Cancer classification by tissue of origin thresholding

Abstract

To provide methods for predicting the presence or absence of cancer.SOLUTION: Provided is a method for predicting the presence or absence of cancer in a test sample, comprising: accessing a test sample having a cancer score and a tissue signal for a first tissue label; selecting one of a plurality of layers based on the tissue signal for the first tissue label, the plurality of layers including a high signal layer for the first tissue label and a low signal layer for the first tissue label; and predicting whether the test sample is associated with the presence or absence of cancer by comparing the cancer score to a binary threshold cutoff for the selected layer.SELECTED DRAWING: Figure 15

Description

[Technical field]

[0001] The present invention relates to cancer classification by tissue of origin thresholding. [Background technology]

[0002] Deoxyribonucleic acid (DNA) methylation plays an important role in the control of gene expression. Aberrant DNA methylation has been implicated in numerous disease processes, including cancer. DNA methylation profiling using methylation sequencing (e.g., whole genome bisulfite sequencing (WGBS)) is increasingly recognized as a valuable diagnostic tool for the detection, diagnosis, and / or monitoring of cancer. For example, specific patterns of differentially methylated regions and / or locus-specific methylation patterns are useful as molecular markers for non-invasive diagnosis using circulating cell-free (cf)DNA. However, there remains a need in the art for improved methods for analyzing methylation sequencing data obtained from cell-free DNA for the detection, diagnosis, and / or monitoring of disease, e.g., cancer. Summary of the Invention [Means for solving the problem]

[0003] Early detection of disease states (e.g., cancer) in subjects is important because it allows earlier treatment and therefore higher chances of survival. Sequencing of DNA fragments in cell-free (cf)DNA samples can be used to identify features that can be used for disease classification. For example, in cancer assessment, features based on cell-free DNA obtained from blood samples (e.g., the presence or absence of somatic variants, methylation status, or other genetic abnormalities) can provide insight into whether a subject may have cancer, and further insight into what type of cancer the subject has. To that end, the present specification includes systems and methods for analyzing cell-free DNA sequencing data to determine the likelihood that a subject has disease.

[0004] The analytical system processes multiple sequencing data obtained from multiple samples (e.g., multiple cancer and non-cancer samples) to identify features that are then utilized for cancer classification. Using the sequencing data, the analytical system can train and deploy a cancer classifier to generate a cancer prediction for the test sample.

[0005] Regarding which training samples to use to train the cancer classifier, the analysis uses training samples that have already been identified and labeled as having one or more cancer types, as well as training samples from healthy individuals that have been labeled as non-cancer. Each training sample includes a set of fragments. For each training sample, the analysis system generates a feature vector, for example, by assigning a score to each of the identified features. The analysis system may group the training samples into one or more sets of training samples for iterative training of the cancer classifier. The analysis system inputs each set of feature vectors into the cancer classifier and adjusts classification parameters in the cancer classifier such that a function of the cancer classifier calculates a cancer prediction that accurately predicts the label of the training samples in the set based on the feature vectors and the classification parameters. After repeating the above steps for each set of training samples, the cancer classifier is fully trained.

[0006] During the development, the analysis system generates a feature vector for the test samples in a similar manner as the training samples, e.g., by assigning a score to each of the features in the feature vector for each of the test samples. The analysis system then inputs the feature vector of the test samples into a cancer classifier, which returns a cancer prediction. In one embodiment, the cancer classifier may be configured as a binary classifier that returns a cancer prediction of the likelihood of having or not having cancer. In another embodiment, the cancer classifier is configured as a multi-class classifier that returns a cancer prediction with a predictive value for the cancer type being categorized.

