a cancer classifier, the classifier is limited by sample noise. For example, the sample may be misdiagnosed, may include confounding tissue signals, or may include multiple genetic features from different clonal populations in a heterogeneous tumor.

[0003] This disclosure relates to addressing the above challenges. The background description provided herein is intended to generally introduce the background information of the disclosure. Unless otherwise indicated herein, the materials described in this section are not prior art to the claims of this application and will not be recognized as prior art or an inspiration of prior art by being incorporated in this section. Summary of the Invention

[0004] Early detection of a subject's disease state, such as cancer, is important because it allows for earlier treatment and thus a greater chance of survival. Sequencing of DNA fragments in a cell-free (cf) DNA sample can be used to identify features that can be used for disease classification. For example, in cancer assessment, cell-free DNA-based features from a blood sample, such as the presence or absence of somatic variants, methylation status, or other genetic aberrations, can provide insights into whether a subject is likely to have cancer and further insights into what type of cancer the subject may have. To this end, this specification includes systems and methods for analyzing cell-free DNA (cfDNA) sequencing data to determine the likelihood that a subject has a disease.

[0005] The present disclosure solves the above problems by providing an improved system and method for cancer classification. The system trains a mixture model to deconvolve the component contributions (also referred to as "fractions") of DNA fragment pairs to training samples. The mixture model can be trained in a manner independent of labeled or known component fractions. The system generates methylation features for each of a plurality of training samples. The methylation features are based on the count or percentage of alternative methylation variants expressed in the methylation sequence reads of the training sample at each of a plurality of genomic regions. The system can use maximum likelihood estimation to train the mixture model to deconvolve the component contributions. The mixture model can include a component submodel and a deconvolution submodel. The component submodel predicts component likelihoods based on the methylation features. The component likelihood is the predicted likelihood that a nucleic acid molecule represented by the methylation features is derived from a component. When training the mixture model, the number of component submodels can be adjusted. The deconvolution submodel predicts component fractions based on the component likelihoods.

[0006] The mixture model can have various applications within the broader scheme of cancer classification.

[0007] In one or more embodiments, the mixture model can be used for tissue purity assessment. Tissue purity assessment can involve examining whether a sample labeled as having known component contributions actually has such component contributions. Tissue purity assessment can also include evaluating whether a cancer sample has a significant tissue contribution for training a cancer classifier. Tissue purity assessment can also incorporate contamination detection. For example, if a sample is said to be derived from a specific tissue but has a predicted tissue fraction that is different, the sample can be considered contaminated.

[0008] In one or more embodiments, the mixture model can be used to predict the cancer type of a test subject, e.g., the type of the primarily affected organ. The mixture model can be used to output the main component fractions corresponding to the predicted cancer type. In one or more embodiments, the mixture model can be used to quantify the cancer signal in a test subject.

[0009] In one or more embodiments, the analysis system can utilize the component submodel to determine the origin of methylation sequence reads. In one or more embodiments, the analysis system can extract the learned information from the mixture model. In one embodiment, the analysis system can learn the correlations between methylation variants at genomic regions that contribute to the component fractions. For example, the analysis system can identify methylation variants that provide information about a specific component. In another embodiment, the analysis system can identify methylation variants at genomic regions associated with non-cancerous impurities.

[0010] Clause 1. A method for training a machine learning hybrid model for identifying tissue types, the method comprising: obtaining a training sample set that includes at least one thousand methylation sequence reads obtained by sequencing DNA fragments; modifying each training sample to produce a corresponding sample methylation signature by determining, for each of a plurality of genomic regions, a first set of methylation sequence reads that overlap the genomic region and a second set of methylation sequence reads that include methylation variations at the genomic region, the sample methylation signature being generated at least in part based on the first set of methylation sequence reads and the second set of methylation sequence reads; generating a training data set that includes the sample methylation signatures; and training the machine learning hybrid model using the training data set, the machine learning hybrid model being configured to identify the contribution of each of a plurality of source tissue types to DNA fragments in a sample.

[0011] Clause 2. The method of Clause 1, wherein at least one training sample is known to include a first source tissue type of the plurality of source tissue types, and wherein training the hybrid model includes training the hybrid model to identify the contribution of the first source tissue type to DNA fragments in the one training sample.

[0012] Clause 3. The method according to any one of Clauses 1 to 2, wherein at least one training sample is known to have the contribution of DNA fragments from each of the plurality of source tissue types, and wherein training the hybrid model includes training the hybrid model to identify the contribution of each of the plurality of source tissue types to DNA fragments in the one training sample.

[0013] Clause 4. The method according to any one of Clauses 1 to 3, wherein at least one training sample is one of the following: a liquid biopsy sample, a tissue biopsy sample, and a purified sample.

[0014] Clause 5. The method according to any one of Clauses 1 to 4, wherein at least one genomic region consists of one CpG site.

[0015] Clause 6. The method according to any one of Clauses 1 to 5, wherein at least one genomic region includes a plurality of CpG sites.

[0016] Clause 7. The method according to any one of Clauses 1 to 6, further comprising: determining the average sequencing depth of each of a set of initial genomic regions based on the methylation sequence reads of the training samples; and filtering out genomic regions having an average sequencing depth below a threshold depth to select the plurality of genomic regions.

[0017] Clause 8. The method according to any one of Clauses 1 to 7, wherein the methylation variation at the genomic region is one of two methylation patterns at the genomic region.

[0018] Clause 9. The method according to Clause 8, wherein the two methylation patterns at a genomic region include methylation and non-methylation.

[0019] Clause 10. The method according to any one of Clauses 1 to 9, wherein the methylation variation at the genomic region is one of more than two methylation patterns at the genomic region.

[0020] Clause 11. The method according to any one of Clauses 1 to 10, wherein modifying each training sample to generate the corresponding sample methylation signature further comprises: for each of the plurality of genomic regions, determining a third set of methylated sequence reads having a reference state at the genomic region, wherein the reference state is any methylation pattern that does not belong to the methylation variation, and wherein the sample methylation signature is further generated based on the third set of methylated sequence reads.

[0021] Clause 12. The method according to any one of Clauses 1 to 11, wherein the tissue types include combinations of the following: non-cancerous impurities; squamous cell carcinoma tissue; skin cancer tissue; melanoma tissue; lung cancer tissue; lung adenocarcinoma tissue; lung squamous cell carcinoma tissue; peritoneal cancer tissue; gastrointestinal cancer tissue; pancreatic cancer tissue; cervical cancer tissue; ovarian cancer tissue; liver cancer tissue; liver tumor tissue; hepatocellular carcinoma tissue; bladder cancer tissue; testicular cancer tissue; breast cancer tissue; brain cancer tissue; colon cancer tissue; rectal cancer tissue; colorectal cancer tissue; endometrial or uterine cancer tissue; salivary gland cancer tissue; kidney cancer or renal cell carcinoma tissue; prostate cancer tissue; vulvar cancer tissue; thyroid cancer tissue; anal canal cancer tissue; penile cancer tissue; head and neck cancer tissue; esophageal cancer tissue; and nasopharyngeal carcinoma (NPC) tissue.

[0022] Clause 13. The method according to Clause 12, wherein the non-cancerous impurities include one or more of lymphocytes, macrophages, fibroblasts, vascular endothelial cells, or non-cancerous tissues.

[0023] Clause 14. The method according to any one of Clauses 12 to 13, wherein the methylation signature of the non-cancerous impurities is obtained from a reference database that contains a plurality of methylation signatures of non-cancerous impurities.

[0024] Clause 15. The method according to any one of Clauses 1 to 14, wherein the machine learning hybrid model is trained according to maximum likelihood estimation.

[0025] Clause 16. The method according to any one of Clauses 1 to 15, wherein training the machine learning hybrid model comprises adjusting the number of tissue types as a hyperparameter of the machine learning hybrid model.

[0026] Clause 17. The method according to Clause 16, wherein adjusting the number of the tissue types as a hyperparameter of the machine learning hybrid model comprises: for each number of source tissue types within a range: training the machine learning hybrid model with that number as the hyperparameter, determining the maximum likelihood by cross - validating the trained machine learning hybrid model with a held - out sample set, and imposing a penalty on the maximum likelihood based on that number; and selecting the optimal number from the range as the hyperparameter based on the penalized maximum likelihood.

[0027] Clause 18. The method according to any one of Clauses 1 to 17, wherein training the machine learning hybrid model comprises training according to one or more machine learning algorithms.

[0028] Clause 19. The method according to any one of Clauses 1 to 18, wherein the machine learning hybrid model comprises a first set of tissue type models, each tissue type model modeling the methylation characteristics of DNA fragments of a source tissue type, and wherein training the machine learning hybrid model comprises training the first set of tissue type models.

[0029] Clause 20. The method according to Clause 19, wherein training the first sub - models comprises training each tissue type model according to a beta distribution.

[0030] Clause 21. The method according to any one of Clauses 1 to 20, wherein the machine learning hybrid model comprises a deconvolution model for deconvolving the contribution of the source tissue type for each training sample, and wherein training the machine learning hybrid model comprises training the deconvolution model.

[0031] Clause 22. The method according to Clause 21, wherein training the deconvolution model comprises training the deconvolution model according to a binomial distribution.

[0032] Clause 23. The method according to any one of Clauses 1 to 22, wherein the machine learning hybrid model comprises: one or more sub - models at a first level for predicting the contribution of a macroscopic source tissue type; and one or more sub - models at a second level for predicting the contribution of the source tissue types under the macroscopic tissue type; and wherein a first - level sub - model predicts the contribution of a macroscopic tissue type, and a set of one or more second - level sub - models predicts the contribution of a set of tissue types under the one macroscopic tissue type, the contribution of the set of tissue types being equal to the contribution of the one macroscopic tissue type.

[0033] Clause 24. A method for identifying the contribution of a source tissue type to DNA fragments in a test sample, the method comprising: obtaining the test sample, the test sample comprising at least one thousand methylated sequence reads of the DNA fragments in the test sample; for each of a plurality of genomic regions: determining a first set of methylated sequence reads overlapping the genomic region and determining a second set of methylated sequence reads having an alternative methylation signature at the genomic region, generating a sample methylation signature of the training sample across the plurality of genomic regions based on the first set of methylated sequence reads and the second set of methylated sequence reads; and applying a machine learning mixture model to the sample methylation signature to identify the contribution of each source tissue type to the DNA fragments in the test sample, optionally wherein the machine learning mixture model is trained according to the method of any one of Clauses 1 to 23.

[0034] Clause 25. The method according to Clause 24, further comprising: reporting the contribution of the identified source tissue type to a client device.

[0035] Clause 26. The method according to any one of Clauses 24 to 25, further comprising: identifying the source tissue type with the greatest contribution and reporting treatment recommendations for treating cancer of the primary source having the identified source tissue type.

[0036] Clause 27. The method according to any one of Clauses 24 to 26, further comprising: comparing the contribution of the source tissue type with a previous contribution of the predicted source tissue type before treatment; determining a treatment effect based on the comparison; and reporting treatment recommendations based on the treatment effect.

[0037] Clause 28. The method according to any one of Clauses 24 to 27, further comprising: determining that the treatment is successful if the contribution of a first source tissue type is less than the previous contribution of the first source tissue type predicted before the treatment; and reporting the success of the treatment.

[0038] Clause 29. The method according to any one of Clauses 24 to 28, further comprising: determining that the treatment has failed if the contribution of a first source tissue type is greater than the previous contribution of the first source tissue type predicted before the treatment; and reporting the failure of the treatment.

[0039] Clause 30. The method according to Clause 29, further comprising: reporting a treatment modification based on the failure of the treatment.

[0040] Clause 31. The method according to any one of Clauses 24 to 30, wherein the test sample is predicted to have cancer.

[0041] Clause 32. The method as described in Clause 31, wherein the test sample is predicted to have cancer by the cancer classifier.

[0042] Clause 33. The method as described in Clause 32, wherein the cancer classifier is trained on a cancer cohort of training samples and a non-cancer cohort of training samples, and wherein each training sample from the cancer cohort and the non-cancer cohort contains at least one thousand methylation sequence reads of DNA fragments in the training sample.

[0043] Clause 34. The method as described in Clause 24, further comprising: applying the machine learning hybrid model to the sample methylation features to determine a subset of methylation sequence reads originating from non-cancer impurity types; excluding the subset of methylation sequence reads originating from the non-cancer impurity types, thereby generating a feature set of methylation sequence reads; and applying the cancer classifier to the feature set of methylation sequence reads to predict cancer in the test sample.

[0044] Clause 35. A method for training a cancer classifier, the method comprising: obtaining a cancer cohort of training samples and a non-cancer cohort of training samples, and wherein each training sample from the cancer cohort and the non-cancer cohort contains at least one thousand methylation sequence reads of DNA fragments in the training sample; generating sample methylation for each training sample by: for each of a plurality of genomic regions: determining a first set of methylation sequence reads overlapping the genomic region, determining a second set of methylation sequence reads having alternative methylation features at the genomic region, and wherein the sample methylation features are partially based on the first set of methylation sequence reads and the second set of methylation sequence reads, and applying a machine learning hybrid model optionally trained according to the method of any one of Clauses 1 to 23 to the sample methylation features of each training sample in the cancer cohort to identify a subset of methylation sequence reads originating from non-cancer impurity types; for each training sample in the cancer cohort, excluding the subset of methylation sequence reads originating from the non-cancer impurity types, thereby generating a feature set of methylation sequence reads; generating a feature set of methylation sequence reads for each training sample in the non-cancer cohort; and training the cancer classifier using the feature set of methylation sequence reads of the training samples from the cancer cohort and the feature set of methylation sequence reads of the training samples from the non-cancer cohort.

[0045] Clause 36. The method as described in Clause 35, wherein the cancer classifier is trained as a machine learning model.

[0046] Clause 37. A method for training a cancer classifier, the method comprising: obtaining, for each training sample among a plurality of training samples, a set of methylation sequence reads obtained by sequencing DNA fragments of the training sample, wherein the plurality of training samples includes training samples obtained from a first cohort of subjects diagnosed with cancer and training samples obtained from a second cohort of subjects not diagnosed with cancer; for each training sample in the first cohort and for each methylation sequence read of the training sample: applying a tissue type model trained to predict the likelihood that the methylation sequence read is from non-cancerous tissue, and excluding the methylation sequence read if the predicted likelihood is higher than a threshold; generating, based on the non-excluded methylation sequence reads of the training sample, methylation features for each training sample in the first cohort; generating, based on the methylation sequence reads of the training sample, methylation features for each training sample in the second cohort; and training a cancer classifier using the methylation features of the training samples to detect the presence of a cancer signal.

[0047] Clause 38. The method according to clause 37, wherein the tissue type model is trained according to a beta distribution using methylation features of non-cancer training samples, the methylation features being generated from methylation sequence reads obtained by sequencing DNA fragments of the non-cancer training samples.

[0048] Clause 39. The method according to clause 38, wherein the methylation features of the non-cancer training samples are obtained from a reference database.

[0049] Clause 40. The method according to any one of clauses 37 to 39, wherein the cancer classifier is trained as a machine learning model.

[0050] Clause 41. The method according to any one of clauses 37 to 40, wherein the methylation features are based on the identification of one or more methylation variations at each of a plurality of genomic regions.

[0051] Clause 42. A method for identifying the source tissue type of a DNA fragment, the method comprising: obtaining methylation sequence reads of the DNA fragment, the methylation sequence reads comprising methylation features on one or more genomic regions; predicting, by applying each of a plurality of tissue type models to the methylation features, the likelihood that the methylation features of the methylation sequence read are from each of a plurality of source tissue types, wherein the plurality of tissue type models are optionally part of a mixture model trained according to the method of any one of clauses 1 to 23; determining that the DNA fragment is from the source tissue type with the highest likelihood; and returning a report of the determination.

[0052] Clause 43. A non-transitory computer-readable storage medium storing instructions that, when executed by a computer processor, cause the computer processor to perform the method according to any one of Clauses 1 to 42.

[0053] Clause 44. A system comprising: a computer processor; and a non-transitory computer-readable storage medium according to Clause 43.

[0054] Clause 45. A computer program product comprising a non-transitory computer-readable storage medium storing a machine learning hybrid model for predicting the proportion of tissue types from a test sample, wherein the product is manufactured by the method according to any one of Clauses 1 to 23.

[0055] Clause 46. A computer program product comprising a non-transitory computer-readable storage medium storing a machine learning cancer classifier for predicting cancer in a test sample, wherein the product is manufactured by the method according to any one of Clauses 33 to 41.

[0056] Clause 47. A treatment kit comprising: a collection container for collecting a DNA sample from a subject; optionally, one or more reagents for separating DNA fragments in the DNA sample; optionally, one or more probes targeting one or more loci determined to indicate a cancer status; and a non-transitory computer-readable storage medium according to Clause 43 or a computer program product according to Claim 45 or 46. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is an exemplary flowchart depicting the overall workflow for classifying a sample for cancer according to one or more embodiments.

[0058] Figure 2A is an exemplary flowchart of an apparatus for sequencing a nucleic acid sample according to one or more embodiments.

[0059] Figure 2B is a block diagram of an analysis system for processing a DNA sample according to one or more embodiments.

[0060] Figure 3A is an exemplary flowchart depicting the process of sequencing fragments of cell-free (cf) DNA to obtain a methylation status vector according to one or more embodiments.

[0061] Figure 3B is according to one or more embodiments for Figure 2AAn exemplary illustration of the process of sequencing fragments of cell-free (cf) DNA to obtain a methylation status vector.

[0062] Figure 4A is an exemplary flowchart that depicts the process of training a mixture model for predicting the proportion of components in a sample according to one or more embodiments.

[0063] Figure 4B is an exemplary flowchart that depicts the process of deploying a mixture model for predicting the proportion of components in a sample according to one or more embodiments.

[0064] Figure 5A Shows methylation features according to one or more embodiments, which may be derived from individual CpG sites as genomic regions.

[0065] Figure 5B Shows methylation features according to one or more embodiments, which may be derived from multiple CpG sites as genomic regions.

[0066] Figure 6 Shows an example methylation feature matrix for training a mixture model according to an example implementation.

[0067] Figure 7 Shows an exemplary architecture of a mixture model according to one or more embodiments.

[0068] Figure 8A is an exemplary flowchart that depicts the process of generating a control group data structure to identify abnormally methylated fragments according to one or more embodiments.

[0069] Figure 8B is an exemplary flowchart that depicts the process of determining that a fragment is abnormally methylated based on a control group data structure according to one or more embodiments.

[0070] Figure 9A is an exemplary flowchart that depicts the process of training a cancer classifier according to one or more embodiments.

[0071] Figure 9B Shows examples of generating feature vectors for training a cancer classifier according to one or more embodiments.

[0072] Figure 10 is an example result of a two-component mixture model according to an example implementation.

[0073] Figure 11 is an example result of a five-component mixture model according to an example implementation.

[0074] Figure 12 These are exemplary results of deconvolving positive hormone receptor status of breast tissue, negative hormone receptor status of breast tissue, prostate tissue, and uterine tissue using a mixture model according to an exemplary embodiment.

[0075] Figure 13A These are exemplary results of deconvolving non-cancerous impurities of colorectal tissue from a colorectal sample using a mixture model according to an exemplary embodiment.

[0076] Figure 13B These are exemplary results of deconvolving non-cancerous impurities of bladder tissue from a bladder sample using a mixture model according to an exemplary embodiment.

[0077] These figures depict multiple embodiments for illustrative purposes only. Those skilled in the art will readily recognize from the following discussion that alternative embodiments of the structures and methods illustrated herein may be employed without departing from the principles described herein. DETAILED DESCRIPTION I. OVERVIEW

[0078] Early detection and classification of cancer is an important technology. Being able to detect cancer before it becomes symptomatic is beneficial to all parties involved, including the patient, doctor, and loved ones. For the patient, early cancer detection allows them a greater chance of a favorable outcome; for the doctor, early cancer detection allows for more treatment options that may lead to a favorable outcome; for loved ones, early cancer detection increases the likelihood of not losing a friend or family member to the disease.

[0079] Recently, early cancer detection technology has advanced towards analyzing genetic fragments (e.g., DNA) of the human body (e.g., blood) to determine if any of those fragments are derived from cancer cells. These new technologies enable doctors to identify cancers present in a patient that may not have been detectable otherwise (e.g., by conventional screening methods). For example, consider the case of a person at high risk of developing breast cancer. Traditionally, this person would visit their doctor regularly for a mammogram, which produces an image of their breast tissue (e.g., an X-ray image), and the doctor would use this to identify cancerous tissue. Unfortunately, even with the highest resolution mammogram, doctors can only identify a tumor when it is approximately one millimeter in size. This means that the cancer has been present in this person for some time and has gone undiagnosed and untreated. For most cancers, such visual determination is common - that is, a cancer can only be identified when the tumor has grown large enough to be detected by some imaging technique.