[0007] The present disclosure provides methods and systems for detecting cancer and / or determining the tissue of origin of cancer. In some embodiments, the invention comprises a method or system for detecting cancer comprising: receiving sequencing data of a plurality of biological samples containing cfDNA fragments, the biological samples comprising cancer samples and non-cancer samples; classifying, for each non-cancer sample of the plurality of biological samples, the biological sample using a multi-class classifier based on features derived from the sequencing data, the multi-class classifier predicting a probability likelihood for each of a plurality of tissue-of-origin classes, the plurality of tissue-of-origin classes further comprising one or more tissue-of-origin subtype classes; determining, for each subtype class, whether the predicted probability likelihood exceeds a subtype boundary point, the subtype boundary point indicating a threshold of specificity for the subtype class; and determining a threshold cutoff for predicting the presence or absence of cancer, the threshold cutoff being determined based on a distribution of probability scores corresponding to non-cancer samples, the distribution of probability scores excluding probability scores associated with one or more non-cancer samples identified as having a probability likelihood above the subtype boundary point.

[0008] In some embodiments, the distribution of probability scores is generated by a binary classifier trained on training samples derived from cancer and non-cancer samples.

[0009] In some embodiments, the training samples are divided into multiple cross-validation training sets and used to train a binary classifier to detect the presence of cancer, where the binary classifier generates a probability score for each training sample indicating the presence or absence of cancer.

[0010] In some embodiments, the binary classifier is associated with a first threshold cutoff, and determining the threshold cutoff for predicting the presence or absence of cancer comprises modifying the first threshold cutoff based on excluding probability scores associated with one or more non-cancer samples identified as having a probability likelihood above the subtype boundary point.

[0011] In some embodiments, the threshold cutoff comprises applying a desired level of specificity to the distribution of probability scores, where the threshold cutoff comprises a threshold probability score.

[0012] In some embodiments, the method or system includes receiving test sequencing data of a test biological sample containing cfDNA fragments, analyzing the test sequencing data to determine a test probability score for the presence or absence of cancer, determining whether the test probability score is above a threshold cutoff, and predicting the presence of cancer in response to a determination that the test probability score is above the threshold cutoff.

[0013] In some embodiments, the method or system further comprises predicting the absence of cancer in response to determining that the test probability score is not above the threshold cutoff.

[0014] In some embodiments, the method or system further comprises, in response to determining that the test probability score is above a threshold cutoff, evaluating the test sequencing data with respect to the cancer tissue of origin using a multi-class classifier.

[0015] In some embodiments, the multi-class classifier is trained with training samples derived from cancer samples and non-cancer samples.

[0016] In some embodiments, the method or system further comprises determining each subtype boundary point by an iterative optimization process that optimizes the trade-off between clinical specificity and clinical sensitivity of the corresponding tissue of origin subtype class.

[0017] In some embodiments, the tissue-of-origin subtype class comprises a hematological class indicative of one or more hematological conditions, hi some embodiments, each subtype cut-off point for each hematological class is determined based on a measure of clinical aggressiveness of the corresponding hematological condition.

[0018] In some embodiments, the measures of clinical aggressiveness include one or more of early stage of disease progression, survival rate, rate of disease progression, and severity of disease.

[0019] In some embodiments, the hematological class includes an NHL indolent class, a myeloid class, and a circulating lymphoid class. In some embodiments, the hematological class includes at least one of a circulating lymphoid class, an NHL indolent class, an NHL aggressive class, a Hodgkin's lymphoma class, a myeloid class, a plasma cell class, a heme_1 class, and a heme_3 class. In some embodiments, the circulating lymphoid class includes one or more subclasses selected from the group consisting of hairy cell leukemia, low-grade b-cell, lymphoplasmacytic, chronic lymphocytic leukemia (CLL), SLL, b-cell lymphoblastic, and mantle cell. In some embodiments, the NHL indolent class includes one or more subclasses selected from the group consisting of MALT_NMZL and follicular lymphoma. In some embodiments, the NHL aggressive class includes one or more subclasses selected from the group consisting of mature t-cell neoplasm, mediastinal LBCL, high-grade b-cell, and DLBCL. In some embodiments, the myeloid class comprises one or more subclasses selected from the group consisting of polycythemia vera (PV), MDS, CML, and AML. In some embodiments, the plasma cell class comprises one or more subclasses selected from the group consisting of plasma cell neoplasm and plasma cell myeloma.