[0080] Cancer detection by analyzing genetic fragments of a patient (e.g., blood) alleviates this problem. For ease of explanation, it is assumed that once cancer cells form, DNA fragments will break off and enter the bloodstream of a person. This occurs when the number of cancer cells is still very small and before they can be observed by imaging techniques. Therefore, with appropriate methods, a system for analyzing DNA fragments in the bloodstream can identify the presence of cancer in the human body based on the shed cancer DNA fragments. More importantly, the system can identify the cancer before more traditional cancer detection techniques.

[0081] Cancer detection based on DNA fragment analysis can be achieved through next-generation sequencing (“NGS”) technology. Broadly speaking, NGS is a group of technologies that can perform high-throughput sequencing of genetic material. As discussed in more detail herein, NGS mainly consists of (1) sample preparation, (2) DNA sequencing, and (3) data analysis. Sample preparation is the laboratory method necessary for preparing DNA fragments for sequencing. Sequencing is the process of reading the ordered nucleotides in a sample, and data analysis is the process of processing and analyzing the genetic information in the sequencing data to identify the presence of cancer.

[0082] Although these steps of NGS may contribute to early cancer detection, these steps themselves also pose complex and detrimental problems to cancer detection. Therefore, any improvement in sample preparation, DNA sequencing, and / or data analysis (including preprocessing, algorithmic processing, and summary or presentation of predictions or conclusions) will more broadly improve cancer detection techniques and early cancer detection.

[0083] For ease of explanation, for example, (1) the problems brought about by sample preparation include DNA sample quality, sample contamination, fragmentation bias, and accurate indexing. Solving these problems will provide better genetic data for cancer detection.

[0084] Similarly, (2) the problems brought about by sequencing include, for example, errors in accurate transcription of fragments (e.g., reading “C” as “A”, etc.), incorrect or difficult fragment assembly and overlap, inconsistent coverage uniformity, the trade-off between sequencing depth and cost and specificity, and insufficient sequencing length. Likewise, solving any of these problems will provide improved genetic data for cancer detection.

[0085] (3) The problems in data analysis are the most formidable and complex. The large amount of data generated by NGS sequencing technology poses many challenges. The sequencing data of a single sample can be in the order of hundreds of thousands (up to millions) of sequence reads, with a total data volume in the terabyte range. Multiply this by the thousands (up to tens of thousands) of samples collected for training the analysis model. Analyzing such a large amount of data effectively and efficiently is highly demanding both programmatically and computationally. For example, analyzing NGS sequencing involves several baseline processing steps, such as aligning reads with each other, aligning and mapping reads to a reference genome, removing duplicate reads, detecting sample contamination, identifying and recognizing variant genes, identifying and recognizing abnormally methylated genes, generating functional annotations, etc. Performing any of these processing steps on terabyte-scale genetic data is computationally costly even for the most powerful computer architectures and completely impossible for the normal human brain. In addition, since the processes of sample preparation and sequence reading are error-prone, a large portion of the genetic data obtained from these processes, i.e., the resulting genetic data, may be of low quality or unusable for cancer identification. For example, a large amount of genetic data may include contaminated samples, transcriptional errors, mismatched regions, over-represented regions, etc., which may not be suitable for high-accuracy cancer detection. Identifying and processing low-quality genetic data in the large amount of genetic data obtained from NGS sequencing is also extremely difficult both programmatically and computationally and is not practical to be performed by the human brain. Overall, any process that helps to more efficiently process large-scale sequencing data will improve cancer detection using NGS sequencing. In addition, such a process is a solution designed for various obstacles generated in NGS sequences and is thus an unconventional and innovative activity in the field of technology exploration.

[0086] In particular, under (3) data analysis, accurately identifying abnormal DNA from NGS data to determine the presence of cancer is also a formidable task. To achieve effective detection, algorithms are being sought to compensate for errors generated, for example, by sample preparation and sequencing, and to overcome the problems of large-scale data analysis brought about by NGS technology. That is, when designing one or more machine learning models or other computational processing algorithms to achieve early cancer detection based on next-generation sequencing technology, they must be configured to take into account the problems brought about by those technologies. Some of those technologies and models are discussed below, and specific improvements to state-of-the-art technologies and models are further discussed. In addition, such technologies are unconventional and innovative activities in the field of technology exploration.

[0087] Tissue samples pose a particular challenge. Generally, the process of collecting tissue samples requires a tissue biopsy to remove tissue from a subject. The tissue is sliced and placed on a slide, stained (e.g., with hematoxylin and eosin (H&E)), annotated by a pathologist, and dissected based on the pathologist's annotation. The dissected portion is used to isolate nucleic acid fragments belonging to the tissue, for example by lysing the cells and isolating the nucleic acid fragments. Such a process relies on the precise annotation by the pathologist and the accuracy of the dissection of the target portion to obtain nucleic acid fragments. Impurities can be introduced into the tissue sample at various points in this process. For example, if the annotation and / or dissection is inaccurate, the dissected portion may be over-inclusive, including tissue outside the targeted portion. In a cancer tissue sample, this may result in the tissue sample including tumor tissue and healthy tissue, thus reducing the purity of the tissue sample. Or in other cases, the tissue sample may include additional tissue types that are not targeted, thus also reducing the purity of the sample.

[0088] One or more current inventions describe a mixture model that can assess the tissue purity in a sample, thus overcoming the above challenges. Samples considered impure (i.e., the contamination tolerance is higher than the acceptable level) can be removed from the training dataset to prevent bias in downstream analysis. For example, when training a cancer classifier to classify tumor samples of different cancer types, the mixture model can assess the purity of the cancer samples used to train the cancer classifier.

[0089] Another particular challenge arises in the context of tumor heterogeneity. In this case, cancer samples derived from a subject diagnosed with cancer may include heterogeneous tumors. Such tumors are a mixture of two or more cancer subtypes, each of which may have its own tumor biology and / or gene mutations. Using such samples for training under the guise of a single tumor biology may blur the classification lines, thus biasing the prediction accuracy of one or more analysis models.

[0090] One or more inventions describe a mixture model that can resolve different cancer subtypes with different methylation profiles in a heterogeneous tumor sample. The mixture model can be trained to deconvolute different subtypes in a heterogeneous tumor sample. Most samples with two or more cancer subtypes may not be used for training the analysis model. In other example applications, the mixture model can be trained to predict cancer subtypes, for example as a downstream analysis when detecting the likelihood of cancer in a test sample. I.A. Cancer Classification Workflow

[0091] Figure 1is an exemplary flowchart that depicts an overall workflow 100 for classifying a sample for cancer according to one or more embodiments. The workflow 100 is performed by one or more entities (e.g., including healthcare providers, sequencing devices, analysis systems, etc.). The purposes of the workflow include detecting and / or monitoring cancer in an individual. From a healthcare perspective, the workflow 100 can be used to complement other existing cancer diagnostic tools. The workflow 100 can be used to provide early cancer detection and / or routine cancer monitoring to better inform the treatment plan for an individual diagnosed with cancer. The overall workflow 100 can include more / fewer steps than Figure 1 those shown.

[0092] A healthcare provider performs sample collection 110. An individual in need of cancer classification visits their healthcare provider. The healthcare provider collects a sample for cancer classification. Examples of biological samples include but are not limited to a tissue biopsy, blood, whole blood, plasma, serum, urine, cerebrospinal fluid, feces, saliva, sweat, tears, pleural fluid, pericardial fluid, or peritoneal fluid of the subject. The sample includes genetic material belonging to the individual, which can be extracted and sequenced for cancer classification. Once the sample is collected, the sample is provided to a sequencing device. In addition to the sample, the healthcare provider can also collect other information related to the individual, such as physiological sex, age, race, smoking status, any previous diagnoses, etc.

[0093] A sequencing device performs sample sequencing 120. A laboratory clinician can perform one or more processing steps on the sample to prepare it for sequencing. Once prepared, the clinician loads the sample into the sequencing device. In combination with Figure 2A and Figure 2B , examples of devices for sequencing are further described. The sequencing device generally extracts and separates nucleic acid fragments and sequences these fragments to determine the sequence of nucleobases corresponding to these fragments. Sequencing can also include amplification of nucleic acid material. Different sequencing procedures include Sanger sequencing, fragment analysis, and next-generation sequencing. Sequencing can be whole-genome sequencing or targeted sequencing with a target panel. In the context of DNA methylation, bisulfite sequencing (e.g., further described in Figure 3A and Figure 3B ) can determine the methylation status by bisulfite conversion of unmethylated cytosine at CpG sites. Sample sequencing 120 generates sequences of multiple nucleic acid fragments in the sample. In one or more embodiments, these sequences can include methylation status vectors, where each methylation status vector describes the methylation status of CpG sites on the fragment.

[0094] An analysis system performs pre-analysis processing 130. In Figure 2BAn exemplary analysis system is described. The pre - analysis processing 130 may include, but is not limited to, deduplicating sequence reads, determining metrics related to coverage, determining whether a sample is contaminated, removing contaminated fragments, identifying sequencing errors, etc.

[0095] The analysis system performs one or more analyses 140. These analyses are statistical analyses or apply one or more trained models to at least predict the cancer status of the individual from whom the sample is derived. Different genetic features can be evaluated and considered, such as methylation of CpG sites, single nucleotide polymorphisms (SNPs), insertions or deletions (indels), other types of genetic mutations, etc. In the context of methylation, the analysis 140 may include tissue purity assessment 142 (e.g., further described in Figure 4A , 4B , 5A, 5B, 6, and 7), feature extraction 144, and application of a cancer classifier 146 to determine a cancer prediction (e.g., further described in Figure 9A and 9B ). The tissue purity assessment 142 involves applying a mixture model to deconvolute the proportions of tissue components that contribute DNA fragments to the sample. Generally, the tissue purity assessment 142 can be used to determine the proportion of methylated sequence reads shed from cancerous tissue in the sample compared to the proportion of methylated sequence reads shed from non - cancer cells. The tissue purity assessment 142 may be particularly useful in deconvoluting heterogeneous tumors that contain multiple clonal populations with potentially different genetic features. The mixture model can also be used to quantify the cancer signal or grade the cancer status based on the determined proportions. The cancer classifier 146 inputs the extracted features to determine a cancer prediction. The cancer prediction may be a label or a value. The label can indicate a specific cancer status. For example, a binary label can indicate the presence or absence of cancer, and a multi - class label can indicate one or more of the multiple cancer types being screened, cancer stage, etc. The value can indicate the likelihood of a specific cancer status, e.g., the likelihood of having cancer and / or the likelihood of a specific cancer type.

[0096] The analysis system returns the prediction 150 to the healthcare provider. The prediction 150 can include a binary prediction of the presence or absence of cancer, a specific cancer type, cancer stage, tissue proportions, etc. The healthcare provider can formulate or adjust a treatment plan based on the returned prediction 150. The optimization of treatment is further described in Section V.C, "Treatment". I.B. Overview of Methylation

[0097] According to the present specification, cfDNA fragments from an individual are processed, for example, by converting unmethylated cytosine to uracil, followed by sequencing and aligning the sequence reads to a reference genome to identify the methylation status of specific CpG sites within these DNA fragments. Each CpG site may be methylated or unmethylated. Identification of aberrantly methylated fragments can provide clues about the cancer status of a subject compared to a healthy individual. As is well known in the art, aberrant DNA methylation (compared to healthy controls) can have different effects, which may contribute to cancer. Various challenges arise in the identification of aberrantly methylated cfDNA fragments. First, to determine that a DNA fragment is aberrantly methylated, it needs to be compared to a group of control individuals for the determination to be convincing. Thus, if the number of individuals in the control group is small, the determination may lose credibility due to the statistical variability resulting from the small size of the control group. Additionally, there may be differences in methylation status among a group of control individuals, and it is difficult to account for such differences when determining whether a subject's DNA fragment is aberrantly methylated. On the other hand, methylation of a cytosine at one CpG site can have a causal effect on the methylation of subsequent CpG sites. Incorporating this dependence into consideration is itself another challenge.

[0098] Methylation typically occurs in deoxyribonucleic acid (DNA) when a hydrogen atom on the pyrimidine ring of a cytosine base is converted to a methyl group, forming 5-methylcytosine. In particular, methylation can occur at dinucleotides composed of cytosine and guanine, which are referred to herein as "CpG sites". In other cases, methylation may also occur on cytosines that are not part of CpG sites or on another nucleotide that is not cytosine; however, these occurrences are relatively rare. In the present disclosure, for clarity, methylation is discussed with reference to CpG sites. Aberrant DNA methylation can be identified as hypermethylation or hypomethylation, both of which can indicate cancer status. Throughout the present disclosure, hypermethylation and hypomethylation of a DNA fragment can be characterized if the number of CpG sites contained in the DNA fragment exceeds a threshold and a threshold percentage of those CpG sites are either methylated or unmethylated.

[0099] The principles described herein also apply to methylation detection in non-CpG contexts, including methylation of non-cytosines. In such embodiments, the laboratory wet assays used to detect methylation may differ from those described herein. Further, the methylation status vectors discussed herein may contain elements that are typically sites where methylation has or has not occurred (even if those sites are not specifically CpG sites). With this adjustment, the remaining processes described herein can remain the same, and thus the inventive concepts described herein also apply to those other forms of methylation. I.C. Definition

[0100] The term "cell-free nucleic acid" or "cfNA" refers to nucleic acid fragments that circulate in an individual's body (e.g., in the blood), which fragments are derived from one or more healthy cells and / or one or more unhealthy cells (e.g., cancer cells). The term "cell-free DNA" or "cfDNA" refers to deoxyribonucleic acid fragments that circulate in an individual's body (e.g., in the blood). In addition, cfNA or cfDNA in an individual's body may also be from other non-human sources.

[0101] The term "genomic nucleic acid", "genomic DNA" or "gDNA" refers to nucleic acid molecules or deoxyribonucleic acid molecules obtained from one or more cells. In various embodiments, gDNA can be extracted from healthy cells (e.g., non-tumor cells) or tumor cells (e.g., biopsy samples). In some embodiments, gDNA can be extracted from cells derived from the hematopoietic cell lineage (e.g., white blood cells).

[0102] The term "circulating tumor DNA" or "ctDNA" refers to nucleic acid fragments derived from tumor cells or other types of cancer cells, which fragments may be released into an individual's body fluids (e.g., blood, sweat, urine or saliva) due to biological processes such as apoptosis or necrosis of dying cells or active release from surviving tumor cells.

[0103] The terms "DNA fragment", "fragment" or "DNA molecule" generally can refer to any deoxyribonucleic acid fragment, i.e., cfDNA, gDNA, ctDNA, etc.

[0104] The term "abnormal fragment", "abnormally methylated fragment" or "fragment with an abnormal methylation pattern" refers to a fragment with abnormal methylation of CpG sites. Abnormal methylation of a fragment can be determined using a probability model to identify the unexpected degree of the fragment methylation pattern observed in a control group.

[0105] The term "abnormal fragment with extreme methylation" or "UFXM" refers to a hypomethylated fragment or a hypermethylated fragment. Hypomethylated fragments and hypermethylated fragments respectively refer to fragments containing at least a certain number (e.g., 5) of CpG sites and having more than a certain threshold percentage (e.g., 90%) of these sites in a methylated or unmethylated state.

[0106] The term "abnormal score" refers to a score for a CpG site based on the number of abnormal fragments (or UFXM in some embodiments) overlapping with that CpG site in a sample. This abnormal score is used in the context of characterizing a sample for classification.

[0107] As used herein, the terms "about" or "approximately" can mean that a particular value is within an acceptable error range as determined by a person of ordinary skill in the art, which depends in part on how the value is measured or determined, e.g., due to limitations of the measurement system. For example, in accordance with the practice in the art, "about" can mean within 1 or more standard deviations. "About" can mean a range of ±20%, ±10%, ±5%, or ±1% of a given value. The terms "about" or "approximately" can mean within the same order of magnitude of a value, within 5-fold, or within 2-fold. When describing a particular numerical value in the present application and claims, unless otherwise stated, it should be assumed that the term "about" means within the acceptable error range of that particular value. The term "about" can have the meaning commonly understood by a person of ordinary skill in the art. The term "about" can mean ±10%. The term "about" can mean ±5%.

[0108] As used herein, the terms "biological sample", "patient sample", or "sample" refer to any sample taken from a subject that can reflect the biological state associated with the subject and includes cell-free DNA. Examples of biological samples include, but are not limited to, the subject's blood, whole blood, plasma, serum, urine, cerebrospinal fluid, feces, saliva, sweat, tears, pleural fluid, pericardial effusion, or peritoneal fluid. A biological sample can include any tissue or material from a living or dead subject. A biological sample can be a cell-free sample. A biological sample can contain nucleic acids (e.g., DNA or RNA) or fragments thereof. The term "nucleic acid" can refer to deoxyribonucleic acid (DNA), ribonucleic acid (RNA), or any hybrid or fragment thereof. The nucleic acids in the sample can be cell-free nucleic acids. The sample can be a liquid sample or a solid sample (e.g., a cell or tissue sample). A biological sample can be a body fluid, such as blood, plasma, serum, urine, vaginal secretion, fluid in hydrocele (e.g., hydrocele of the testis), vaginal lavage fluid, pleural fluid, ascites, cerebrospinal fluid, saliva, sweat, tears, sputum, bronchoalveolar lavage fluid, nipple discharge, aspirate from different parts of the body (e.g., thyroid, breast), etc. A biological sample can be a fecal sample. In various embodiments, most of the DNA in a biological sample in which cell-free DNA has been enriched (e.g., a plasma sample obtained by a centrifugation protocol) can be cell-free (e.g., greater than 50%, 60%, 70%, 80%, 90%, 95%, or 99% of the DNA can be cell-free). A biological sample can be processed to physically disrupt the tissue or cell structure (e.g., centrifugation and / or cell lysis) to release intracellular components into a solution that can further contain enzymes, buffers, salts, detergents, etc., which can be used to prepare a sample for analysis.

[0109] As used herein, the terms "control," "control sample," "reference," "reference sample," "normal," and "normal sample" describe a sample taken from a subject that does not suffer from a particular disease or is otherwise healthy. In one example, the methods disclosed herein can be applied to a subject suffering from a tumor, wherein the reference sample is a sample taken from healthy tissue of the subject. The reference sample can be obtained from the subject or from a database. The reference can be, for example, a reference genome, which is used to compare the sequences of nucleic acid fragments obtained from sequencing of the subject's sample. The reference genome can refer to a haploid or diploid genome, to which the sequences of nucleic acid fragments from biological samples and constitutional samples can be compared and compared. An example of a constitutional sample can be the DNA of white blood cells obtained from the subject. For a haploid genome, each locus has only one nucleotide. For a diploid genome, heterozygous loci can be identified; each heterozygous locus has two alleles, either of which can match the locus for comparison.

[0110] As used herein, the term "cancer" or "tumor" refers to an abnormal mass of tissue, in which the growth of the mass exceeds and is out of coordination with the growth of normal tissue.

[0111] As used herein, the phrase "healthy" refers to a subject that is in good health. A healthy subject may be shown to be free of any malignant or non-malignant disease. A "healthy individual" may have other diseases or conditions unrelated to the condition being measured that are not generally considered "healthy".

[0112] As used herein, the term "methylation" refers to a modification of deoxyribonucleic acid (DNA) in which 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 consisting of cytosine and guanine, which are referred to herein as "CpG sites". In other cases, methylation may also occur on cytosine that is not part of a CpG site or on another nucleotide that is not cytosine; however, these situations are rare. Abnormal cfDNA methylation can be identified as hypermethylation or hypomethylation, both of which can indicate a cancer state. Abnormal DNA methylation (compared to healthy controls) can have different effects, which may cause cancer. The principles described herein also apply to methylation detection in CpG and non-CpG backgrounds, including methylation of non-cytosine. Further, the methylation state vector may contain some elements, which are usually vectors of sites where methylation has occurred or not occurred (even if those sites are not specifically CpG sites).

[0113] As used interchangeably herein, the term "methylated fragment" or "nucleic acid methylated fragment" refers to a string of sequences for which the methylation status of each CpG site among a plurality of CpG sites in the string is determined by performing methylation sequencing on a nucleic acid (e.g., a nucleic acid molecule and / or a nucleic acid fragment). In a methylated fragment, the position and methylation status of each CpG site in the nucleic acid fragment are determined based on the alignment of these sequence reads (e.g., obtained by sequencing these nucleic acids) to a reference genome. A nucleic acid methylated fragment contains the methylation status (e.g., a methylation status vector) of each CpG site among a plurality of CpG sites, which specifies the position of the nucleic acid fragment in the reference genome (e.g., as specified according to the position of the first CpG site in the nucleic acid fragment or another similar metric using a CpG index) and the number of CpG sites in the nucleic acid fragment. Based on the methylation sequencing of a nucleic acid molecule, a CpG index can be used to perform the alignment of sequence reads to a reference genome. As used herein, the term "CpG index" refers to a list (e.g., CpG 1, CpG 2, CpG 3, etc.) of each CpG site among a plurality of CpG sites in a reference genome (e.g., a human reference genome), which can be in electronic format. The CpG index further contains the corresponding genomic position in the corresponding reference genome of each respective CpG site in the CpG index. Thus, each CpG site in each respective nucleic acid methylated fragment can be indexed to a specific position in the corresponding reference genome determined by using the CpG index.

[0114] As used herein, the term "methylation variant" refers to a different methylation pattern among a set of k (e.g., 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 25, etc.) consecutive CpG sites. Methylation variants can be used to distinguish DNA from a particular cancer type from DNA from non-cancerous plasma. A given methylation variant has a single defined "variant state", i.e., a single specific and detectable pattern of methylation status. A "reference state" is any methylation status pattern that is not a variant state. As one or more examples, a set of 5 consecutive CpGs can have one variant state of 5 methylated CpGs (e.g., MMMMM, where "M" represents methylated), another variant state of 5 unmethylated CpGs (e.g., UUUUU, where "U" represents unmethylated), or a third variant state including a mixed methylation status (e.g., UMMMU). The reference state can include all other patterns that are not considered one of the variant states.

[0115] As used herein, the term "true positive" (TP) refers to a subject having a particular disease. A "true positive" can refer to a subject having a tumor, cancer, pre-cancerous lesion (e.g., pre-cancerous injury), localized or metastatic cancer, or a non-malignant disease. A "true positive" can also refer to a subject having a particular disease and identified as having that disease by the assays or methods disclosed herein. As used herein, the term "true negative" (TN) refers to a subject not having a particular disease or not having a detectable disease. A "true negative" can refer to a subject not having a particular disease or a detectable disease (e.g., tumor, cancer, pre-cancerous lesion (e.g., pre-cancerous injury), localized or metastatic cancer, non-malignant disease) or a subject who is otherwise healthy. A "true negative" can refer to a subject not having a particular disease or not having a detectable disease, or identified as not having that disease by the assays or methods disclosed herein.

[0116] As used herein, the term "reference genome" refers to any known, sequenced, or characterized specific genome, whether partial or complete, of any organism or virus, which can be used to align identified sequences from a subject. Online genome browsers hosted by the National Center for Biotechnology Information ("NCBI") or the University of California, Santa Cruz ("UCSC") provide exemplary reference genomes for human subjects and many other organisms. A "genome" refers to the complete genetic information of an organism or virus expressed as a nucleic acid sequence. As used herein, a reference sequence or reference genome is typically an assembled or partially assembled genomic sequence from one or more individuals. In some embodiments, the reference genome is an assembled or partially assembled genomic sequence from one or more human individuals. A reference genome can be considered a representative instance of the gene set of a species. In some embodiments, the reference genome includes sequences assigned to chromosomes. Exemplary human reference genomes include, but are not limited to, NCBI build 34 (UCSC equivalent: hg16), NCBI build 35 (UCSC equivalent: hg17), NCBI build 36.1 (UCSC equivalent: hg18), GRCh37 (UCSC equivalent: hg19), and GRCh38 (UCSC equivalent: hg38).

[0117] As used herein, the term "sequence read" or "read" refers to a nucleotide sequence generated by any sequencing process described herein or known in the art. A read can be generated from one end of a nucleic acid fragment ("single-end read") or sometimes from both ends of the nucleic acid (e.g., paired-end read, mate-pair read). In some embodiments, sequence reads (e.g., single-end or paired-end reads) can be generated from one or both strands of a targeted nucleic acid fragment. The length of the sequence read is typically associated with a particular sequencing technology. For example, sequence reads provided by high-throughput methods can range in size from a few dozen to several hundred base pairs (bp). In some embodiments, the mean, median, or average length of these sequence reads is about 15 bp to 900 bp long (e.g., about 20 bp, about 25 bp, about 30 bp, about 35 bp, about 40 bp, about 45 bp, about 50 bp, about 55 bp, about 60 bp, about 65 bp, about 70 bp, about 75 bp, about 80 bp, about 85 bp, about 90 bp, about 95 bp, about 100 bp, about 110 bp, about 120 bp, about 130 bp, about 140 bp, about 150 bp, about 200 bp, about 450 bp, about 300 bp, about 350 bp, about 400 bp, about 450 bp, or about 500 bp). In some embodiments, the mean, median, or average length of these sequence reads is about 1000 bp, 2000 bp, 5000 bp, 10,000 bp, or 50,000 bp or longer. For example, nanopore sequencing can provide sequence reads ranging in size from a few dozen to several hundred to several thousand base pairs. Illumina parallel sequencing can provide sequence reads with a less variable size, e.g., most of these sequence reads are less than 200 bp. A sequence read (or sequencing read) can refer to sequence information corresponding to a nucleic acid molecule (e.g., a string of nucleotides). For example, a sequence read can correspond to a string of nucleotides from a portion of a nucleic acid fragment (e.g., about 20 to about 150), can correspond to a string of nucleotides from one or both ends of a nucleic acid fragment, or can correspond to the nucleotides of an entire nucleic acid fragment. A sequence read can be obtained in a variety of ways, e.g., using a sequencing technology or using a probe, e.g., in a hybridization array or a capture probe, or an amplification technique, e.g., polymerase chain reaction (PCR) or linear amplification using a single primer or isothermal amplification.

[0118] As used herein, the term "sequencing" and the like generally refer to any and all biochemical processes that can be used to determine the sequence of a biological macromolecule (e.g., nucleic acid or protein). For example, sequencing data can include all or part of the nucleotide bases in a nucleic acid molecule (e.g., a DNA fragment).

[0119] As used herein, the term "sequencing depth" is used interchangeably with the term "coverage" and refers to the number of times the consensus sequence reads corresponding to unique nucleic acid target molecules aligned to a locus align to that locus; for example, the sequencing depth is equal to the number of unique nucleic acid target molecules covering that locus. The locus can be as small as a single nucleotide, or as large as a chromosomal arm, or as large as an entire genome. The sequencing depth can be expressed as "Yx", such as 50x, 100x, etc., where "Y" refers to the number of times the locus is covered by the sequence corresponding to the target nucleic acid; for example, the number of times information on independent sequences covering that particular locus is obtained. In some embodiments, the sequencing depth corresponds to the number of genomes that have been sequenced. The sequencing depth can also be applied to multiple loci or an entire genome, in which case Y can refer, respectively, to the mean or average number of times a locus or a haploid genome or an entire genome has been sequenced. When referring to the average depth, the actual depths of different loci included in the dataset can cover a range of values. Ultra-deep sequencing can refer to a sequencing depth of at least 100x at a particular locus.

[0120] As used herein, the term "sensitivity" or "true positive rate" (TPR) refers to the number of true positives divided by the sum of the number of true positives and false negatives. Sensitivity can characterize the ability of an assay or method to correctly identify the proportion of a population that truly has a certain disease. For example, sensitivity can characterize the ability of a method to correctly identify the number of subjects with cancer within a population. In another example, sensitivity can characterize the ability of a method to correctly identify one or more markers indicative of cancer.

[0121] As used herein, the term "specificity" or "true negative rate" (TNR) refers to the number of true negatives divided by the sum of the number of true negatives and false positives. Specificity can characterize the ability of an assay or method to correctly identify the proportion of a population that truly does not have a certain disease. For example, specificity can characterize the ability of a method to correctly identify the number of subjects without cancer within a population. In another example, specificity characterizes the ability of a method to correctly identify one or more markers indicative of cancer.

[0122] As used herein, the term "subject" refers to any living or non-living organism, including but not limited to humans (e.g., male humans, female humans, fetuses, pregnant females, children, etc.), non-human animals, plants, bacteria, fungi, or protists. Any human or non-human animal can be a subject, including but not limited to mammals, reptiles, avians, amphibians, fish, ungulates, ruminants, bovines (e.g., cows), equines (e.g., horses), ovines and caprines (e.g., sheep, goats), suines (e.g., pigs), camelids (e.g., camels, llamas, alpacas), monkeys, apes (e.g., gorillas, chimpanzees), ursids (e.g., bears), poultry, dogs, cats, mice, rats, fish, dolphins, whales, and sharks. In some embodiments, the subject is a male or female at any stage (e.g., adult male, adult female, or child). A subject from whom a sample is collected or who is treated with any of the methods or compositions described herein can be of any age and can be an adult, infant, or child.

[0123] As used herein, the term "tissue" can correspond to a group of cells that aggregate together as a functional unit. More than one cell type can be found in a single tissue. Different types of tissues may be composed of different types of cells (e.g., hepatocytes, alveolar cells, or blood cells), but can also correspond to tissues from different organisms (maternal vs. fetal) or to healthy cells vs. tumor cells. The term "tissue" generally can refer to any group of cells found in the human body (e.g., heart tissue, lung tissue, kidney tissue, nasopharyngeal tissue, oropharyngeal tissue). In some aspects, the term "tissue" or "tissue type" or "source tissue type" can be used to refer to the tissue from which a cell-free nucleic acid molecule or fragment is derived. In one example, a viral nucleic acid fragment can be derived from blood tissue. In another example, a viral nucleic acid fragment can be derived from tumor tissue.

[0124] As used herein, the term "genomic" refers to a characteristic of an organism's genome. Examples of genomic characteristics include but are not limited to: those characteristics related to the primary nucleic acid sequence of all or part of the genome (e.g., the presence or absence of nucleotide polymorphisms, indels, sequence rearrangements, mutation frequencies, etc.), the copy number of one or more specific nucleotide sequences within the genome (e.g., copy number, allelic frequency fraction, ploidy of a single chromosome or the entire genome, etc.), the epigenetic state of all or part of the genome (e.g., covalent nucleic acid modifications such as methylation, histone modifications, nucleosome positioning, etc.), and the expression profile of an organism's genome (e.g., gene expression levels, isoform expression levels, gene expression ratios, etc.).

[0125] The purpose of the terms used herein is only to describe a specific situation and is not intended to be limiting. As used herein, the singular forms "a / an" and "the" are also intended to include the plural forms unless the context clearly indicates otherwise. In addition, with respect to the scope in which the terms "including", "includes", "having", "has", "with" or their variants are used in the detailed description and / or claims, such terms are intended to mean inclusion in a manner similar to the term "comprising". I.D. Example Analysis System

[0126] Figure 2A is an exemplary flowchart of a device for sequencing a nucleic acid sample according to one or more embodiments. This illustrative flowchart includes devices such as sequencer 220 and analysis system 200. Sequencer 220 and analysis system 200 can work in series to perform one or more steps in the process.

[0127] In multiple embodiments, sequencer 220 receives an enriched nucleic acid sample 210. As Figure 2A shown, sequencer 220 can include a graphical user interface 225 that enables a user to interact with a specific task (e.g., start sequencing or terminate sequencing) and one or more loading stations 230 for loading sequencing cartridges including these enriched fragment samples and / or for loading buffers necessary to perform the sequencing assay. Thus, once the user of sequencer 220 has provided the necessary reagents and sequencing cartridges to the loading station 230 of sequencer 220, the user can start sequencing by interacting with the graphical user interface 225 of sequencer 220. Once started, sequencer 220 performs sequencing and outputs sequence reads of the enriched fragments in nucleic acid sample 210.

[0128] In some embodiments, sequencer 220 is communicatively coupled to analysis system 200. Analysis system 200 includes a number of computing devices for processing sequence reads to meet various application requirements, such as evaluating the methylation status of one or more CpG sites, variant identification, or quality control. Sequencer 220 can provide the sequence reads to analysis system 200 in the BAM file format. Analysis system 200 can be communicatively coupled to sequencer 220 via wireless, wired, or a combination of wireless and wired communication technologies. Generally, analysis system 200 is configured with a processor and a non-transitory computer-readable storage medium storing computer instructions that, when executed by the processor, cause the processor to process sequence reads or perform one or more steps of any method or process disclosed herein.

[0129] In some embodiments, methods known in the art can be used to align sequence reads with a reference genome to determine alignment position information. The alignment position can generally describe the start position and end position of a region in the reference genome that corresponds to the start nucleotide base and the end nucleotide base of a given sequence read. For methylation sequencing, the concept of alignment position information can be generalized to indicate the first CpG site and the last CpG site included in a sequence read based on its alignment with the reference genome. The alignment position information can further indicate the methylation status and positions of all CpG sites in a given sequence read. The region in the reference genome may be associated with a gene or a fragment of a gene; thus, the analysis system 200 can label the sequence read with one or more genes that are aligned with the sequence read. In one embodiment, the fragment length (or size) can be determined by the start position and the end position.

[0130] In multiple embodiments, for example, when using a paired-end sequencing protocol, a sequence read consists of a pair of reads, denoted as R_1 and R_2. For example, the first read R_1 can be sequenced from one end of a double-stranded DNA (dsDNA) molecule, while the second read R_2 is sequenced from the other end of the dsDNA molecule. Thus, the nucleotide base pairs of the first read R_1 and the second read R_2 can always be aligned with the nucleotide bases of the reference genome in a consistent manner (e.g., in opposite directions). The alignment position information derived from the read pair R_1 and R_2 may include a start position in the reference genome that corresponds to one end of the first read (e.g., R_1), and an end position in the reference genome that corresponds to one end of the second read (e.g., R_2). In other words, the start position and the end position in the reference genome can represent the possible positions of the nucleic acid fragment in the reference genome. An output file in SAM (Sequence Alignment / Map) format or BAM (binary) format can be generated and output for further analysis.

[0131] Figure 2B is a block diagram of an analysis system 200 for processing a DNA sample according to one or more embodiments. The analysis system employs one or more computing devices for analyzing a DNA sample. The analysis system 200 includes a sequence processor 240, a sequence database 245, a model database 255, a model 250, a parameter database 265, and a scoring engine 260. In some embodiments, the analysis system 200 performs some or all of the processes described throughout this disclosure.

[0132] The sequence processor 240 generates a methylation status vector for the fragments from the sample. For each CpG site on a fragment, the sequence processor 240 generates a methylation status vector for each fragment that indicates the location of the fragment in the reference genome, the number of CpG sites within the fragment, and the methylation status (methylated, unmethylated, or indeterminate) of each CpG site in the fragment. This process is implemented by the process 200 in Figure 2A . The sequence processor 240 may store the methylation status vectors of the respective fragments in the sequence database 245. The data in the sequence database 245 may be organized such that the methylation status vectors from the same sample are associated.

[0133] Further, multiple different models 250 may be stored in or retrieved from the model database 255 for testing the sample. Generally, a model receives an input and generates an output based on a function that operates on the input according to one or more parameters. In one instance, the model is a trained mixture model for deconvolving the component proportions based on the methylation features of the sample. The mixture model may include various submodels each having its own parameters and (multiple) functions. In another instance, the model is a trained cancer classifier that uses a feature vector derived from aberrant fragments to determine the cancer prediction of the test sample. The training and use of the cancer classifier will be further discussed in Part IV, "Cancer Classifier for Determining Cancer". The analysis system 200 may train one or more models 250 and store the various trained parameters in the parameter database 265. The analysis system 200 stores the model 250 together with the associated functions in the model database 255.

[0134] During inference, the scoring engine 260 uses one or more models 250 to return an output. The scoring engine 260 accesses the models 250 in the model database 255 and the trained parameters from the parameter database 265. For each model, the scoring engine receives an input suitable for the model and calculates an output based on the received input, the parameters, and the functions associated with the input and output in each model. In some use cases, the scoring engine 260 further calculates a metric related to the confidence of the output calculated by the model. In other use cases, the scoring engine 260 calculates other intermediate values in the model. II. Methylation Sequencing of DNA Fragments

[0135] Figure 3Ais an exemplary flowchart that depicts process 300 for sequencing fragments of cfDNA to obtain a methylation status vector according to one or more embodiments. To analyze DNA methylation, the analysis system first obtains 310 a sample containing a plurality of cfDNA molecules from an individual. In additional embodiments, process 300 can be applied to sequence other types of DNA molecules. Process 300 is Figure 1 an embodiment of sample sequencing 120.

[0136] The analysis system can isolate 310 each cfDNA molecule from the sample. These cfDNA molecules can be processed 320 to convert unmethylated cytosine to uracil. In one embodiment, the method uses bisulfite treatment of DNA, which converts unmethylated cytosine to uracil without converting methylated cytosine. For example, commercially available kits can be used for bisulfite conversion, such as EZ DNA Methylation TM -Gold, EZDNAMethylation TM -Direct or EZ DNAMethylation TM -Lightning kits (from Zymo Research Corporation (Irvine, CA)). In another embodiment, the conversion of unmethylated cytosine to uracil is accomplished by 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 (Ipswich, MA, NEBiolabs).

[0137] From the converted cfDNA molecules, a sequencing library can be prepared 330. During library preparation, unique molecular identifiers (UMIs) can be added to nucleic acid molecules (e.g., DNA molecules) by adapter ligation. A UMI can be a short nucleic acid sequence (e.g., 4 - 10 base pairs) that is added to both ends of a DNA fragment (e.g., a DNA molecule cleaved by physical shearing, enzymatic digestion, and / or chemical cleavage) during adapter ligation. A UMI can be a degenerate base pair that serves as a unique tag for identifying sequence reads derived from a particular DNA fragment. During PCR amplification after adapter ligation, the UMI is replicated along with the ligated DNA fragment. This provides a way to identify sequence reads from the same original fragment in downstream analysis.

[0138] Optionally, multiple hybridization probes can be used to enrich cfDNA molecules or genomic regions in a sequencing library that can provide information about the cancer state. These hybridization probes are short oligonucleotides that can hybridize to specifically designated cfDNA molecules or target regions and enrich those fragments or regions for subsequent sequencing and analysis. Hybridization probes can be used to perform targeted, high-depth analysis of a specified set of CpG sites of interest to a researcher. Hybridization probes can be tiled across one or more target sequences at a coverage of 1X, 2X, 3X, 4X, 5X, 6X, 7X, 8X, 9X, 10X, or greater than 10X. For example, hybridization probes tiled at 2X coverage contain overlapping probes such that each part of the target sequence hybridizes to two independent probes. Hybridization probes can be tiled across one or more target sequences at a coverage less than 1X.

[0139] In one embodiment, hybridization probes are designed to enrich DNA molecules that have been treated (e.g., using bisulfite) to convert unmethylated cytosine to uracil. During enrichment, the hybridization probes (also referred to herein as "probes") can be used to target and capture nucleic acid fragments that can provide information about the presence or absence of cancer (or disease), cancer state, or cancer classification (e.g., cancer type or tissue of origin). These probes can be designed to anneal (or hybridize) to the target (complementary) strand of the DNA. The target strand can be the "plus" strand (e.g., the strand that is transcribed into mRNA and subsequently translated into protein) or the complementary "minus" strand. The length of the probes can range from dozens, hundreds, to thousands of base pairs. The probes can be designed based on a set of methylation sites. The probes can be designed based on a set of target genes to analyze specific mutations or target regions in a genome (e.g., of a human or other organism) suspected of corresponding to certain cancers or other types of diseases. Additionally, the probes can cover overlapping portions of the target region.

[0140] Once prepared, the sequencing library or a portion thereof can be sequenced 340 to obtain a plurality of sequence reads. These sequence reads may be in a computer-readable digital format for processing and interpretation by computer software. These sequence reads may be aligned to a reference genome to determine alignment position information. The alignment position information can indicate the start and end positions of a region in the reference genome that corresponds to the start and end nucleotide bases of a given sequence read. The alignment position information can also include the sequence read length, which can be determined from the start and end positions. A region in the reference genome may be associated with a gene or a fragment of a gene. A sequence read can consist of a pair of reads, denoted as R1 and R SFor example, the first read R1 can sequence from one end of a nucleic acid fragment, while the second read R2 sequences from the other end of the nucleic acid fragment. Thus, the nucleotide base pairs of the first read R1 and the second read R2 can always maintain a consistent alignment pattern (e.g., opposite directions) with the nucleotide bases of the reference genome. The alignment position information derived from the read pair R1 and R2 may include a starting position in the reference genome that corresponds to one end of the first read (e.g., R1), and a terminating position in the reference genome that corresponds to one end of the second read (e.g., R2). In other words, the starting position and the terminating position in the reference genome can represent the possible positions corresponding to the nucleic acid fragment in the reference genome. An output file in SAM (Sequence Alignment / Map) format or BAM (binary) format can be generated and output for further analysis, such as methylation status determination.

[0141] From these sequence reads, the analysis system determines the position and methylation status of each CpG site based on alignment with the reference genome. The analysis system generates a methylation status vector for each fragment, which specifies the position of the fragment in the reference genome (e.g., as specified according to the position of the first CpG site in each fragment or another similar metric), the number of CpG sites in the fragment, and the methylation status of each CpG site in the fragment (methylated (e.g., represented as M), unmethylated (e.g., represented as U), or indeterminate (e.g., represented as I)). The observed statuses can be the methylated and unmethylated states; while the unobserved status is indeterminate. The indeterminate methylation status may result from sequencing errors and / or differences in methylation status between complementary strands of the DNA fragment. The methylation status vector may be stored in transient or permanent computer memory for subsequent use and processing. Further, the analysis system may remove duplicate reads or duplicate methylation status vectors from a single sample. The analysis system may identify a particular fragment with one or more CpG sites that has a number or percentage of indeterminate methylation statuses higher than a threshold, and may then exclude such fragments, or selectively include such fragments but construct a model to account for such indeterminate methylation statuses.