[0020] In some embodiments, the sequencing data comprises methylation sequencing data generated by methylation sequencing of cfDNA fragments.In some embodiments, the methylation sequencing comprises WGBS.In some embodiments, the methylation sequencing comprises targeted sequencing.In some embodiments, the feature derived from the methylation sequencing data represents methylation pattern, clonal fraction, or growth or turnover rate.

[0021] In some embodiments, the plurality of tissue-of-origin classes comprises one or more solid or liquid cancerous tissue-of-origin selected from the group consisting of breast cancer, uterine cancer, cervical cancer, ovarian cancer, bladder cancer, urothelial carcinoma of the renal pelvis, non-urothelial renal cancer, prostate cancer, anorectal cancer, colorectal cancer, esophageal cancer, gastric cancer, hepatobiliary cancer arising from hepatocytes, hepatobiliary cancer arising from cells other than hepatocytes, pancreatic cancer, upper gastrointestinal squamous cell carcinoma, upper gastrointestinal cancer other than squamous cell, head and neck cancer, lung cancer, lung adenocarcinoma, small cell lung cancer, squamous cell lung cancer and cancer other than adenocarcinoma or small cell lung cancer, neuroendocrine cancer, melanoma, thyroid cancer, sarcoma, multiple myeloma, lymphoma, and leukemia. In some embodiments, the plurality of tissue-of-origin classes comprises a non-cancer class.

[0022] In other aspects, the present disclosure describes methods and systems for detecting and classifying cancer, the methods or systems comprising: receiving sequencing data of a biological sample comprising cfDNA fragments; analyzing the sequencing data using a multi-class classifier based on features derived from the sequencing data, the multi-class classifier predicting a probability likelihood for each of a plurality of tissue-of-origin classes, the plurality of tissue-of-origin classes comprising one or more cancer tissue-of-origin classes and one or more hematological tissue-of-origin subtype classes; and determining a cancer classification based on the probability likelihood predicted by the multi-class classifier, the cancer classification comprising presence or absence of cancer, cancer tissue of origin, or hematological tissue of origin.

[0023] In other embodiments, a method for predicting the presence or absence of cancer in a test sample includes accessing a test sample having a cancer score and a tissue signal for a first tissue label, selecting one of a plurality of tiers based on the tissue signal for the first tissue label, the plurality of tiers including a tier of high signal for the first tissue label and a tier of low signal for the first tissue label, and predicting whether the test sample is associated with the presence or absence of cancer by comparing the cancer score against a binary threshold cutoff for the selected tier.

[0024] In some embodiments, the test sample comprises a test feature vector determined according to methylation sequencing data of the test sample.

[0025] In some embodiments, the cancer score is determined by applying a binary cancer classifier to the test feature vector.

[0026] In some embodiments, the tissue signal is a tissue of origin (TOO) prediction determined by applying a multi-class cancer classifier to the test feature vector.

[0027] In some embodiments, the TOO prediction comprises a prediction value for each of a plurality of tissue labels, each prediction value indicating the likelihood that the test sample corresponds to the cancer type associated with the tissue label.

[0028] In some embodiments, the step of selecting one of the plurality of tiers based on the tissue signal for the first tissue label includes determining whether the tissue signal for the first tissue label is at or above a predicted value threshold, selecting a high signal tier in response to a determination that the tissue signal for the first tissue label is at or above the predicted value threshold, and selecting a low signal tier in response to a determination that the tissue signal for the first tissue label is below the predicted value threshold.

[0029] In some embodiments, the TOO prediction indicates one or more top predictions of one or more tissue labels of the plurality of tissue labels, and the top prediction of a tissue label indicates that the test sample is predicted to have a cancer type associated with the top predicted tissue label.