[0142] Figure 3B is according to one or more embodiments of Figure 3AExemplary illustration of process 300 for sequencing cfDNA molecules to obtain a methylation status vector. For example, the analysis system receives a cfDNA molecule 312 which, in this example, contains three CpG sites. As shown, the first and third CpG sites of the cfDNA molecule 312 are methylated 314. During processing step 320, the cfDNA molecule 312 is converted to generate a converted cfDNA molecule 322. During processing 320, the unmethylated second CpG site has its cytosine converted to uracil. However, the first and third CpG sites are not converted.

[0143] After conversion, a sequencing library 330 is prepared and sequencing 340 is performed to generate sequence reads 342. The analysis system aligns the sequence reads 342 with a reference genome 344. The reference genome 344 provides context as to where in the human genome the cfDNA fragment originated. In this simplified example, the analysis system performs an alignment 350 of the sequence reads 342 such that the three CpG sites correspond to CpG sites 33, 24, and 25 (these are arbitrary reference identifiers used for ease of description). Thereby, the analysis system can generate information regarding the methylation status of all CpG sites on the cfDNA molecule 312 and the locations in the human genome to which these CpG sites map. As shown, the methylated CpG sites on the sequence reads 342 are read as cytosine. In this example, cytosine appears only at the first and third CpG sites of the sequence reads 342, from which it can be inferred that the first and third CpG sites in the original cfDNA molecule are methylated. The second CpG site can be read as thymine (U is converted to T during the sequencing process), from which it can be inferred that the second CpG site in the original cfDNA molecule is unmethylated. With these two pieces of information, methylation status and location, the analysis system generates 360 a methylation status vector 352 for the cfDNA fragment 312. In this example, the resulting methylation status vector 352 is <M 23 , U 24 , M 25 >, where M corresponds to methylated CpG sites, U corresponds to unmethylated CpG sites, and the subscript numbers correspond to the location of each CpG site in the reference genome.

[0144] One or more alternative sequencing methods can be used to obtain sequence reads from the nucleic acids of a biological sample. The one or more sequencing methods can include any form of sequencing that can be used to obtain a certain number of sequence reads measured from nucleic acids (e.g., cell-free nucleic acids), including but not limited to high-throughput sequencing systems such as the Roche 454 platform, the SOLID platform from Applied Biosystems, Helicos True Single Molecule DNA sequencing technology, the sequencing-by-hybridization platform from Affymetrix, the Single Molecule Real-Time (SMRT) technology from Pacific Biosciences, the 454 Life Sciences, the synthesis sequencing platforms from Illumina / Solexa and Helicos Biosciences, the ligation sequencing platform from Applied Biosystems. The ION TORRENT technology from Life technologies and nanopore sequencing can also be used to obtain sequence reads from nucleic acids (e.g., cell-free nucleic acids) in a biological sample. Synthesis sequencing and sequencing based on reversible terminators (e.g., Illumina Genome Analyzer; Genome Analyzer II; HISEQ 3000; HISEQ 4500 (Illumina, San Diego Calif.)) can be used to obtain sequence reads from cell-free nucleic acids of a biological sample obtained from a trained subject to form a genotype dataset. Millions of cell-free nucleic acid (e.g., DNA) fragments can be sequenced in parallel. In one example of such sequencing technology, a flow cell is used, which includes an optically transparent slide with eight independent lanes on the surface, to which oligonucleotide anchors (e.g., adapter primers) are bound. The cell-free nucleic acid sample can include a signal or tag for easy detection. Obtaining sequence reads from cell-free nucleic acids of a biological sample can include obtaining quantitative information on the signal or tag through a variety of techniques, such as flow cytometry, quantitative polymerase chain reaction (qPCR), gel electrophoresis, gene chip analysis, microarray, mass spectrometry, cell fluorescence analysis, fluorescence microscopy, confocal laser scanning microscopy, laser scanning cytometry, affinity chromatography, manual batch separation mode, electric field suspension, sequencing, and combinations thereof.

[0145] One or more sequencing methods may include whole genome sequencing analysis. Whole genome sequencing analysis may include a physical analysis that generates sequence reads of the whole genome or a majority of the whole genome, which may be used to determine large-scale variations such as copy number variations or copy number aberrations. Such physical analysis may employ whole genome sequencing technology or whole exome sequencing technology. Whole genome sequencing analysis may have an average sequencing depth of at least 1x, 2x, 3x, 4x, 5x, 6x, 7x, 8x, 9x, 10x, at least 20x, at least 30x, or at least 40x on the genome of a tested subject. In some embodiments, the sequencing depth is about 30,000x. One or more sequencing methods may include targeted panel sequencing analysis. For the targeted genome, the targeted panel sequencing analysis may have an average sequencing depth of at least 50,000x, at least 55,000x, at least 60,000x, or at least 70,000x. The targeted genome may include 450 to 500 genes. The targeted genome may include a range of 500 ± 5 genes, a range of 500 ± 10 genes, or a range of 500 ± 25 genes.

[0146] One or more sequencing methods may include paired-end sequencing. One or more sequencing methods may generate multiple sequence reads. These multiple sequence reads may have an average length range of 10 to 700, 50 to 400, or 100 to 300. One or more sequencing methods may include methylation sequencing analysis. Methylation sequencing may be i) whole genome methylation sequencing or ii) targeted DNA methylation sequencing using multiple nucleic acid probes. For example, methylation sequencing is whole genome bisulfite sequencing (e.g., WGBS). Methylation sequencing may be targeted DNA methylation sequencing using multiple nucleic acid probes that target the most informative regions of the methylome, which may be a unique methylation database, and may be previous whole genome and targeted sequencing profiling analyses.

[0147] Methylation sequencing may detect one or more 5-methylcytosine (5mC) and / or 5-hydroxymethylcytosine (5hmC) in the corresponding nucleic acid methylation fragment. Methylation sequencing may include converting one or more unmethylated cytosines, or one or more methylated cytosines, in the corresponding nucleic acid methylation fragment into the corresponding one or more uracils. During methylation sequencing, one or more uracils may be detected as one or more corresponding thymines. The conversion of one or more unmethylated cytosines or one or more methylated cytosines may include chemical conversion, enzymatic conversion, or a combination thereof.

[0148] For example, bisulfite conversion involves converting cytosine to uracil without converting methylated cytosine (e.g., 5-methylcytosine or 5-mC). In some DNA, approximately 95% of the cytosine in the DNA may be unmethylated, and the resulting DNA fragments may include many uracils represented as thymine. Enzymatic conversion processes can be used to process nucleic acids prior to sequencing and can be processed in a variety of ways. An example of bisulfite-free conversion involves a bisulfite-free sequencing method with base-level resolution, TET-assisted pyridine borane sequencing (TAPS), for non-destructive and direct detection of 5-methylcytosine and 5-hydroxymethylcytosine without affecting unmodified cytosine. For a CpG site among multiple corresponding CpG sites in a nucleic acid methylation fragment, if the methylation status is determined to be methylated by methylation sequencing, then the CpG site is in a methylated state; if it is determined to be unmethylated by methylation sequencing, then the CpG site is in an unmethylated state.

[0149] Methylation sequencing analysis (e.g., WGBS and / or targeted methylation sequencing) can have an average sequencing depth including but not limited to about 1,000x, 2,000x, 3,000x, 5,000x, 10,000x, 15,000x, 20,000x, or 30,000x. Methylation sequencing can have a sequencing depth greater than 30,000x, e.g., at least 40,000x or 50,000x. Whole-genome bisulfite sequencing methods can have an average sequencing depth of 20x to 50x, while targeted methylation sequencing methods have an average effective depth of 100x to 1000x, where the effective depth can be the equivalent coverage required for whole-genome bisulfite sequencing to obtain the same number of sequence reads as in targeted methylation sequencing.

[0150] For further details regarding methylation sequencing (e.g., WGBS and / or targeted methylation sequencing), see, for example, U.S. Patent Application No. 16 / 352,602, filed on March 13, 2019, titled "Methylation Fragment Anomaly Detection", U.S. Patent Application No. 16 / 719,902, filed on December 18, 2019, titled "Systems and Methods for Estimating Cell Source Fractions Using Methylation Information", and U.S. Patent Application No. 17 / 191,914, filed on March 4, 2021, titled "Systems and Methods for Cancer Condition Determination Using Autoencoders". Other methods for methylation sequencing, including those disclosed herein and / or any modifications, alternatives, or combinations thereof, can be used to obtain the methylation patterns of the fragments. Methylation sequencing can be used to identify one or more methylation state vectors, e.g., as described in U.S. Patent Application No. 16 / 352,602, filed on March 13, 2019, titled "Anomalous Fragment Detection and Classification", or according to any of the techniques disclosed in U.S. Patent Application No. 15 / 931,022, filed on May 13, 2020, titled "Model-Based Featurization and Classification". Each reference is incorporated by reference in its entirety.

[0151] Methylation sequencing of nucleic acids and one or more methylation status vectors obtained therefrom can be used to obtain multiple nucleic acid methylation fragments. Each set corresponding to multiple nucleic acid methylation fragments (e.g., for each corresponding genotype dataset) can contain more than 100 nucleic acid methylation fragments. In each set corresponding to multiple nucleic acid methylation fragments, the average number of nucleic acid methylation fragments can contain 1,000 or more nucleic acid methylation fragments, 5,000 or more nucleic acid methylation fragments, 10,000 or more nucleic acid methylation fragments, 20,000 or more nucleic acid methylation fragments, or 30,000 or more nucleic acid methylation fragments. In each set corresponding to multiple nucleic acid methylation fragments, the average number of nucleic acid methylation fragments can be 10,000 nucleic acid methylation fragments - 50,000 nucleic acid methylation fragments. The multiple nucleic acid methylation fragments can contain 1,000 or more, 10,000 or more, 100,000 or more, 1 million or more, 10 million or more, 100 million or more, 500 million or more, 1 billion or more, 2 billion or more, 3 billion or more, 4 billion or more, 5 billion or more, 6 billion or more, 7 billion or more, 8 billion or more, 9 billion or more, or 10 billion or more nucleic acid methylation fragments. The average length of the multiple nucleic acid methylation fragments can be 140 to 480 nucleotides. III. Deconvolution of Component Ratios

[0152] Deconvolution of component ratios involves determining the breakdown of the component ratios from which the nucleic acid fragments in a sample are derived. Component ratios represent the breakdown of the components that contribute nucleic acid fragments to a sample. In one or more embodiments, the component ratios specify the percentage of the sample attributable to each component. For example, the deconvolution process determines a first percentage of the nucleic acid fragments in the sample that are derived from a first component, a second percentage of the nucleic acid fragments in the sample that are derived from a second component, and so on for the remaining components. In other embodiments, the component ratios can specify the ranking of the components according to predicted ratios. For example, component 3 is the largest contributor of the nucleic acid fragments in the sample, followed by component 1, and then component 2 (in an example where a mixture model predicts the ratios of three components).

[0153] The deconvolution process models the methylation characteristics of the various components. Methylation characteristics represent the methylated sequence reads of a component. In some embodiments, the methylation characteristics contain the counts of methylation variations over multiple genomic regions. In other embodiments, the methylation characteristics can further contain the counts of reference states over multiple genomic regions. The reference state at a genomic region can include the remaining methylation patterns not assigned to a variant state. As for components, each component can be a tissue type, a cell type, or another subtype thereof.

[0154] The deconvolution process can be applied at various stages of the overall cancer classification workflow 100 described in Figure 1 For example, in one exemplary implementation, at step 142 of workflow 100, the deconvolution process can be used to evaluate sample purity. In another exemplary implementation, the deconvolution process can be used in cancer classification 146 to determine a specific cancer type. As yet another exemplary implementation, the deconvolution process can be used to quantify cancer signals, e.g., for monitoring cancer status and / or progression. Other exemplary implementations can utilize the component ratios in other ways. III.A. Mixture Model

[0155] The analysis system can train a mixture model for deconvolving component ratios. The mixture model generally predicts component ratios based on the methylation sequence reads of a sample. The mixture model typically includes multiple component sub-models. The component sub-models model the methylation characteristics of the components. The number of components to be deconvolved in the training sample set can be adjusted as a hyperparameter of the mixture model. During training, the component sub-models can take as input the methylation characteristics of the sample and output component likelihoods, which indicate the likelihood that the methylation characteristics of the sample are from that component.

[0156] Figure 4A and Figure 4B show an overview of using the mixture model. The analysis system (or its various components) can perform the methods of Figure 4A and Figure 4B In other embodiments, any steps of the methods can be performed by other general-purpose computing devices. In other embodiments, the methods can include more steps, fewer steps, different steps, or some combination of these steps.

[0157] Figure 4A is an exemplary flowchart that depicts a method 400 for training a mixture model for predicting component ratios in a sample according to one or more embodiments.

[0158] The analysis system obtains 410 or more training samples. Each training sample can include methylation sequence reads, e.g., as by sample sequencing 120 in Figure 1 or more specifically Figure 3AMethylation sequence reads generated by sample processing 300. Each methylation sequence read may include methylation information of a nucleic acid fragment of the sample. The sample may be a liquid biopsy sample, a tissue-derived sample, an isolated cell sample, another biological sample, or some combination thereof. The sample may include circulating tumor cells, disseminated tumor cells, other types of somatic cells, etc. The training sample may have a known or unknown component ratio. For example, the first sample may be a tissue-derived sample that includes nucleic acid fragments derived from isolated tissue. The first sample will have a component ratio that is mainly the isolated tissue. Another sample may be a purified plasma sample from a healthy subject that mainly contains free nucleic acid fragments. The sample may have a component ratio that is mainly non-cancerous free nucleic acid fragments. In some example embodiments, the analysis system obtains 1,000 training samples, 2,000 training samples, 3,000 training samples, 4,000 training samples, 5,000 training samples, 6,000 training samples, 7,000 training samples, 8,000 training samples, 9,000 training samples, 10,000 training samples, 20,000 training samples, 30,000 training samples, 40,000 training samples, 50,000 training samples, or more than 50,000 training samples.

[0159] The analysis system generates 420 methylation features for each training sample, for example, by counting methylation variations across multiple genomic regions. The analysis system may partition the genome into individual genomic regions. Each genomic region may span multiple methylation sites, e.g., CpG sites. The analysis system may span the genome, i.e., for human training samples, use whole-genome sequencing of the entire human genome. In other embodiments, the methylation sequence reads are generated from targeted methylation assays of a subset of genomic regions in the targeted genome. The analysis system may also screen out genomic regions with an average sequencing depth below a threshold depth across the entire training sample. The methylation features may be based on other types of methylation features (e.g., methylation density at each methylation site among multiple methylation sites, methylation density across a genomic region, count of highly methylated fragments overlapping a genomic region, count of highly unmethylated fragments overlapping a genomic region, etc.).

[0160] At each genomic region, an analysis system can count methylated sequence reads, and each methylated sequence read can be characterized as one of multiple methylation variants. For example, a set of CpGs can be associated with multiple methylation variants, and each methylation variant indicates a different cancer. Each methylation variant can have a defined variant state. In the case of having two variant states, one variant state can be considered the major variant state, and the other variant state can be considered the minor variant state. The analysis system can count a first count of methylated sequence reads (or nucleic acid fragments) having the major variant state and a second count of methylated sequence reads (or nucleic acid fragments) having the minor variant state. In the case of having more than two variant states, there can be a major variant state, a minor variant state, and multiple other variant states. The analysis system can count the methylated sequence reads having each variant state, i.e., the first count for the major variant state and the count for each other variant state. In other embodiments, the analysis system can count the first count for the major variant state and a second count for one or more other variant states together. In some embodiments, the analysis system can count only the alternative variant states. Figure 5A and Figure 5B shows the counting of variant states.

[0161] In one or more embodiments, the analysis system can perform genomic region selection. The analysis system can exclude genomic regions according to one or more selection criteria. An example selection criterion sets a limit on the number of methylation variants of the genomic regions to be included. For example, if more than two methylation variants occur at a genomic region within a population at a significant proportion, the analysis system can exclude such a genomic region. The significant proportion (which can also be referred to as the prevalence) can be 1%, 2%, 3%, 4%, 5%, 6%, 7%, 8%, 9%, 10%, 11%, 12%, 13%, 14%, 15%, 20%, 25%, or 30%. Another selection criterion can rank and exclude genomic regions according to the ability of the genomic region to distinguish between components. For example, when training a mixture model, the analysis system can generate an information gain score for each genomic region based on the informativeness of the genomic region in deconvolution. Genomic regions with scores below a threshold can be excluded.

[0162] Figure 5A shows the counting of methylation variants from a single CpG locus 505 as a genomic region. In a single CpG locus genomic region, there are typically two methylation states: a reference state and a variant state (related to the methylation variant). The analysis system can count a first count of methylated sequence reads having the reference state and a second count of methylated sequence reads having the variant state.

[0163] For example, in Figure 5A , there are six segments 510 overlapping with a single CpG site 505. At each CpG site (represented as a diamond) on the segment, the segment has a methylation state. The methylation state can include methylated (shown as solid), unmethylated (shown as hollow), and unknown (shown as diagonal hatching). Unknown methylation may include an uncertain state caused by mutation or sequencing error. As a single site, there are two possible methylation states (M or U) in this genomic region. One methylation state is considered the reference state, and the other methylation state is considered the variant state. For example, at the CpG site 505 as a genomic region, the reference genome specifies that this CpG site is mainly methylated, so the reference state is methylated. As for counting, the analysis system can count four segments (or methylated sequence reads) as having the methylated reference state, and two segments (or methylated sequence reads) as having the unmethylated variant state.

[0164] Figure 5B shows counting of methylation variants for a genomic region 515 spanning multiple CpG sites 517. These CpG sites 517 include CpG site 1, 2, and 3. As in Figure 5A , the solid diamond represents methylation, the hollow diamond represents unmethylation, and the diagonal hatched diamond represents unknown. In the illustrated embodiment, the methylation pattern of segment 1520A can include methylation at CpG site 1, unmethylation at CpG site 2, and methylation at CpG site 3. For example, this can be the reference state. The analysis system can count the number of segments having the methylated, unmethylated, and methylated reference states at the genomic region 515. In this example, four segments have the reference state. Segment 2 520B has a methylation variant different from the reference state. All three CpG sites 517 in the methylation pattern of segment 2 520B are methylated. The analysis system can count one total segment (or methylated sequence read) as having this first methylation variant. Segment 4 520D has another methylation variant different from the reference state and the first methylation variant. All three CpG sites 517 in the methylation pattern of segment 4 520D are unmethylated. The analysis system can count one total segment (or methylated sequence read) as having this second methylation variant. The analysis system can also sum the counts of all methylation variants. In such an example, the count of all methylation variants is two, including segment 2 520A having the first methylation variant and segment 4 520D having the second methylation variant.

[0165] Return Figure 4A, the analysis system generates 420 methylation features based on the counts of methylation variations. In one embodiment, a methylation feature is the count of (multiple) methylation variations for each genomic region among a plurality of genomic regions. In another embodiment, a methylation feature is the percentage of (multiple) methylation variations for each genomic region among a plurality of genomic regions. To calculate the percentage of (multiple) methylation variations for a genomic region, the analysis system counts the total count of methylation sequence reads present at the genomic region and the count of methylation sequence reads having (multiple) alternative methylation variations. The percentage is equal to the count of methylation sequence reads having (multiple) methylation variations divided by the total count of methylation sequence reads present at the genomic region. For example, there are 10,000 genomic regions, and the methylation feature is a vector of length 10,000. The first value of the vector corresponds to the count or percentage of (multiple) methylation variations at the first genomic region, the second value of the vector corresponds to the count or percentage of (multiple) methylation variations at the second genomic region, and so on up to 10,000 genomic regions.

[0166] In some embodiments, the methylation feature may further include the count or percentage of the reference state for each genomic region. In an example with 10,000 genomic regions, the methylation feature will contain 20,000 values, including 10,000 counts or percentages of the reference state across the genomic regions and 10,000 counts or percentages of methylation variations.

[0167] In some embodiments, the methylation feature may further include one or more counts or percentages for each methylation variation at a genomic region. For example, in a population at a genomic region, two or more alternative methylation variations are expressed at a statistically significant proportion of the population (e.g., higher than 5%, 10%, 15%, etc.). The methylation feature may also include the count or percentage for each methylation variation at the genomic region, e.g., a first count for a first methylation variation, a second count for a second methylation variation with a methylation pattern different from the first methylation variation, and so on.

[0168] In one or more embodiments, the methylation feature may be normalized. The normalization may be based on the sequencing depth of the sample. Normalization based on the sequencing depth normalizes the variance of the genetic material across samples.

[0169] The analysis system trains a 430 mixture model using the methylation features of the training samples. The analysis system may train the mixture model to predict the component proportions (also referred to as deconvolution of the component proportions) based on the input methylation features. The analysis system may perform maximum likelihood estimation to train the model to fit the methylation features of the training samples.

[0170] An analysis system can train a mixture model as a machine learning model. Training of the machine learning model can include supervised training, unsupervised training, or semi-supervised training. Supervised training typically requires the use of known labels or values, which can be used to calculate the error or loss of the predictions of the model. The analysis system training the model can adjust the parameters of the model to minimize the error. Using the mixture model, the analysis system can utilize training samples with known component ratios to provide a supervised loss when training the mixture model. Unsupervised training typically requires learning patterns in the training data without known labels or values to supervise the learning. Example unsupervised techniques include clustering, anomaly detection, latent variable learning, other types of machine learning techniques that do not rely on known labels or values, and the like. In the context of the mixture model, the analysis system can utilize training samples without known component ratios and allow the mixture model to learn patterns within the training data. Semi-supervised learning typically may require the use of training data where some of the training data has known labels or values. Types of machine learning models that can be implemented include decision trees, neural networks, multi-layer perceptrons, support vector machines, models that rely on other types of machine learning, their derivatives, and combinations thereof. In Figure 6 and Figure 7 embodiments of the mixture model architecture are described in

[0171] Figure 6 An exemplary methylation feature matrix 610 for training a mixture model according to an example embodiment is shown. The methylation feature matrix 610 represents the methylation features of the training samples for training the mixture model. The methylation feature matrix 610 contains multiple samples 620 over multiple genomic regions 630. The methylation features of each training sample 620 are represented by a row of the methylation feature matrix 610.

[0172] Figure 7 An exemplary architecture of a mixture model 700 according to one or more embodiments is shown. The mixture model 700 includes multiple group sub-models and a deconvolution model 740. Each group sub-model is configured to input the methylation features of a sample and output the predicted likelihood that the methylation features are from a component. The predicted likelihood can be expressed as a percentage, e.g., the likelihood that the methylation features are from component 1 is 65%. For example, the component 1 sub-model 710 outputs a component 1 likelihood 715, the component 2 sub-model 720 outputs a component 2 likelihood 725, and so on for the group sub-models until the component K sub-model 730 outputs a component K likelihood 735. The component likelihoods are input into the deconvolution, and the deconvolution outputs the component ratios of the methylation features.

[0173] The analysis system can train the component submodels and the deconvolution submodel 740 in parallel. Parallel training requires inputting the methylation feature matrix 610 into all the submodels simultaneously, where the analysis system adjusts the parameters of each submodel in parallel. In the supervised learning training method, the analysis system can feed one or more training samples into the mixture model 700 to predict the component ratios of the training samples. The analysis system can calculate the loss of the training samples by comparing the known component ratios of the samples with the predicted component ratios of the samples. In the unsupervised learning method for training, the analysis system can feed the training samples into the mixture model 700 to cluster similar methylation features mainly from the same component. In such an unsupervised method, the analysis system can be independent of the true component ratios of the training samples.

[0174] In other embodiments, the analysis system can train one or more of the submodels individually. Individual training requires using a subset of the training data to train the submodel. Once trained, the analysis system can fix the parameters of that submodel while training other submodels. For example, the analysis system can train each of the component submodels individually and then keep the parameters of these component submodels while training the deconvolution submodel 740. The analysis system can use training samples mainly from one component to train the corresponding component submodel of that component. Then, the analysis system can use training samples with mixing ratios to train the deconvolution model 740.

[0175] In one or more embodiments, the mixture model 700 can include further dividing the component submodels into further sub-component submodels. For example, the component 1 submodel 710 can be connected to one or more sub-component submodels (not shown). The sub-component submodels can further determine the sub-component likelihoods that attribute methylation features to the sub-components. For example, component 1 can represent breast tissue shed from potential breast tumor cells. As a further division of the breast tissue, component 1 can include the human epidermal growth factor receptor 2 (HER2) positive and negative states. The component 1 submodel 710 can provide a component 1 likelihood 715 of 65% breast tissue. Based on the methylation features of the sample, the first sub-component submodel can output that if the methylation feature is from breast tissue (component 1), then the methylation feature has a 35% likelihood of being in the HER2 positive state. The deconvolution submodel 740 inputs the component likelihoods and any sub-component likelihoods to determine the component and sub-component proportions. Continuing with the above example, the deconvolution submodel 740 of the mixture model 700 can output the proportion of breast tissue as component 1 and the proportion of the HER2 positive state. In some embodiments with multiple sub-components, the deconvolution submodel 740 can provide the sub-component proportions for each sub-sub-component, where the sum of the sub-component proportions is the component proportion. The mixture model 700 can include a tree of components and sub-components that spans multiple layers in depth. For example, a three-layer depth can include: a first component includes a set of sub-components, and one of these sub-components includes a second set of sub-components. The analysis system can train the sub-component submodels individually, or can train the sub-component submodels in parallel with other submodels of the mixture model 700.

[0176] Return to Figure 4A , when training the mixture model 430, the analysis system can adjust the number K of components as a hyperparameter 435. K represents the number of components for which the mixture model predicts the component proportions (e.g., as Figure 7 shown). The analysis system can iteratively train the mixture model while adjusting the number K of components and cross-validating the trained mixture model. Cross-validation can include using a validation set of training samples with known component proportions. The analysis system can input the methylation features of the training samples to predict the component proportions. The analysis system can determine the error or loss as the difference between the known component proportions and the predicted component proportions. The analysis system can score each mixture model trained with a different number K of components based on the error or loss from the validation set. In one or more embodiments, the analysis system can apply a penalty as a factor of the number of components in order to optimize the number of components in the mixture model. If there are too few components, the model may fit the training data poorly. If there are too many components, the model may overfit the training data. The analysis system can also score how well the mixture model fits the training data.

[0177] In one or more embodiments, the analysis system can inform the mixture model what each component corresponds to. In such embodiments, the analysis system can train the mixture model without knowing the labels of the training data. The analysis system can later utilize the known labels or proportions of the training data to inform what each component corresponds to. The analysis system can utilize training samples with known component proportions, such as the following training sample: containing 65% nucleic acid fragments derived from breast tissue and 35% nucleic acid fragments corresponding to non-cancerous cell-free nucleic acid fragments. The analysis system can input the training sample into the mixture model, which outputs a predicted component proportion of 70% for component 1, a predicted component proportion of 25% for component 2, and a predicted component proportion of 5% for component 3 (in an example mixture model trained on at least three components). The analysis system can match the largest known component proportion and the largest predicted component proportion to correspond to the same component. The analysis system can similarly process other known component proportions and other predicted component proportions in the training sample. The analysis system can utilize multiple labeled training samples to confirm the labeling of the components in the mixture model. For example, most of the training samples indicate that component 1 of the mixture model matches lung tissue, and so on for the remaining components of the mixture model.

[0178] Figure 4B is an exemplary flowchart that depicts a method 440 for deploying a mixture model for predicting component proportions in a sample. The mixture model can be trained according to a method 400 for training a mixture model.

[0179] The analysis system obtains 450 a test sample for predicting component proportions. The test sample contains genetic material that includes methylation information. The analysis system may or may not know the disease or cancer status of the subject providing the test sample. The methylation information can include methylation sequence reads, e.g., methylation sequence reads sequenced by the processes described in Figure 3A and Figure 3B .

[0180] The analysis system generates 460 methylation features for the test sample by counting methylation variations over a plurality of genomic regions. The plurality of genomic regions match the plurality of genomic regions on which the mixture model was trained. Similar to the training samples, the analysis system can normalize the methylation features of the test sample, e.g., based on sequencing depth.

[0181] The analysis system applies 470 the mixture model to the methylation features to predict the component proportions of the test sample. The analysis system inputs the methylation features of the test sample into the mixture model, which outputs the component proportions. III.B. Application of the Mixture Model

[0182] Based on the training method of the hybrid model, the analysis system can adopt the hybrid model in various applications.

[0183] In one or more embodiments, the hybrid model can be used for tissue purity assessment (e.g., in Figure 1 step 142). The analysis system can utilize the hybrid model to evaluate the purity of a sample before using the sample (e.g., for training a cancer classifier, etc.). For example, the analysis system can obtain a plurality of training samples diagnosed with a specific cancer type. The analysis system can use the hybrid model to predict the component ratio of the training samples. In one embodiment, the analysis system can exclude training samples with a tissue related to the specific cancer type below a threshold ratio. For example, the training sample is derived from a subject diagnosed with lung cancer, but the hybrid model predicts that the training sample has 10% nucleic acid fragments derived from lung tissue. If the analysis system retains training samples with nucleic acid fragments of the corresponding tissue from the diagnosed cancer type higher than 30%, the analysis system can not use the training samples below the threshold when training the cancer classifier. This can advantageously focus the training of the cancer classifier on training samples with a large amount of cancer signals, thereby improving the sensitivity of the cancer classifier. Utilizing the hybrid model for tissue purity can also identify training samples with misdiagnosis. For example, as determined by the hybrid model, a training sample derived from a subject diagnosed with prostate cancer may have a large amount of liver tissue in a liquid biopsy sample. The analysis system can not use such training samples in the training of the cancer classifier to prevent the cancer classifier from being skewed.

[0184] In one or more embodiments, a mixture model can be used to predict the cancer type of a test subject. An analysis system can utilize the mixture model after or as part of a cancer classifier. The analysis system can use the cancer classifier to generate a cancer prediction. The cancer prediction can include a binary state (between cancer and non-cancer), a cancer type (between multiple cancer types), or some combination thereof. In one embodiment, in response to predicting the likelihood of cancer presence (binary state), the analysis system can use the mixture model to determine the component proportions as a way to determine the cancer type. The largest component proportion that is not non-cancerous tissue or other healthy tissue can be determined as the cancer type of the test sample. In other embodiments, the analysis system can use the component proportions of the mixture model to confirm the predicted cancer type of the cancer classifier. For example, the cancer classifier predicts that the test sample may have head and neck cancer. The analysis system can use the mixture model to identify whether head and neck tissue is present in the test sample in a substantial proportion. If not, the analysis system can record the inconsistency, e.g., reduce the confidence in the prediction. The analysis system can also use the mixture model to determine whether multiple cancer types may be present. For example, if the mixture model provides a substantial proportion of two different components (which are not non-cancerous or other healthy tissue), the analysis system can determine that the test sample may have two cancer types.

[0185] In one or more embodiments, a mixture model can be used to quantify the cancer signal in a test subject. The analysis system can use the mixture model to predict the component proportions of a test subject diagnosed with cancer or predicted to have cancer. The analysis system can quantify the cancer signal based on the component proportions. In some embodiments, the analysis system can use the component likelihoods output by the component sub-models to quantify the cancer signal. Quantification of the cancer signal can provide insights into which stage the cancer may be in or the aggressiveness of the cancer. For example, the analysis system can analyze test samples from an individual at different time points to, e.g., evaluate changes in the cancer signal based on the component proportions predicted by the mixture model. If the analysis system identifies an increase in the cancer signal, the analysis system can determine that the cancer is progressing. The analysis system can further determine the aggressiveness of the cancer based on the rate of increase of the cancer signal. The analysis system can also use the quantification of the cancer signal to evaluate various treatment plans. If there is a slowdown in growth or a reduction in quantification, the analysis system can determine that the treatment plan being implemented is successful. Otherwise, the analysis system can determine that the treatment plan is not successful and can further recommend adjustments to the treatment plan. The recommendations for adjusting the treatment plan can include changing the therapy, changing the dose or regimen, increasing the duration or frequency of the treatment plan, etc.

[0186] In one or more embodiments, an analysis system may utilize component sub-models in addition to a mixture model. Generally, a component sub-model represents the distribution of methylation variant allele fractions of genetic material from a component. Thereby, the analysis system may utilize the component sub-model to predict the likelihood that a methylation signature originated from a component. The analysis system may also generate a distribution of methylation signatures for a component. Using the component sub-model, the analysis system may predict what component a methylation sequence read originated from. The analysis system may utilize such predictions to screen cancer training samples to remove non-cancer impurity sequence reads to retain methylation sequence reads more likely to be related to cancer signals when training a cancer classifier. Non-cancer impurities may include one or more of the following: lymphocytes, macrophages, fibroblasts, vascular endothelial cells, or non-cancerous tissue. In one or more embodiments, the analysis system may obtain methylation signatures of non-cancer impurities from a reference database. The reference database may include the distribution of methylation signatures observed from non-cancerous tissue or cells (e.g., from healthy subjects not diagnosed with cancer (or any other genetic-based disease)). In some embodiments, the reference database may be populated with methylation signatures observed during a previous assessment of genetic material. In some embodiments, the reference database may be populated with values obtained from a third party (such as from a public collection of methylation signatures).

[0187] In one or more embodiments, the analysis system may extract learned information from the mixture model. In one embodiment, the analysis system may learn the correlations between methylation variations at genomic regions contributing to the component proportions. For example, the analysis system may identify methylation variations that provide information about a particular component. In another embodiment, the analysis system may identify methylation variations at genomic regions associated with non-cancer impurities. The analysis system may screen to remove the use of such methylation variations at these genomic regions as these genomic regions may be confounded with non-cancer impurities. IV. Cancer Classifier for Determining Cancer

[0188] Cancer classification involves extracting genetic features and applying one or more models to the extracted features to determine cancer predictions. An analysis system aggregates the extracted features into a feature vector, which can then be input into a trained cancer classifier to determine a cancer prediction based on the input feature vector. The cancer prediction can include a label and / or a value. The label can be binary, indicating the presence or absence of cancer in a test subject, and / or multi-class, indicating one or more specific cancer types out of multiple cancer types being screened. The value can be a grading of the cancer status (e.g., the stage of the cancer) or a quantification of the cancer signal. In particular, the cancer classifier can be a machine learning model that includes multiple classification parameters and a function that represents the relationship between the input feature vector and the cancer prediction as the output. Inputting the feature vector together with the classification parameters into the function produces the cancer prediction. In one or more embodiments, cancer classification utilizes multiple models. Some models can be used to determine one or more features from methylation sequence reads for cancer classification. Before deploying the cancer classifier, the analysis system trains the cancer classifier. IV.A. Identifying Abnormal Fragments

[0189] The analysis system can determine the abnormal fragments of a sample using the methylation status vector of the sample. For each fragment in the sample, the analysis system can use the methylation status vector corresponding to the fragment to determine whether the fragment is an abnormal fragment. In some embodiments, the analysis system calculates a p-value score for each methylation status vector, which describes the probability of observing that methylation status vector or other less likely methylation status vectors in a healthy control group. The analysis system can determine the fragments with a p-value score of the methylation status vector lower than a threshold as abnormal fragments. In some embodiments, the analysis system further labels the fragments with at least a certain number of CpG sites and with a methylation or unmethylated percentage exceeding a certain threshold as hypermethylated fragments and hypomethylated fragments, respectively. Hypermethylated fragments or hypomethylated fragments can also be referred to as ultra-methylation abnormal fragments (UFXM). In other embodiments, the analysis system can employ a variety of other probability models to determine abnormal fragments. Examples of other probability models include mixture models, deep probability models, etc. In some embodiments, the analysis system can use any combination of multiple processes described below to identify abnormal fragments. Through the identified abnormal fragments, the analysis system can screen the set of methylation status vectors of the sample for use in other processes, such as for training and deploying the cancer classifier. IV.A.i. p-Value Screening Method

[0190] In some embodiments, the analysis system calculates a p-value score for each methylation status vector compared to the methylation status vectors of the fragments in a healthy control group. The p-value score can describe the probability of observing a methylation status in the healthy control group that matches this methylation status vector or other, even less likely, methylation status vectors. To determine that a DNA fragment is abnormally methylated, the analysis system can use a healthy control group in which most of the fragments are normally methylated. When performing this probability analysis to identify abnormal fragments, the determination is significant when compared to the population of control subjects that make up the healthy control group. To ensure the robustness of the healthy control group, the analysis system can select a threshold number of healthy individuals to obtain samples that include DNA fragments. The following Figure 8A describes a method for generating a data structure of a healthy control group that the analysis system can use to calculate the p-value score. Figure 8B describes a method for calculating the p-value score using the generated data structure.

[0191] According to one embodiment, Figure 8A is a flowchart depicting process 800 for generating a data structure of a healthy control group. To create the data structure of the healthy control group, the analysis system can receive multiple DNA fragments (e.g., cfDNA) from multiple healthy individuals. The analysis system can generate 805 a methylation status vector for each fragment, e.g., via process 300.

[0192] For each methylation status vector of a fragment, the analysis system can subdivide 810 the methylation status vector into a string of CpG sites. In some embodiments, the analysis system subdivides 810 the methylation status vector such that the resulting strings are all less than a given length. For example, a methylation status vector of length 11 might 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 status vector of length 7 is subdivided into strings of length less than or equal to 4, yielding 4 strings of length 4, 5 strings of length 3, 6 strings of length 2, and 7 strings of length 1. If a methylation status vector is shorter than or the same as the specified string length, then the methylation status vector might be converted into a single string that contains all of the CpG sites of the vector.

[0193] The analysis system counts these strings by calculating, for each possible CpG site and methylation state likelihood in the vector, the number of strings in the control group that have that specified CpG site as the first CpG site in the string and that methylation state likelihood. For example, at a given CpG site and considering a string length of 3, there are 2^3 or 8 possible string configurations. At that given CpG site, for each of these 8 possible string configurations, the analysis system counts the number of occurrences of each methylation state vector likelihood in the control group. Continuing this example, this may involve counting the following numbers: 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 >,..., <U x , U x+1 , U x+2 >. The analysis system creates a data structure that stores the statistical counts for each starting CpG site and string likelihood.

[0194] Setting an upper limit on the string length has several benefits. First, depending on the maximum length of the string, the size of the data structure created by the analysis system can increase dramatically. For example, a maximum string length of 4 means that for each CpG site, at least 2^4 numbers need to be counted for strings of length 4. Increasing the maximum string length to 5 means that there are an additional 2^4 or 16 numbers to count for each CpG site, doubling the numbers to be counted (and the computer memory required) compared to the previous string length. Reducing the string size can help keep the creation and performance of the data structure (e.g., for subsequent access, as described below) reasonable in terms of computation and storage. Second, a statistical consideration for limiting the maximum string length can be to avoid overfitting downstream models that use string counts. If long strings of CpG sites do not strongly influence the outcome biologically (e.g., the prediction of anomalies that are predictive of the presence of cancer), then calculating probabilities based on large strings of CpG sites can be problematic because it uses a large amount of data that may not be available and may be too sparse for the model to perform properly. For example, calculating the probability of an anomaly / cancer condition based on the first 100 CpG sites can use the counts of strings of length 100 in the data structure, ideally some of which exactly match these first 100 methylation states. If only sparse counts of strings of length 100 are available, there may not be enough data to determine whether a given string of length 100 in the test sample is anomalous.

[0195] According to one embodiment,Figure 8B is a flowchart depicting a process 830 for identifying aberrant methylation segments from an individual. In process 830, an analysis system generates 840 methylation status vectors from cfDNA segments of a subject, such as by process 300. The analysis system can process each methylation status vector as follows.

[0196] For a given methylation status vector, the analysis system enumerates 845 all possible methylation status vectors that have the same starting CpG site and the same length (i.e., set of CpG sites) as the given methylation status vector. Since each methylation status is typically either methylated or unmethylated, there are actually two possible states at each CpG site, so the number of different possibilities of a methylation status vector can depend on a power of 2, such that a methylation status vector of length n will be associated with 2 n methylation status vector possibilities. For a methylation status vector that contains one or more CpG sites with an indeterminate state, the analysis system may only consider those CpG sites with an observable state to enumerate 830 the possibilities of the methylation status vector.

[0197] The analysis system calculates the probability of observing each methylation state vector likelihood of 850 by accessing the healthy control group data structure for the identified starting CpG sites and methylation state vector lengths. In some embodiments, when calculating the probability of observing a given likelihood, Markov chain probabilities are used to model the joint probability calculation. The Markov model can be trained at least in part based on an assessment of the methylation state of each CpG site in the corresponding fragments (e.g., nucleic acid methylation fragments) of those nucleic acid methylation fragments in the healthy non-cancer cohort dataset that have corresponding multiple CpG sites. For example, a Markov model (e.g., a hidden Markov model or HMM) is used to determine the probability of observing a sequence of a certain methylation state (including, e.g., "M" or "U") for a nucleic acid methylation fragment among multiple nucleic acid methylation fragments, given a set of probabilities (for each state in the sequence, this set of probabilities determines the likelihood of observing the next state in the sequence). This set of probabilities can be obtained by training the HMM. Such training can involve calculating statistical parameters (e.g., the probability that the first state can transition to the second state (transition probability) and / or the probability of observing a given methylation state at the corresponding CpG site (emission probability)) given an initial training dataset of the observed methylation state sequence (e.g., methylation pattern). The HMM can be trained using supervised training (e.g., using samples where the underlying sequence and the observed states are both known) and / or unsupervised training (e.g., Viterbi learning, maximum likelihood estimation, expectation maximization training, and / or Baum-Welch training). In other embodiments, computational methods other than Markov chain probabilities are used to determine the probability of observing each methylation state vector likelihood. For example, such computational methods can include a learning representation. The p-value threshold can be between 0.01 and 0.10, or between 0.03 and 0.06. The p-value threshold can be 0.05. The p-value threshold can be less than 0.01, less than 0.001, or less than 0.0001.