[0030] In some embodiments, the step of selecting one of the plurality of tiers includes determining whether the first tissue label is top predictive, selecting a high signal tier in response to a determination that the first tissue label is top predictive, and selecting a low signal tier in response to a determination that the first tissue label is not top predictive.

[0031] In some embodiments, the step of selecting one of the plurality of tiers includes determining whether the first tissue label is a second top predictive, and in response to a determination that the first tissue label is the second top predictive, selecting a high signal tier, and in response to a determination that the first tissue label is not the second top predictive, selecting a low signal tier.

[0032] In some embodiments, the plurality of layers includes a moderate signal layer for a moderate tissue signal.

[0033] In some embodiments, the test sample has a tissue signal for a second tissue class, and selecting one of the plurality of layers is further based on the tissue signal for the second tissue label.

[0034] In some embodiments, the binary threshold cutoff for each stratum is determined by obtaining a holdout set of samples, each sample having a cancer score and a tissue signal for the first tissue label; stratifying the holdout set into a plurality of strata based on the tissue signal for the first tissue label of the holdout set of samples; sweeping the domain of the cancer score with the plurality of candidate binary threshold cutoffs by calculating, for each stratum of the plurality of strata, a true positive rate and a false positive rate for each candidate binary threshold cutoff based on the cancer scores of the samples in the stratum, and selecting a binary threshold cutoff from the plurality of candidate binary threshold cutoffs for the stratum based on the false positive budget and the calculated false positive rate for the stratum.

[0035] In another embodiment, a method for detecting and classifying cancer is disclosed, the method comprising: receiving sequencing data of a biological sample comprising cfDNA fragments; applying a multi-class classifier to features derived from the sequencing data, the multi-class classifier predicting a probability likelihood for each of a plurality of hematological tissue of origin subtype classes; and determining a hematological tissue of origin associated with the biological sample based on the probability likelihood predicted by the multi-class classifier. In some embodiments, a system comprising a hardware processor and a non-transitory computer readable storage medium storing executable instructions that, when executed by the hardware processor, cause the processor to perform the steps of the method.

[0036] In other embodiments, the multi-class classifier further predicts a probability likelihood for the non-cancer class.

[0037] In other embodiments, the multi-class classifier is trained with training samples derived from samples with hematological conditions and non-cancer samples. [Brief description of the drawings]