[0198] The analysis system calculates a p-value score for the methylation state vector using the calculated probability of each likelihood. In some embodiments, this includes identifying the calculated probability corresponding to the likelihood that matches the methylation state vector under discussion. In particular, this can be the likelihood having the same set of CpG sites as the methylation state vector or similarly having the same starting CpG site and length as the methylation state vector. The analysis system can sum the calculated probabilities of any likelihood whose probability is less than or equal to the identified probability to generate a p-value score.

[0199] This p-value can represent the probability of observing the methylation status vector of the fragment or other more unlikely methylation status vectors in the healthy control group. Thus, a low p-value score can generally correspond to a methylation status vector that is rare in healthy individuals and causes the fragment to be labeled as abnormally methylated relative to the healthy control group. A high p-value score can generally be associated with a methylation status vector that is expected to be present in relatively healthy individuals. For example, if the healthy control group is the non-cancer group, a low p-value can indicate that the fragment is abnormally methylated relative to the non-cancer group and thus may indicate the presence of cancer in the test subject.

[0200] As described above, the analysis system can calculate a p-value score for each of the plurality of methylation status vectors, where each methylation status vector represents a cfDNA fragment in the test sample. To identify which fragments are abnormally methylated, the analysis system may screen the set of 865 methylation status vectors based on their p-value scores. In some embodiments, the screening is performed by comparing the p-value scores to a threshold and retaining only those fragments that are below the threshold. This threshold p-value score can be on the order of 0.1, 0.01, 0.001, 0.0001, or similar.

[0201] According to the example results of process 800, the analysis system can produce a median (range) of 2,800 (1,500 - 12,000) fragments with an abnormal methylation pattern for the cancer-free participants in the training, and a median (range) of 3,000 (1,200 - 420,000) fragments with an abnormal methylation pattern for the cancer participants in the training. The set of screened fragments with an abnormal methylation pattern can be used for downstream analysis as described in Section IV.B below.

[0202] In some embodiments, the analysis system uses an 860 sliding window to determine the likelihood of the methylation status vector and calculate the p-value. The analysis system only needs to enumerate the likelihoods and calculate the p-values for windows that contain consecutive CpG sites, where the length of the window (of CpG sites) is shorter than at least some of the fragments (otherwise, the window would be meaningless), rather than enumerating the likelihoods and calculating the p-values for all methylation status vectors. The window length can be static, user-determined, dynamic, or otherwise selected.

[0203] When calculating the p-value of a methylation status vector larger than the window, the window can start identifying the set of consecutive CpG sites within the window in the vector from the first CpG site in the vector. The analysis system can calculate a p-value score for the window including the first CpG site. Then, the analysis system can "slide" the window to the second CpG site in the vector and calculate another p-value score for the second window. Thus, for a window size l and a methylation vector length m, each methylation status vector can generate m - l + 1 p-value scores. After completing the p-value calculation for each part of the vector, the lowest p-value score among all the sliding windows can be used as the overall p-value score for the methylation status vector. In other embodiments, the analysis system aggregates the p-value scores of the methylation status vector to generate an overall p-value score.

[0204] Using a sliding window can help reduce the number of methylation status vector possibilities that need to be enumerated and their corresponding probability calculations that would otherwise have to be performed. As a practical example, a fragment may have up to 54 CpG sites. The analysis system can use a window of size 5 (for example) and perform 50 p-value calculations for each of the 50 windows of the methylation status vector of the fragment, without having to calculate the probabilities of 2^54 (about 1.8×10^16) possibilities to generate a single p-value. Each of these 50 calculations can enumerate 2^5 (32) methylation status vector possibilities, for a total of 50×2^5 (1.6×10^3) probability calculations. This can greatly reduce the calculations that need to be performed, with no significant impact on the accurate identification of abnormal fragments.

[0205] In embodiments of an indeterminate state, the analysis system may calculate a p-value score that sums the CpG sites with an indeterminate state in the methylation status vector of the fragment. The analysis system can identify all possibilities that are consistent with all methylation states in the methylation status vector except for the indeterminate state. The analysis system may assign a probability that is the sum of the probabilities of these identified possibilities to the methylation status vector. For example, the analysis system can calculate the probability of the methylation status vector <M1, I2, U3> as the sum of the probabilities of the possibilities of the methylation status vectors <M1, M2, U3> and <M1, U2, U3>, since the methylation states of CpG sites 1 and 3 have been observed and are consistent with the methylation states of the fragment at CpG sites 1 and 3. This method of summing the CpG sites with an indeterminate state can use probability calculations of up to 2^i possibilities, where i represents the number of indeterminate states in the methylation status vector. In additional embodiments, a dynamic programming algorithm may be employed to calculate the probability of a methylation status vector with one or more indeterminate states. Advantageously, the dynamic programming algorithm runs in linear computational time.

[0206] In some embodiments, by caching at least some computations, the computational load of computing probabilities and / or p-value scores may be further reduced. For example, the analysis system may cache the probability computations of the methylation state vector (or its window) likelihoods in transient or permanent memory. If other segments have the same CpG sites, caching the probabilities of these likelihoods may allow for efficient computation of p-value scores without having to recompute the probabilities of underlying likelihoods. Equivalently, the analysis system may compute p-value scores for each likelihood of the methylation state vector associated with a set of CpG sites in the vector (or its window). The analysis system may cache these p-value scores for use in determining the p-value scores of other segments that include the same CpG sites. Overall, the p-value scores computed for multiple likelihoods with the same CpG sites in the methylation state vector may be used to determine the p-value score of another one of these multiple likelihoods under the same set of CpG sites.

[0207] Prior to training the regional model or cancer classifier, one or more nucleic acid methylation segments may be screened. Screening the nucleic acid methylation segments may include removing each corresponding nucleic acid methylation segment that does not meet one or more selection criteria (e.g., below or above a selection criterion) from the corresponding plurality of nucleic acid methylation segments. The one or more selection criteria may include a p-value threshold. The output p-value of the corresponding nucleic acid methylation segment may be determined based at least in part on comparing the corresponding methylation pattern of the corresponding nucleic acid methylation segment with the distribution of the corresponding methylation patterns of the nucleic acid methylation segments in the healthy non-cancer cohort dataset that have the corresponding plurality of CpG sites of the corresponding nucleic acid methylation segment.

[0208] Filtering multiple nucleic acid methylation fragments can include removing each corresponding nucleic acid methylation fragment that does not meet a p-value threshold. Using the methylation patterns observed in the first batch of multiple nucleic acid methylation fragments, a filter is applied to the methylation pattern of each corresponding nucleic acid methylation fragment. Each corresponding methylation pattern of each corresponding nucleic acid methylation fragment (e.g., fragment one, …, fragment N) can include a corresponding one or more methylation sites (e.g., CpG sites), which are identified by methylation site identifiers and corresponding methylation patterns, represented as a sequence containing multiple 1s and multiple 0s, where each “1” represents a methylated CpG site among these one or more CpG sites, and each “0” represents an unmethylated CpG site among these one or more CpG sites. The methylation patterns observed in the first batch of multiple nucleic acid methylation fragments can be used to construct a methylation state distribution for the status of these CpG sites, which are jointly represented by the first batch of multiple nucleic acid methylation fragments (e.g., CpG site A, CpG site B, …, CpG site ZZZ). More details regarding nucleic acid methylation fragment processing are disclosed in U.S. Provisional Patent Application No. 17 / 191,914, filed on March 4, 2021, titled “Systems and Methods for Cancer Condition Determination Using Autoencoders,” which is hereby incorporated by reference in its entirety.

[0209] When the abnormal methylation score of a corresponding nucleic acid methylation fragment is below an abnormal methylation score threshold, the nucleic acid methylation fragment may fail to meet a selection criterion among one or more selection criteria. In such cases, the abnormal methylation score can be determined by a mixture model. For example, the mixture model can determine the likelihood of the methylation state vector (e.g., methylation pattern) of a corresponding nucleic acid methylation fragment based on the number of possible methylation state vectors of the same length and at the same corresponding genomic location, so as to detect an abnormal methylation pattern in the nucleic acid methylation fragment. This process can be performed by generating multiple possible methylation states for a vector of a specified length at each genomic location of a reference genome. Using the multiple possible methylation states, the total number of possible methylation states can be determined, and subsequently the probability of each predicted methylation state at the genomic location can be determined. Then, by matching the sample nucleic acid methylation fragment with the predicted (e.g., possible) methylation state and retrieving the calculated probability of the predicted methylation state, the likelihood of the sample nucleic acid methylation fragment corresponding to the genomic location within the reference genome can be determined. Then, based on the probability of the sample nucleic acid methylation fragment, an abnormal methylation score can be calculated.

[0210] When the number of residues of a corresponding nucleic acid methylation fragment is lower than a threshold number, the nucleic acid methylation fragment fails to meet the selection criteria in one or more selection criteria. The threshold number of residues can be between 10 and 50, between 50 and 100, between 100 and 150, or more than 150. The threshold number of residues can be a fixed value between 20 and 90. When the number of CpG sites of a corresponding nucleic acid methylation fragment is lower than a threshold number, the nucleic acid methylation fragment may fail to meet the selection criteria in one or more selection criteria. The threshold number of CpG sites can be 4, 5, 6, 7, 8, 9, or 10. When the genomic start position and genomic end position of a corresponding nucleic acid methylation fragment indicate that the number of nucleotides represented by the nucleic acid methylation fragment in the human genome reference sequence is lower than a threshold number, the nucleic acid methylation fragment fails to meet the selection criteria in one or more selection criteria.

[0211] The screening can remove nucleic acid methylation fragments that have the same corresponding methylation pattern, the same corresponding genomic start position and genomic end position as another nucleic acid methylation fragment among the corresponding multiple nucleic acid methylation fragments. This screening step can remove completely redundant fragments, including, in some cases, PCR duplicate fragments. The screening can remove nucleic acid methylation fragments that have the same corresponding genomic start position and genomic end position as another nucleic acid methylation fragment among the corresponding multiple nucleic acid methylation fragments, and the number of different methylation states is lower than a threshold number. The threshold number for retaining different methylation states of nucleic acid methylation fragments can be 1, 2, 3, 4, 5, or more than 5. For example, if the first nucleic acid methylation fragment has the same corresponding genomic start and end positions as the second nucleic acid methylation fragment, but has at least 1, at least 2, at least 3, at least 4, or at least 5 different methylation states at the corresponding CpG sites (e.g., aligned with the reference genome), then the first nucleic acid methylation fragment is retained. Another example is that if the first nucleic acid methylation fragment has the same methylation state vector (e.g., methylation pattern) as the second nucleic acid methylation fragment, but has different corresponding genomic start and end positions, then the first nucleic acid methylation fragment is also retained.

[0212] The screening can remove detection artifacts in multiple nucleic acid methylation fragments. Removing detection artifacts can include removing sequence reads obtained from sequencing hybridization probes and / or sequence reads obtained from sequences that did not undergo conversion during bisulfite conversion. The screening can remove contaminants (e.g., contaminants generated during sequencing, nucleic acid isolation, and / or sample preparation).

[0213] The screening can be based on the mutual information between the corresponding methylation fragments and the cancer status in multiple training subjects, and a subset of methylation fragments can be removed from the multiple methylation fragments. For example, mutual information can measure the degree of interdependence between two simultaneously sampled conditions of interest. The mutual information can be determined by selecting a set of independent CpG sites (e.g., within all or part of the nucleic acid methylation fragments) from one or more datasets and comparing the probabilities of the methylation status of the CpG site set between two sample groups (e.g., genotype datasets, biological samples, and / or subsets and / or groups of subjects). The mutual information score can represent the probability of the methylation pattern of the first condition and the second condition at the corresponding region within the corresponding frame of the sliding window, thereby indicating the discriminative ability of the corresponding region. When the sliding window moves forward over the selected CpG site set and / or the selected genomic region, the mutual information score can be similarly calculated for each region in each frame of the sliding window. More detailed information regarding mutual information screening is disclosed in U.S. Patent Application 17 / 119,606, filed on December 11, 2020, titled "Cancer Classification Using Patch Convolutional Neural Networks", which is hereby incorporated by reference in its entirety. IV.A.ii. Hypermethylated Fragments and Hypomethylated Fragments

[0214] In some embodiments, the analysis system identifies 870 from the screened set and determines hypomethylated fragments or hypermethylated fragments as abnormal fragments. The analysis system identifies hypermethylated fragments that have more than a threshold number of CpG sites and more than a threshold percentage of these CpG sites are methylated. The analysis system identifies hypomethylated fragments that have more than a threshold number of CpG sites and more than a threshold percentage of these CpG sites are not methylated. Example thresholds for the length of the fragment (or CpG site) include more than 3, 4, 5, 6, 7, 8, 9, 10, etc. Example thresholds for the methylation or non-methylation percentage include more than 80%, 85%, 90%, or 95%, or any other percentage within the range of 50%-100%. IV.B. Training of the Cancer Classifier

[0215] According to one embodiment, Figure 9A is a flowchart depicting the process 900 of training a cancer classifier. The analysis system obtains 910 multiple training samples, each of which has a set of abnormal fragments and a cancer type label. The multiple training samples can include any combination of samples from healthy individuals with an overall label of "non-cancer"; from subjects with an overall label of "cancer" or with a specific label (e.g., "breast cancer", "lung cancer", etc.). Training samples from subjects for a particular cancer type may be referred to as a cohort or cancer type cohort for that cancer type.

[0216] The analysis system determines 920 feature vectors for each training sample based on the set of abnormal segments of the training samples. The analysis system can calculate an abnormal score for each CpG site in the initial CpG site set. The initial CpG site set may be all or a part of the CpG sites in the human genome - its order of magnitude may be 10 4 、10 5 、10 6 、10 7 、10 8 and so on. In one embodiment, the analysis system uses a binary classification score to define an abnormal score for the feature vector based on the presence or absence of abnormal segments in the set of abnormal segments covering the CpG sites. In another embodiment, the analysis system defines an abnormal score based on the count of abnormal segments overlapping with the CpG site. In one example, the analysis system may use a ternary classification score, assigning a first score for the absence of abnormal segments, a second score for the presence of a small number of abnormal segments, and a third score for the presence of multiple abnormal segments. For example, the analysis system counts 5 abnormal segments in the sample overlapping with the CpG site and calculates an abnormal score based on the count of 5. In one or more embodiments, the feature vector further includes (e.g., as identified and counted in Figure 4A and Figure 4B ) one or more features based on methylation variation for the mixture model.

[0217] Once all abnormal scores have been determined for the training samples, the analysis system can determine the feature vector as a vector including multiple elements, where each element is one of the abnormal scores associated with one of the CpG sites in the initial set. The analysis system can normalize the abnormal scores of the feature vector based on the coverage of the sample. Here, the coverage can refer to the median or average sequencing depth across all CpG sites covered by the initial CpG site set used in the classifier (or based on the set of abnormal segments of a given training sample).

[0218] For example, now referring to Figure 9B , this figure shows the matrix 922 of training feature vectors. In this example, the analysis system has identified the CpG sites [S]926 used to generate the feature vectors for the cancer classifier. The analysis system selects the training sample [N]924. The analysis system determines the first abnormal score 928 for the first arbitrary CpG site [s1], and this score will be used in the feature vector of the training sample [n1]. The analysis system examines each abnormal segment in the set of abnormal segments. If the analysis system identifies at least one abnormal segment including the first CpG site, then the analysis system determines the first abnormal score 928 of the first CpG site as 1, as shown in Figure 9BAs shown. Consider a second arbitrary CpG site [s2]. The analysis system similarly examines the set of anomalous fragments to identify if there is at least one anomalous fragment that includes the second CpG site [s2]. If the analysis system does not find any such anomalous fragments that include the second CpG site, then the analysis system determines the second anomalous score 929 for the second CpG site [s2] to be 0, as Figure 9B shown. Once the analysis system has determined all the anomalous scores for the initial set of CpG sites, the analysis system determines a feature vector for the first training sample [n1], which includes the anomalous scores, where the first anomalous score 928 for the first CpG site [s1] is 1 and the second anomalous score 929 for the second CpG site [s2] is 0, plus the subsequent anomalous scores, thus forming the feature vector [1, 0, …].

[0219] Other methods for sample characterization can be found in the following documents: U.S. Application No. 15 / 931,022 titled “Model-Based Featurization and Classification”; U.S. Application No. 16 / 579,805 titled “Mixture Model for Targeted Sequencing”; U.S. Application No. 16 / 352,602 titled “Anomalous Fragment Detection and Classification”; and U.S. Application No. 16 / 723,716 titled “Source of Origin Deconvolution Based on Methylation Fragments in Cell-Free DNA Samples”; all of which are hereby incorporated by reference in their entirety.

[0220] The analysis system may further limit the CpG sites considered for use in the cancer classifier. The analysis system calculates an information gain 930 for each CpG site in the initial set of CpG sites based on the feature vectors of the training samples. According to step 920, each training sample has a feature vector that may contain the anomalous scores for all the CpG sites in the initial set of CpG sites, which set may include up to all the CpG sites in the human genome. However, some CpG sites in the initial set of CpG sites may be less informative or may provide information that duplicates that of other CpG sites when distinguishing between multiple cancer types.

[0221] In one embodiment, the analysis system calculates the information gain of 930 for each cancer type and each CpG site in the initial set to determine whether to include the CpG site in the classifier. The information gain is calculated for training samples that are given a cancer type compared to all other training samples. For example, two random variables, "Aberrant Fragment" ("AF") and "Cancer Type" ("CT") are used. In one embodiment, AF is a binary variable indicating whether there is an aberrant fragment overlapping a given CpG site in a given sample (where the aberrant score / feature vector was determined as above). CT is a random variable indicating whether the cancer belongs to a certain specific type. The analysis system calculates the mutual information of CT with respect to AF given. That is, how much information about the cancer type can be obtained if it is known whether there is an aberrant fragment overlapping a specific CpG site. In practice, for the first cancer type, the analysis system calculates the pairwise mutual information gain with each other cancer type and sums the mutual information gains of all other cancer types.

[0222] For a given cancer type, the analysis system can use this information to rank these sites according to the specificity of the CpG sites for the cancer. This process can be repeated for all cancer types under consideration. If a particular region is typically aberrantly methylated in the training samples of a given cancer but not in the training samples of other cancer types or healthy training samples, then for that given cancer type, the CpG sites overlapped by those aberrant fragments can have a high information gain. The CpG sites of each cancer type after ranking can be added (selected) 940 to the selected CpG site set according to their ranking for use in the cancer classifier by a greedy algorithm.

[0223] In other embodiments, the analysis system can consider other selection criteria to select informative CpG sites for the cancer classifier. One selection criterion can be that the distance between the selected CpG site and any other selected CpG site is greater than a threshold. For example, the selected CpG site is more than a threshold number of base pairs (e.g., 100 base pairs) away from any other selected CpG site such that CpG sites within the threshold distance are not also selected for consideration for inclusion in the cancer classifier.

[0224] In one embodiment, based on the selected CpG site set from the initial set, the analysis system may modify 950 the feature vector of the training samples as needed. For example, the analysis system can truncate the feature vector to remove the aberrant scores corresponding to CpG sites not in the selected CpG site set.

[0225] Using the feature vectors of the training samples, the analysis system can train a cancer classifier in any of a number of ways. The feature vectors can correspond to the initial set of CpG sites in step 920 or to the selected set of CpG sites in step 950. In one embodiment, the analysis system trains 960 a binary cancer classifier based on the feature vectors of the training samples to distinguish cancer from non-cancer. In this way, the training samples used by the analysis system include non-cancer samples from healthy individuals and cancer samples from subjects. Each training sample can have one of two labels, "cancer" or "non-cancer". In this embodiment, the classifier outputs a cancer prediction indicating the likelihood of the presence or absence of cancer.