[0038] [Figure 1A] FIG. 1A is a flowchart describing a process for sequencing fragments of cell-free (cf) DNA to obtain a methylation state vector, according to an embodiment. [Figure 1B] FIG. 1B is a diagram of the process of FIG. 1A for sequencing fragments of cell-free (cf) DNA to obtain a methylation state vector, according to an embodiment. [Figure 2A] FIG. 2A is a flow chart illustrating a process for determining aberrantly methylated fragments obtained from a sample, according to an embodiment. [Figure 2B] FIG. 2B is a flow chart illustrating a process for determining aberrantly methylated fragments obtained from a sample, according to an embodiment. [Figure 3A] FIG. 3A is a flowchart illustrating a process for training a cancer classifier, according to an embodiment. [Figure 3B] FIG. 3B is a diagram of an example of generation of feature vectors used to train a cancer classifier, according to an embodiment. [Figure 4A] FIG. 4A is a flow chart of a device for sequencing a nucleic acid sample, according to one embodiment. [Figure 4B] FIG. 4B is a block diagram of an analysis system, according to an embodiment. [Diagram 5] FIG. 5 is a number of graphs illustrating the cancer prediction accuracy rate of a multi-class cancer classifier for various cancer types, according to an example implementation. [Figure 6] FIG. 6 is a number of graphs illustrating the cancer prediction accuracy rate of a multi-class cancer classifier for various cancer types after first using a binary cancer classifier, according to an exemplary implementation. [Figure 7] FIG. 7 is an illustration of a confusion matrix showing the performance of a trained cancer classifier, according to an exemplary implementation. [Figure 8] FIG. 8 is a graph of the cancer type likelihood of non-cancer samples with specificity above 95%. [Figure 9A] FIG. 9A is a graph of hematological subtypes separated according to methylation sequencing data. [Figure 9B] FIG. 9B is a graph of hematological subtypes separated according to methylation sequencing data. [Figure 10A] FIG. 10A is a flow chart illustrating a process for determining a binary threshold cutoff for a binary cancer classification, according to one or more embodiments. [Figure 10B] FIG. 10B is a flow chart describing the process of thresholding TOO labels to determine a binary threshold cutoff for binary cancer classification, according to one or more embodiments. [Figure 11] FIG. 11 is a confusion matrix showing the performance of a tissue of origin classifier trained with additional hematological cancer subtypes. [Figure 12A] FIG. 12A is a graph showing the cancer prediction accuracy rate of the cancer classifier with and without adjusting the threshold cutoff for multiple cancer types across cancer stages. [Figure 12B] FIG. 12B is a graph showing the cancer prediction accuracy rate of the cancer classifier with and without adjusting the threshold cutoff for multiple cancer types across cancer stages. [Figure 13A] FIG. 13A is an illustration of a process for stratifying hematological signals into two layers, according to one or more embodiments. [Figure 13B] FIG. 13B is a diagram of a process for stratifying hematological signals into three layers, according to one or more embodiments. [Figure 13C] FIG. 13C is an illustration of a process for first stratifying the hematological signal and then the colorectal signal, according to one or more embodiments. [Figure 14] FIG. 14 is an illustration of a process for determining a binary threshold cutoff for TOO stratification, according to one or more embodiments. [Figure 15]FIG. 15 is a flow chart describing a process for predicting the presence or absence of cancer for a test sample using a binary threshold cutoff determined by TOO stratification, according to one or more embodiments. [Figure 16A] FIG. 16A is a graph showing the sensitivity of the classifier at the 99.5% specificity level across hematological subtypes. [Figure 16B] FIG. 16B is a graph showing the sensitivity of the classifier at 95% specificity across stages of Hodgkin's and non-Hodgkin's lymphoma. [Figure 17] FIG. 17 is a confusion matrix illustrating the cancer prediction accuracy rate of the hematology-specific cancer classifier in an exemplary implementation. [Figure 18] FIG. 18 is a series of graphs plotting cancer scores versus distance from the centroid in a UMAP embedding of a hematology-specific cancer classification in a first exemplary implementation. [Figure 19] FIG. 19 is a graph plotting anomaly scores of multiple training samples for hematology-specific cancer classification in a second exemplary implementation. [Figure 20] FIG. 20 is a graph showing the sensitivity of the hematology-specific cancer classifier at 99.5% specificity in the second exemplary implementation. [Figure 21] FIG. 21 is a diagram of a confusion matrix illustrating the cancer prediction accuracy rate of the hematology-specific cancer classifier in the second exemplary implementation. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0039] The figures depict various embodiments for illustrative purposes only. Those skilled in the art will readily recognize from the following description that alternative embodiments of the structures and methods illustrated herein may be used without departing from the principles described herein.

[0040] I. Overview Overview of IA methylation According to the present specification, cfDNA fragments obtained from an individual are processed, for example, by converting unmethylated cytosines to uracils, sequenced, and the sequence reads are compared to a reference genome to identify the methylation status at specific CpG sites in the DNA fragments. Each CpG site may be methylated or unmethylated. Identification of abnormally methylated fragments compared to healthy individuals can provide insight into the cancer status of the subject. As is well known in the art, DNA methylation abnormalities (compared to healthy controls) cause various effects that contribute to cancer. Identification of abnormally methylated cfDNA fragments presents various challenges. First, determining that a DNA fragment is abnormally methylated only carries weight in terms of comparison with a group of control individuals, and as a result, when the control group is small, the determination loses reliability due to statistical variability within the small size of the control group. In addition, methylation status varies among groups of control individuals, which is difficult to take into account when determining that a DNA fragment of a subject is abnormally methylated. On a separate note, cytosine methylation at a CpG site necessarily affects methylation at subsequent CpG sites, and unpacking this dependency is another challenge in itself.