[0226] In another embodiment, the analysis system trains 970 a multi-class cancer classifier to distinguish between multiple cancer types (also referred to as the tissue of origin (TOO) labels). The cancer types can include one or more cancers and may also include non-cancer types (which may also include any additional other diseases or genetic disorders, etc.). To this end, the analysis system can use a cancer type cohort and may or may not include a non-cancer type cohort. In this multi-cancer embodiment, the cancer classifier is trained to determine a cancer prediction (or, more specifically, a TOO prediction) that contains a predicted value for each cancer type being classified. The predicted value may correspond to the likelihood that a given training sample (during inference, corresponding to a test sample) has each cancer type. In one implementation, the predicted values are scored between 0 and 100, where the sum of the predicted values equals 100. For example, the cancer classifier returns a cancer prediction that includes predicted values for breast cancer, lung cancer, and non-cancer. For example, the classifier can return a cancer prediction that states that the test sample has a 65% likelihood of being breast cancer, a 25% likelihood of being lung cancer, and a 10% likelihood of being non-cancer. The analysis system may further evaluate the predicted values to generate a prediction of the presence of one or more cancers in the sample, also referred to as a TOO prediction, which indicates one or more TOO labels, e.g., the first TOO label with the highest predicted value, the second TOO label with the second highest predicted value, etc. Continuing with the above example, given these percentages, in this example, the system may determine that the sample has breast cancer because breast cancer has the highest likelihood. The analysis system can further confirm the TOO prediction by predicting the component proportions of the sample using a Figure 4A and Figure 4B hybrid model. This is further described above in Figure 4B 。

[0227] In these two embodiments, the analysis system trains a cancer classifier by inputting a set of training samples with their feature vectors into the cancer classifier and adjusting the classification parameters, so that the function of the classifier can accurately associate the training feature vectors with their corresponding labels. The analysis system may group the training samples into one or more sets of training samples for iteratively training the cancer classifier in batches. After inputting all the sets of training samples including their training feature vectors and adjusting the classification parameters, the cancer classifier can be fully trained to label the test samples according to their feature vectors within a certain error margin. The analysis system may train the cancer classifier according to any one of many methods. For example, a binary cancer classifier may be an L2-regularized logistic regression classifier trained using a logarithmic loss function. For another example, a multi-cancer classifier may be a multinomial logistic regression classifier. In practice, other techniques may be used to train these two types of cancer classifiers. There are many of these techniques, including function kernel methods, random forest classifiers, mixture models, autoencoder models, machine learning algorithms (such as multi-layer neural networks), etc. that may be used.

[0228] The classifier may include a logistic regression algorithm, a neural network algorithm, a support vector machine algorithm, a naive Bayes algorithm, a nearest neighbor algorithm, a boosting tree algorithm, a random forest algorithm, a decision tree algorithm, a multinomial logistic regression algorithm, a linear model, or a linear regression algorithm. IV.C. Deployment of the Cancer Classifier

[0229] During the use of the cancer classifier, the analysis system may obtain test samples from a subject with an unknown cancer type. The analysis system may use any combination of processes 300 and 830 to process the test samples composed of DNA molecules, thereby obtaining a set of abnormal fragments. The analysis system can determine test feature vectors for use by the cancer classifier according to the similar principles discussed in process 900. The analysis system may calculate an abnormal score for each of the multiple CpG sites used by the cancer classifier. For example, the cancer classifier receives as input a feature vector including the abnormal scores of 1,000 selected CpG sites. Therefore, based on the set of abnormal fragments, the analysis system can determine a test feature vector including the abnormal scores of the 1,000 selected CpG sites. The analysis system may calculate the abnormal score in the same way as for the training samples. In some embodiments, the analysis system defines the abnormal score as a binary classification score based on the presence or absence of hypermethylated or hypomethylated fragments in the set of abnormal fragments covering the CpG sites. The analysis system may generate a test feature vector that includes one or more features based on methylation variations (such as Figure 4A and Figure 4B as described).

[0230] Then, the analysis system can input the test feature vector into the cancer classifier. Then, based on the classification parameters trained in process 900 and the test feature vector, the function of the cancer classifier can generate a cancer prediction. In the first manner, the cancer prediction can be binary classification, selected from the group consisting of "cancer" or "non-cancer"; in the second manner, the cancer prediction is selected from the group consisting of multiple cancer types and "non-cancer". In another embodiment, the cancer prediction includes a prediction value for each of the multiple cancer types. In addition, the analysis system may determine that the test sample is most likely to belong to one of these cancer types. Continuing with the above example, for a cancer prediction of a test sample with a 65% likelihood of breast cancer, a 25% likelihood of lung cancer, and a 10% likelihood of non-cancer, the analysis system may determine that the test sample is most likely to have breast cancer. In another example, where the cancer prediction is binary classification, with a 60% likelihood of non-cancer and a 40% likelihood of cancer, the analysis system determines that the test sample is most likely not to have cancer. In another embodiment, to identify which cancer type the test subject has, the cancer prediction with the highest likelihood may still be compared with a threshold (e.g., 40%, 50%, 60%, 70%). If the cancer prediction with the highest likelihood does not exceed that threshold, the analysis system may return an indeterminate result.

[0231] In another embodiment, the analysis system cascades the cancer classifier trained in step 960 of process 900 with another cancer classifier trained in step 970 of process 900. The analysis system can input the test feature vector into the cancer classifier trained as a binary classifier in step 960 of process 900. The analysis system can receive the output of the cancer prediction. The cancer prediction may be binary classification as to whether the test subject is likely to have cancer. In other embodiments, the cancer prediction includes prediction values describing the likelihood of cancer and the likelihood of non-cancer. For example, the cancer prediction has a cancer prediction value of 85% and a non-cancer prediction value of 15%. The analysis system may determine that the test subject is likely to have cancer. Once the analysis system determines that the test subject is likely to have cancer, the analysis system may input the test feature vector into a multi-class cancer classifier trained to distinguish different cancer types. The multi-class cancer classifier can receive the test feature vector and return a cancer prediction of one of the multiple cancer types. For example, the multi-class cancer classifier provides a cancer prediction indicating that the test subject is most likely to have ovarian cancer. In another embodiment, the multi-class cancer classifier provides prediction values for each of the multiple cancer types. For example, the cancer prediction may include a breast cancer type prediction value of 40%, a colorectal cancer type prediction value of 15%, and a liver cancer prediction value of 45%.

[0232] According to a general implementation example of binary cancer classification, an analysis system can determine a cancer score for a test sample based on sequencing data of the test sample (e.g., methylation sequencing data, SNP sequencing data, other DNA sequencing data, RNA sequencing data, etc.). The analysis system can compare the cancer score of the test sample with a binary threshold critical value for predicting whether the test sample is likely to have cancer. Using a TOO threshold based on one or more TOO subtype categories, the binary threshold critical value can be adjusted. The analysis system may further generate a feature vector for the test sample for use in a multi-class cancer classifier to determine a cancer prediction indicating one or more likely cancer types.

[0233] A classifier can be used to determine the disease state of a test subject (e.g., a subject with an unknown disease state). The method can include obtaining a test genomic data construct in electronic format (e.g., single time point test data) that includes values for each of a plurality of genomic features corresponding to a plurality of nucleic acid fragments in a biological sample obtained from the test subject. Then, the method can include applying the test genomic data construct to a test classifier to thereby determine the state of the disease condition of the test subject. The test subject may not have been previously diagnosed with the disease condition.

[0234] The classifier can be a time classifier that uses at least (i) a first test genomic data construct generated from a first biological sample obtained from the test subject at a first time point and (ii) a second test genomic data construct generated from a second biological sample obtained from the test subject at a second time point.

[0235] A trained classifier can be used to determine the disease state of a test subject (e.g., a subject with an unknown disease state). In this case, the method can include obtaining a test time series data set of the test subject in electronic format, where: for each respective time point among a plurality of time points, the test time series data set includes a corresponding test genotype data construct that includes values for a plurality of genotype features corresponding to a plurality of nucleic acid fragments in a corresponding biological sample obtained from the test subject at the respective time point; for each pair of respective consecutive time points among the plurality of time points, the test time series data set includes an indication of the length of time between the pair of respective consecutive time points. Then, the method can include applying the test genotype data construct to a test classifier to thereby determine the state of the disease condition of the test subject. The test subject may not have been previously diagnosed with the disease condition. V. Applications

[0236] In some embodiments, the methods, analysis systems, and / or classifiers of the present invention can be used to detect the presence of cancer, monitor cancer progression or recurrence, monitor treatment response or efficacy, 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 probability score (e.g., from 0 to 100) that describes the likelihood that a test feature vector is from a subject with cancer. In some embodiments, the probability score is compared to a threshold probability to determine whether the subject has cancer. In other embodiments, the likelihood or probability score can be evaluated at multiple different time points (e.g., before or after treatment) to monitor disease progression or to monitor treatment efficacy (e.g., treatment effect). In still other embodiments, the likelihood or probability score can be used to make or influence clinical decisions (e.g., cancer diagnosis, treatment selection, treatment effect assessment, etc.). For example, in one embodiment, if the probability score exceeds a threshold, a doctor can prescribe an appropriate treatment. V.A. Early Detection of Cancer

[0237] In some embodiments, the methods and / or classifiers of the present invention are used to detect the presence or absence of cancer in a subject suspected of having cancer. For example, a classifier (e.g., as described in Part IV above) can be used to determine a cancer prediction that describes the likelihood that a test feature vector is from a subject with cancer.

[0238] In one embodiment, the cancer prediction is the likelihood (e.g., a score between 0 and 100) that a test sample has cancer (i.e., binary classification). Thus, the analysis system may determine a threshold for determining whether a test subject has cancer. For example, a cancer prediction greater than or equal to 60 may indicate that the subject has cancer. In still other embodiments, a cancer prediction 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 indicates that the subject has cancer. In other embodiments, the cancer prediction can indicate the severity of the disease. For example, a cancer prediction of 80 as compared to a cancer prediction below 80 (e.g., a probability score of 70) can indicate a more severe form or later stage of cancer. Similarly, over time, an increasing cancer prediction (e.g., determined by classifying test feature vectors of multiple samples collected from the same subject at two or more time points) can indicate disease progression, or over time, a decreasing cancer prediction can indicate treatment success.

[0239] In another embodiment, cancer prediction involves a number of prediction values, where each of the multiple cancer types being classified (i.e., multi-class classification) has a prediction value (e.g., a score between 0 and 100). The prediction value may correspond to the likelihood that a given training sample (during inference, corresponding to the training sample) has each cancer type. The analysis system may identify the cancer type with the highest prediction value and indicate that the test subject is likely to have that cancer type. In other embodiments, the analysis system further compares the highest prediction value to a threshold (e.g., 50, 55, 60, 65, 70, 75, 80, 85, etc.) to determine that the test subject is likely to have that cancer type. In other embodiments, the prediction value may also indicate the severity of the disease. For example, a prediction value greater than 80 as compared to a prediction value of 60 may indicate a more severe form or later stage of cancer. Similarly, over time, an increasing prediction value (e.g., determined by classifying test feature vectors from multiple samples taken from the same subject at two or more time points) may indicate disease progression, or over time, a decreasing prediction value may indicate treatment success.

[0240] In accordance with various aspects of the present invention, the methods and systems of the present invention can be trained to detect or classify multiple cancer signatures. For example, the methods, systems, and classifiers of the present 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 cancer types.

[0241] Examples of cancers that can be detected using the methods, systems, and classifiers of the present invention include carcinoma, lymphoma, blastoma, sarcoma, and leukemia or lymphoid malignancies. More specific examples of such cancers include, but are not limited to, squamous cell carcinoma (e.g., epithelial squamous cell carcinoma), skin cancer, melanoma, lung cancer (including small cell lung cancer, non-small cell lung cancer (“NSCLC”), lung adenocarcinoma, and lung squamous cell carcinoma), peritoneal cancer, gastric (gastric or stomach) cancer (including gastrointestinal cancer), pancreatic cancer (e.g., pancreatic ductal adenocarcinoma), cervical cancer, ovarian cancer (e.g., high-grade serous ovarian cancer), liver cancer (e.g., hepatocellular carcinoma (HCC)), hepatoma, liver tumor, bladder cancer (e.g., urothelial bladder cancer), testicular (germ cell tumor) cancer, breast cancer (e.g., HER2-positive, HER2-negative, and triple-negative breast cancer), brain cancer (e.g., astrocytoma, glioma (e.g., glioblastoma)), colon cancer, rectal cancer, colorectal cancer, endometrial cancer or uterine cancer, salivary gland cancer, kidney cancer or renal cancer (e.g., renal cell carcinoma, nephroblastoma (nephroblastoma or Wilms’ tumor)), prostate cancer, vulvar cancer, thyroid cancer, anal canal cancer, penile cancer, head and neck cancer, esophageal cancer, and nasopharyngeal cancer (NPC). Additional examples of cancers include, but are not limited to, retinoblastoma, thecoma, ovatestoma, malignant hematological diseases, including but not limited to non-Hodgkin lymphoma (NHL), multiple myeloma, and acute malignant hematological diseases, endometriosis, fibrosarcoma, choriocarcinoma, laryngeal cancer, Kaposi sarcoma, schwannoma, oligodendroglioma, neuroblastoma, rhabdomyosarcoma, osteosarcoma, leiomyosarcoma, and urothelial cancer.

[0242] In some embodiments, the cancer is one or more of anal rectal 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, uterine cancer, or any combination thereof.

[0243] In some embodiments, one or more cancers can be “high-signal” cancers (defined as cancers with a 5-year cancer-specific mortality rate greater than 50%), such as anal rectal cancer, colorectal cancer, esophageal cancer, head and neck cancer, hepatobiliary cancer, lung cancer, ovarian cancer, and pancreatic cancer, as well as lymphoma and multiple myeloma. High-signal cancers tend to be more aggressive and typically have a higher-than-average concentration of free nucleic acids in test samples obtained from patients. V.B. Cancer and Treatment Monitoring

[0244] In some embodiments, cancer prediction can be evaluated at multiple different time points (e.g., before or after treatment) to monitor disease progression or to monitor treatment efficacy (e.g., treatment effect). For example, the methods included in the present invention involve obtaining a first sample (e.g., a first plasma cfDNA sample) from a cancer patient at a first time point and determining a first cancer prediction therefrom (as described herein), and 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 cancer prediction therefrom (as described herein).

[0245] In certain embodiments, the first time point is before cancer treatment (e.g., before resection surgery or treatment intervention), the second time point is after cancer treatment (e.g., after resection surgery or treatment intervention), and a classifier is used to monitor the efficacy of the treatment. For example, if the second cancer prediction decreases compared to the first cancer prediction, then the treatment is considered successful. However, if the second cancer prediction increases compared to the first cancer prediction, then the treatment is considered unsuccessful. In other embodiments, both the first time point and the second time point are before cancer treatment (e.g., before resection surgery or treatment intervention). In still other embodiments, both the first and second time points are after cancer treatment (e.g., after resection surgery or treatment intervention). In still other embodiments, cfDNA samples may be obtained and analyzed from the cancer patient at the first and second time points, e.g., to monitor cancer progression, to determine whether the cancer is in remission (e.g., after treatment), to monitor or detect residual disease or disease recurrence, or to monitor treatment (e.g., therapeutic) effect.

[0246] Those skilled in the art can easily understand that a test sample can be obtained from a cancer patient at any desired set of time points and analyzed according to the method of the present invention to monitor the cancer status of the patient. In some embodiments, the time duration between the first time point and the second time point ranges 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 50 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, the frequency of obtaining a test sample from a patient can be: at least once every 5 months, at least once every 6 months, at least once a year, at least once every 2 years, at least once every 3 years, at least once every 4 years or at least once every 5 years. V.C. Processing

[0247] In yet another embodiment, cancer prediction can be used to make or influence clinical decisions (e.g., cancer diagnosis, treatment selection, assessment of treatment efficacy, etc.). For example, in one embodiment, if the cancer prediction (e.g., for cancer or for a specific cancer type) exceeds a threshold, a doctor can prescribe an appropriate treatment (e.g., resection surgery, radiotherapy, chemotherapy, and / or immunotherapy). The doctor can prescribe an appropriate treatment plan based on the analysis performed by the analysis system (e.g. Figure 1 Analysis 140).

[0248] A classifier (as described herein) can be used to determine a cancer prediction that a sample feature vector is from a subject with cancer. In one embodiment, when the cancer prediction exceeds a threshold, an appropriate treatment (e.g., resection surgery or therapeutic intervention) is prescribed. For example, in one embodiment, if the cancer prediction is greater than or equal to 60, one or more appropriate treatments are prescribed. In another embodiment, if the cancer prediction 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 treatments are prescribed. In other embodiments, the cancer prediction can indicate the severity of the disease. Then an appropriate treatment matching the disease severity can be prescribed.

[0249] In some embodiments, the treatment is one or more cancer therapeutic agents selected from the group consisting of: chemotherapeutic agents, targeted cancer therapeutic agents, differentiation therapeutic agents, hormone therapeutic agents, and immunotherapeutic agents. For example, the treatment can be one or more chemotherapeutic agents selected from the group consisting of: 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 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, proteasome inhibitors, angiogenesis inhibitors, and monoclonal antibody conjugates. In some embodiments, the treatment is one or more differentiation therapeutic agents including retinoids, such as tretinoin, alitretinoin, and bexarotene. In some embodiments, the treatment is one or more hormone therapeutic agents selected from the group consisting of: antiestrogens, aromatase inhibitors, progesterones, estrogens, antiandrogens, and GnRH agonists or analogs. In one embodiment, the treatment is one or more immunotherapeutic agents 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 (e.g., thalidomide and lenalidomide (REVLIMID)). An experienced physician or oncologist can select an appropriate cancer therapeutic agent based on multiple characteristics, such as tumor type, cancer stage, previous exposure to cancer treatments or therapeutic agents, and other characteristics of the cancer. VI. Kit Embodiments

[0250] The present disclosure also discloses kits for performing the above methods, including methods related to cancer classifiers. The kits can include one or more collection containers for collecting samples containing genetic material from an individual. The samples can include blood, plasma, serum, urine, feces, saliva, other types of body fluids, or any combination thereof. Such kits can include reagents for isolating nucleic acids from the samples. The reagents can further include reagents for sequencing the nucleic acids, including buffers and detection agents. In one or more embodiments, the kits can include one or more sequencing sets that include probes for targeting specific genomic regions, specific mutations, specific genetic variations, or some combination thereof. In other embodiments, the samples collected via the kits are provided to a sequencing laboratory, which can use the sequencing sets to sequence the nucleic acids in the samples.

[0251] The kits can further include instructions for using the reagents contained in the kits. For example, the kits can include instructions for collecting samples, extracting nucleic acids from the test samples. Exemplary instructions can be the order of adding reagents, the centrifugation speed for isolating nucleic acids from the test samples, how to amplify nucleic acids, how to sequence nucleic acids, or any combination thereof. The instructions can further clarify how to operate a computing device as the analysis system 200 to perform the steps of any of the described methods or processes.

[0252] In addition to the above components, the kits can further include a computer-readable storage medium having stored thereon computer software for performing the various methods described throughout this disclosure. One form in which these instructions can exist is as printed information on a suitable medium or substrate (e.g., one or more sheets of paper on which the information is printed), in the packaging of the kit, in a package insert. Another way would be a computer-readable medium, such as a floppy disk, CD, hard drive, network data storage, in which the instructions are stored in the form of computer code. Yet another means that can exist is a web address that can be used to access information at a remote site via the Internet. VII. Example Results VII.A. Sample Collection and Processing

[0253] Study Design and Samples: CCGA (NCT02889978) is a prospective, multi-center, case-control, longitudinal follow-up observational study. De-identified biological samples were collected from approximately 15,000 participants at 342 study sites. The samples were divided into a training set (1,785) and a test set (1,015); the samples were selected to ensure a predetermined distribution of cancer types and non-cancers at each study site in each cohort, and the cancer and non-cancer samples were age-frequency matched by gender.

[0254] Whole-genome bisulfite sequencing: cfDNA was isolated from plasma and analyzed by whole-genome bisulfite sequencing (WGBS; depth 30x). cfDNA was extracted from two tubes of plasma (combined volume up to 10 ml) from each patient using the modified QIAamp Circulating Nucleic Acid Kit (Qiagen; Germantown, MD). Bisulfite conversion of up to 75 ng of plasma cfDNA was performed using the EZ-96 DNA Methylation Kit (Zymo Research, D5003). The converted cfDNA was used to prepare dual-index sequencing libraries using the Accel-NGS Methyl-Seq DNA Library Preparation Kit (Swift BioSciences; Ann Arbor, MI), and the constructed libraries were quantified using the KAPA Library Quantification Kit for Illumina platforms (Kapa Biosystems; Wilmington, MA). Four libraries were pooled with 10% PhiX v3 library (Illumina, FC-110-3001) and clustered on an Illumina NovaSeq 7000 S2 flow cell, followed by 150-bp paired-end sequencing (30x).

[0255] For each sample, the WGBS fragment set was reduced to a small subset of fragments with aberrant methylation patterns. Additionally, hypermethylated or hypomethylated cfDNA fragments were selected. cfDNA fragments with aberrant methylation patterns that were hypermethylated or hypomethylated, i.e., UFXM, were selected. Fragments that occurred frequently in cancer-free individuals or were methylation-unstable were less likely to generate highly discriminatory features for cancer status classification. Therefore, we generated a statistical model and data structure of typical fragments using an independent reference set (i.e., reference genome) of 108 cancer-free, non-smoking participants (age: 58 ± 14 years, 79 cases [73%] female) from the CCGA study. These samples were used to train a Markov chain model (order 3), which estimates the likelihood of a given sequence of CpG methylation states within a fragment. This model has been shown to be calibrated within the normal fragment range (p-value > 0.001) and was used to exclude fragments with a p-value ≥ 0.001 from the Markov model as they were not aberrant enough.

[0256] As described above, a further data reduction step only selects fragments that cover at least 5 CpGs and have an average methylation either >0.9 (hypermethylation) or <0.1 (hypomethylation). This process results in a median (range) of 2,800 (1,500 - 12,000) UFXM fragments for participants without cancer in the training, and a median (range) of 3,000 (1,200 - 420,000) UFXM fragments for participants with cancer in the training. Since this data reduction process only uses reference set data, this stage only needs to be applied to each sample once. VII.B. Mixture Model Results

[0257] The following mixture model results are from a mixture model trained and deployed according to the methods described in Figure 4A and 4B and the methods described throughout this disclosure. Training samples containing methylated sequence reads are used to train the mixture model. The analysis system can adjust or fix the number of components in the mixture model as a hyperparameter. Once training is complete, the analysis system uses a validation set of samples to validate the accuracy of the mixture model.

[0258] Figure 10 are example results of a two-component mixture model according to an example embodiment. The mixture model is trained to discern the component ratios between two components. The analysis system uses a validation set of forty samples with known varying component ratios. The average sequence depth of the samples is 10. The mixture model evaluates over 5,000 methylation variations in the genome. The analysis system applies the trained two-component mixture model to the methylation profiles of the validation set of forty samples.

[0259] The top graph 1010 shows the accuracy of the mixture model in predicting component ratios. In the top graph 1010, the x-axis represents the true component ratio between the two components of the forty validation samples, and the y-axis represents the component ratio predicted by the mixture model. The diagonal line from (0.00, 0.00) to (1.00, 1.00) represents an accurate prediction of the component ratio. Qualitatively, all forty validation samples are on or touch the diagonal line.

[0260] The bottom graph 1020 shows the ability of the mixture model to predict methylation profiles based on known component ratios. In Figure 1020, the x-axis represents the true methylation variant allele fraction, and the y-axis represents the methylation variant allele fraction predicted by the mixture model. The diagonal line from (0.00, 0.00) to (1.00, 1.00) represents an accurate prediction of the methylation variant allele fraction. For a given value on Figure 1020, the accuracy of the prediction can be determined based on the degree of deviation from the diagonal line. Qualitatively, the predictions of the methylation variant allele fraction rarely deviate significantly from the true methylation variant allele fraction.

[0261] Figure 11 These are example results of a five-component mixture model according to an example embodiment. The mixture model was trained to discern the component proportions between five components. The analysis system utilized a validation set that included sixteen samples in five classes. Each class of samples had the same known component proportions. Specifically, class 1 included 60% of the first component (dark green) and 10% of each of the remaining components; class 2 included 60% of the second component (orange) and 10% of each of the remaining components; class 3 included 60% of the third component (blue) and 10% of each of the remaining components; class 4 included 60% of the fourth component (pink) and 10% of each of the remaining components; and class 5 included 60% of the fifth component (light green) and 10% of each of the remaining components. The mixture model evaluated over 5,000 methylation variations in the genome. The analysis system applied the trained five-component mixture model to the methylation profiles of the validation set of five classes of samples. The top graph 1110 shows the five classes of samples and their known component proportion breakdowns. Each class is grouped along the x-axis. The y-axis represents percentages or proportions. The bottom graph 1120 shows the mixture model's predictions of the component proportions for the five classes. The mixture model accurately predicted that the first component (dark green) was more than 50% in all samples of class 1, the second component (orange) was more than 50% in all samples of class 2, the third component (blue) was more than 50% in all samples of class 3, the fourth component (pink) was more than 50% in all samples of class 4, and the fifth component (light green) was more than 50% in all samples of class 5. Although the validation samples can be artificially generated to have precise component proportions, the mixture model was able to accurately identify the major component proportion for each of the eighty samples.

[0262] Figure 12This is an example result of deconvolving the positive hormone receptor status of breast tissue, the negative hormone receptor status of breast tissue, prostate tissue, and uterine tissue using a mixture model according to an example embodiment. The mixture model is trained to allow customization of the number of components. The mixture model is applied to validation samples known to have one of the cancer types corresponding to the components. Each category is grouped along the x-axis. The y-axis represents a percentage or proportion. The red component represents a non-cancer signal or a general signal common in many samples. In the prostate category (all samples diagnosed with prostate cancer) and the uterine category (all samples diagnosed with uterine cancer), the mixture model mainly predicts that the samples (except for the red component) have a major proportion of the orange component (for the prostate category) and the purple component (for the uterine category). Although the mixture model is independent of the labels of the components, the mixture model successfully performs deconvolution and attributes similar methylation features to the same component. There are a few uterine samples where the predicted purple component may be less than another non-red component, but the mixture model predicts that the purple component contributes significantly in the vast majority of uterine samples. For breast HER2– and breast HER2+, the mixture model determines that there are enough differences to warrant two independent components between the two categories. The mixture model faces some challenges in some samples that are predicted to have a significant contribution from other components. This may be due to mislabeling by the healthcare staff of the diagnostic samples or because the two tissues have similar methylation features.

[0263] Figure 13A This is an example result of deconvolving non-cancer impurities in colorectal tissue from colorectal samples using a mixture model according to an example embodiment. When training the mixture model, the analysis system adjusts the number of components in the colorectal samples shown in FIG. 1310 to three. These three components are represented by red, green, and blue. The red component seems to represent the putative non-cancer impurities present in most samples. Each column represents a colorectal sample. The grey columns represent samples enriched with disseminated tumor cells. Notably, the mixture model is able to evaluate the purity of such samples. Using the mixture model to evaluate tissue purity, the purity assessment is increased from 50% to 87%. The two components of blue and green may indicate heterogeneity in colorectal cancer tissue.

[0264] Figure 13B This is an example result of deconvolving non-cancer impurities in bladder tissue from bladder samples using a mixture model according to an example embodiment. When training the mixture model, the analysis system adjusts the number of components in the bladder samples shown in FIG. 1320 to two. These two components are represented by red and blue. The red component seems to represent the putative non-cancer impurities present in most samples. Each column represents a bladder sample. The grey columns represent samples enriched with disseminated tumor cells. Notably, the mixture model is able to evaluate the purity of such samples. Using the mixture model to evaluate tissue purity, the purity assessment is increased from 19% to 67%. VIII. Other Considerations

[0265] The foregoing detailed description of the embodiments refers to the accompanying drawings, which illustrate specific embodiments of the present disclosure. Other embodiments having different structures and operations are also within the scope of the present disclosure. Terms such as "the present invention" are used to refer to certain specific instances among many alternative aspects or embodiments of the applicant's invention shown in this specification, and their use or non-use is not intended to limit the scope of the applicant's invention or the scope of the claims.

[0266] Embodiments of the present invention may also relate to an apparatus for performing the operations herein. This apparatus may be specially constructed for the desired purposes and / or this apparatus may comprise a general-purpose computing device selectively activated or reconfigured by a computer program stored in a computer. Such a computer program may be stored in a non-transitory, tangible computer-readable storage medium, or stored in any medium type suitable for storing electronic instructions, which may be coupled to a computer system bus. In addition, any computing system mentioned in this specification may include a single processor, or may be an architecture employing a multi-processor design to improve computing power.

[0267] Any step, operation, or process described herein as being performed by an analysis system may be performed (performed or implemented) by one or more hardware or software modules of the apparatus alone or in combination with other computing devices. In one embodiment, the software module is implemented by a computer program product, which comprises a computer-readable medium containing computer program code that can be executed by a computer processor to perform any or all of the steps, operations, or processes described.

Claims

1. A method for training a machine learning hybrid model for tissue type identification, characterized in that, The method includes: Obtaining a training sample set that includes at least one thousand methylated sequence reads obtained by sequencing deoxyribonucleic acid (DNA) fragments; Modifying each training sample to generate a corresponding sample methylation signature by determining, for each of a plurality of genomic regions, a first set of methylated sequence reads that overlap the genomic region and a second set of methylated sequence reads that include a methylation variant at the genomic region, the sample methylation signature being generated at least in part based on the first set of methylated sequence reads and the second set of methylated sequence reads; Generating a training data set that includes these sample methylation signatures; And Training the machine learning hybrid model using the training data set, the machine learning hybrid model being configured to identify the contribution of each of a plurality of source tissue types to DNA fragments in a sample.

2. The method according to claim 1, wherein At least one training sample is known to include a first source tissue type among the plurality of source tissue types, wherein training the hybrid model includes training the hybrid model to identify the contribution of the first source tissue type to DNA fragments in the one training sample.

3. The method according to any one of claims 1 to 2, characterized in that, At least one training sample is known to have the contribution of DNA fragments from each of the plurality of source tissue types, wherein training the hybrid model includes training the hybrid model to identify the contribution of each of the plurality of source tissue types to DNA fragments in the one training sample.

4. The method according to any one of claims 1 to 3, characterized in that, At least one training sample is one of the following: a liquid biopsy sample, a tissue biopsy sample, and a purified sample.

5. The method according to any one of claims 1 to 4, characterized in that, At least one genomic region consists of one CpG site.

6. The method according to any one of claims 1 to 5, characterized in that At least one genomic region includes a plurality of CpG sites.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: Determining the average sequencing depth of each of the genomic regions in an initial set of genomic regions based on the methylated sequence reads of the training samples; and Filtering out genomic regions with an average sequencing depth below a threshold depth to select the plurality of genomic regions.

8. The method according to any one of claims 1 to 7, characterized in that, The methylation variant at a genomic region is one of two methylation patterns at the genomic region.

9. The method according to claim 8, wherein The two methylation patterns at a genomic region include methylation and non-methylation.

10. The method according to any one of claims 1 to 9, characterized in that, The methylation variant at a genomic region is one of more than two methylation patterns at the genomic region.

11. The method according to any one of claims 1 to 10, characterized in that, Modifying each training sample to generate the corresponding sample methylation signature further includes: For each of the plurality of genomic regions, determining a third set of methylated sequence reads that have a reference state at the genomic region, wherein the reference state is any methylation pattern that does not belong to the methylation variant, and wherein the sample methylation signature is further generated based on the third set of methylated sequence reads.

12. The method according to any one of claims 1 to 11, characterized in that, These tissue types include the following combinations: non-cancerous impurities; squamous cell carcinoma tissue; skin cancer tissue; melanoma tissue; lung cancer tissue; lung adenocarcinoma tissue; lung squamous cell carcinoma tissue; peritoneal cancer tissue; gastrointestinal cancer tissue; pancreatic cancer tissue; cervical cancer tissue; ovarian cancer tissue; liver cancer tissue; hepatoma tissue; hepatic cancer tissue; bladder cancer tissue; testicular cancer tissue; breast cancer tissue; brain cancer tissue; colon cancer tissue; rectal cancer tissue; Colorectal cancer tissue; endometrial cancer or uterine cancer tissue; salivary gland cancer tissue; kidney cancer or renal cancer tissue; prostate cancer tissue; vulvar cancer tissue; thyroid cancer tissue; anal canal cancer tissue; penile cancer tissue; head and neck cancer tissue; esophageal cancer tissue; and nasopharyngeal carcinoma (NPC) tissue.

13. The method according to claim 12, wherein The non-cancerous impurities include one or more of lymphocytes, macrophages, fibroblasts, vascular endothelial cells, or non-cancerous tissues.

14. The method according to any one of claims 12 to 13, characterized in that, The methylation signature of the non-cancerous impurities is obtained from a reference database that contains multiple methylation signatures of non-cancerous impurities.

15. The method according to any one of claims 1 to 14, characterized in that, The machine learning mixture model is trained according to maximum likelihood estimation.

16. The method according to any one of claims 1 to 15, characterized in that, Training the machine learning mixture model includes adjusting the number of tissue types as a hyperparameter of the machine learning mixture model.

17. The method according to claim 16, wherein Adjusting the number of tissue types as a hyperparameter of the machine learning mixture model includes: For each number of source tissue types within a range: Training the machine learning mixture model with that number as the hyperparameter, determining the maximum likelihood by cross-validating the trained machine learning mixture model with a held-out sample set, and applying a penalty to the maximum likelihood based on that number; and selecting the optimal number from the range as the hyperparameter based on the penalized maximum likelihood.

18. The method according to any one of claims 1 to 17, characterized in that, Training the machine learning mixture model includes training according to one or more machine learning algorithms.

19. The method according to any one of claims 1 to 18, characterized in that, The machine learning mixture model includes a first set of tissue type models, each tissue type model modeling the methylation signature of DNA fragments of a source tissue type, and wherein training the machine learning mixture model includes training the first set of tissue type models.

20. The method according to claim 19, wherein Training these first submodels includes training each tissue type model according to a beta distribution.

21. The method according to any one of claims 1 to 20, characterized in that, The machine learning mixture model includes a deconvolution model that is used to deconvolve the contribution of the source tissue type for each training sample, and wherein training the machine learning mixture model includes training the deconvolution model.

22. The method according to claim 21, wherein, Training the deconvolution model includes training the deconvolution model according to a binomial distribution.

23. The method according to any one of claims 1 to 22, characterized in that, The machine learning mixture model includes: one or more submodels at a first level for predicting the contribution of a macroscopic source tissue type; and one or more submodels at a second level for predicting the contribution of source tissue types under that macroscopic tissue type; and wherein a first-level submodel predicts the contribution of a macroscopic tissue type, and a set of one or more second-level submodels predicts the contribution of a set of tissue types under that one macroscopic tissue type, and the contribution of the set of tissue types is equal to the contribution of that one macroscopic tissue type.

24. A method for identifying the contribution of a source tissue type to DNA fragments in a test sample, characterized in that, The method includes: Obtaining the test sample, the test sample including at least one thousand methylation sequence reads of DNA fragments in the test sample; For each genomic region among a plurality of genomic regions: Determining a first set of methylation sequence reads overlapping the genomic region, and Determining a second set of methylation sequence reads having an alternative methylation signature at the genomic region, Generating a sample methylation signature of the training sample across the plurality of genomic regions based on the first set of methylation sequence reads and the second set of methylation sequence reads; and Apply a machine learning mixture model to the sample methylation features to identify the contribution of each source tissue type to the DNA fragments in the test sample. Optionally, the machine learning mixture model is trained according to the method described in any one of claims 1 to 23.

25. The method according to claim 24, wherein The method further comprises: Report the contribution of the identified source tissue type to a client device.

26. The method according to any one of claims 24 to 25, characterized in that The method further comprises: Identify the source tissue type with the greatest contribution, and Report treatment recommendations for treating cancer of the primary source of the identified source tissue type.

27. The method according to any one of claims 24 to 26, characterized in that, The method further comprises: Compare the contribution of the source tissue type with the previous contribution of the source tissue type predicted before treatment; Determine the treatment effect based on the comparison; and Report treatment recommendations based on the treatment effect.

28. The method according to any one of claims 24 to 27, characterized in that, The method further comprises: If the contribution of the first source tissue type is less than the previous contribution of the first source tissue type predicted before the treatment, determine that the treatment is successful; and Report the success of the treatment.

29. The method according to any one of claims 24 to 28, characterized in that, The method further comprises: If the contribution of the first source tissue type is greater than the previous contribution of the first source tissue type predicted before the treatment, determine that the treatment has failed; and Report the failure of the treatment.

30. The method according to claim 29, wherein, The method further comprises: Report treatment modifications based on the failure of the treatment.

31. The method according to any one of claims 24 to 30, characterized in that, The test sample is predicted to have cancer.

32. The method according to claim 31, wherein The test sample is predicted to have cancer by a cancer classifier.

33. The method according to claim 32, wherein, The cancer classifier is trained on a cancer cohort of training samples and a non-cancer cohort of training samples, wherein each training sample from the cancer cohort and the non-cancer cohort contains at least one thousand methylation sequence reads of the DNA fragments in the training sample.

34. The method according to claim 24, wherein The method further comprises: Apply the machine learning mixture model to the sample methylation features to determine a subset of methylation sequence reads originating from non-cancer impurity types; Exclude the subset of methylation sequence reads originating from the non-cancer impurity types, thereby generating a feature set of methylation sequence reads; And Apply a cancer classifier to the feature set of methylation sequence reads to predict cancer in the test sample.

35. A method for training a cancer classifier, characterized in that, The method comprises: Obtain a cancer cohort of training samples and a non-cancer cohort of training samples, wherein each training sample from the cancer cohort and the non-cancer cohort contains at least one thousand methylation sequence reads of the DNA fragments in the training sample; Generate sample methylation for each training sample by: For each genomic region among a plurality of genomic regions: Determine a first set of methylation sequence reads overlapping the genomic region, Determine a second set of methylation sequence reads having alternative methylation features at the genomic region, and wherein the sample methylation features are partially based on the first set of methylation sequence reads and the second set of methylation sequence reads, and Apply a machine learning mixture model optionally trained according to the method described in any one of claims 1 to 23 to the sample methylation features of each training sample in the cancer cohort to identify a subset of methylation sequence reads originating from non-cancer impurity types; For each training sample in the cancer cohort, a subset of the methylation sequence reads originating from the non-cancer impurity type is excluded, thereby generating a feature set of methylation sequence reads; Generate a feature set of methylation sequence reads for each training sample in the non-cancer cohort; and Train the cancer classifier using the feature set of methylation sequence reads from the training samples in the cancer cohort and the feature set of methylation sequence reads from the training samples in the non-cancer cohort.

36. The method according to claim 35, wherein The cancer classifier is trained as a machine learning model.

37. A method for training a cancer classifier, characterized in that, The method includes: Obtaining, for each training sample among a plurality of training samples, a set of methylation sequence reads obtained by sequencing DNA fragments of the training sample, wherein the plurality of training samples includes training samples obtained from a first cohort of subjects diagnosed with cancer and training samples obtained from a second cohort of subjects not diagnosed with cancer; For each training sample in the first cohort and for each methylation sequence read of the training sample: Apply a tissue type model trained to predict the likelihood that the methylation sequence read originates from non-cancerous tissue, and If the predicted likelihood is higher than a threshold, exclude the methylation sequence read; Generate methylation features for each training sample in the first cohort based on the non-excluded methylation sequence reads of the training sample; Generate methylation features for each training sample in the second cohort based on the methylation sequence reads of the training sample; and Train a cancer classifier using the methylation features of the training sample to detect the presence of a cancer signal.

38. The method according to claim 37, wherein Train the tissue type model according to a beta distribution using the methylation features of non-cancer training samples, the methylation features being generated from methylation sequence reads obtained by sequencing DNA fragments of these non-cancer training samples.

39. The method according to claim 38, wherein The methylation features of the non-cancer training samples are obtained from a reference database.

40. The method according to any one of claims 37 to 39, characterized in that, The cancer classifier is trained as a machine learning model.

41. The method according to any one of claims 37 to 40, characterized in that, The methylation feature is based on the identification of one or more methylation variations at each of a plurality of genomic regions.

42. A method for identifying the source tissue type of a DNA fragment, characterized in that, The method includes: Obtaining a methylation sequence read of the DNA fragment, the methylation sequence read including methylation features on one or more genomic regions; Predicting the likelihood that the methylation features of the methylation sequence read originate from each of a plurality of source tissue types by applying each of a plurality of tissue type models to the methylation features, wherein the plurality of tissue type models are optionally part of a hybrid model trained according to the method of any one of claims 1 to 23; Determining the source tissue type from which the DNA fragment most likely originated; and Returning a report of the determination.

43. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores instructions that, when executed by a computer processor, cause the computer processor to perform the method of any one of claims 1 to 42.

44. A system, characterized in that, The system includes: a computer processor; and a non-transitory computer-readable storage medium as claimed in claim 43.

45. A computer program product, characterized in that, The computer program product comprises a non-transitory computer-readable storage medium storing a machine learning hybrid model for predicting the tissue type proportions from a test sample, wherein the product is manufactured by the method according to any one of claims 1 to 23.

46. A computer program product, characterized in that, The computer program product comprises a non-transitory computer-readable storage medium storing a machine learning cancer classifier for predicting cancer in a test sample, wherein the product is manufactured by the method according to any one of claims 33 to 41.

47. A treatment kit, characterized in that, The therapeutic kit comprises: A collection container for collecting a DNA sample from a subject; Optionally, one or more reagents for separating DNA fragments in the DNA sample; Optionally, one or more probes targeting one or more loci determined to indicate a cancer status; And The non-transitory computer-readable storage medium according to claim 43 or the computer program product according to claim 45 or 46.

Images