[0041] 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 to form 5-methylcytosine. Specifically, methylation tends to occur at the dinucleotides of cytosine and guanine, referred to herein as "CpG sites." In other cases, methylation occurs at cytosine that is not part of a CpG site or at another nucleotide that is not a cytosine, but these occur rarely. For the purposes of clarity in this disclosure, methylation is discussed with reference to CpG sites. Aberrant DNA methylation can be identified as hypermethylation or hypomethylation, either of which may be indicative of a cancer status. Throughout this disclosure, hypermethylation and hypomethylation are characterized for a DNA fragment if the DNA fragment contains a number of CpG sites above a threshold, and a percentage of the CpG sites above a threshold are either methylated or unmethylated.

[0042] Those skilled in the art will understand that the principles described herein are equally applicable to the detection of methylation in non-CpG contexts, including non-cytosine methylation. In such embodiments, the wet lab assays used to detect methylation may vary from those described herein. Furthermore, the methylation state vectors discussed herein may include elements that are typically sites where methylation occurs or does not occur (even if these sites are not specifically CpG sites). With that substitution, the rest of the process described herein remains the same, and as a result, the inventive concepts described herein are applicable to other forms of methylation.

[0043] IB definition The term "individual" refers to a human individual. The term "healthy individual" refers to an individual who is presumed to be cancer or disease-free. The term "subject" refers to an individual who is known to have or is likely to have cancer or disease.

[0044] The term "cell-free nucleic acid" or "cfNA" refers to nucleic acid fragments that circulate in an individual's body (e.g., in blood) and originate from one or more healthy cells and / or one or more cancer cells. The term "cell-free DNA" or "cfDNA" refers to deoxyribonucleic acid fragments that circulate in an individual's body (e.g., in blood). In addition, cfNA or cfDNA in an individual's body may be derived from other non-human sources.

[0045] The terms "genomic nucleic acid," "genomic DNA," or "gDNA" refer to a nucleic acid molecule or deoxyribonucleic acid molecule 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 is extracted from cells derived from a blood cell lineage, e.g., white blood cells.

[0046] The term "circulating tumor DNA" or "ctDNA" refers to nucleic acid fragments that originate from tumor cells or other types of cancer cells and are released into an individual's bodily fluids (e.g., blood, sweat, urine, or saliva) as a result of biological processes, such as apoptosis or necrosis of dying cells, or are actively released by viable tumor cells.

[0047] The terms "DNA fragment," "fragment," or "DNA molecule" may generally refer to any deoxyribonucleic acid fragment, i.e., cfDNA, gDNA, ctDNA, etc.

[0048] The term "sequence read" refers to a nucleotide sequence obtained from a nucleic acid molecule derived from a test sample obtained from an individual. Sequence reads can be obtained by various methods known in the art.

[0049] The term "sequencing depth" or "depth" refers to the total number of sequence reads or read segments at a given genomic location or locus from a test sample obtained from an individual.

[0050] The terms "aberrant fragment," "aberrantly methylated fragment," or "fragment with an aberrant methylation pattern" refer to a fragment with aberrant methylation of a CpG site. The aberrant methylation of a fragment can be determined using a probabilistic model that identifies the surprise of observing the methylation pattern of the fragment in a control group.

[0051] The term "abnormal fragments with extreme methylation" or "UFXM" refers to hypomethylated or hypermethylated fragments. Hypomethylated and hypermethylated fragments refer to fragments that have at least some number of CpG sites (e.g., 5) that are methylated or unmethylated above some threshold percentage (e.g., 90%), respectively.

[0052] The term "abnormality score" refers to the score of a CpG site based on the number of abnormal fragments (or in some embodiments, UFXMs) from the sample that overlap that CpG site. The abnormality score is used in the context of characterizing a sample for classification.

[0053] II. Sample Processing II.A. Generation of methylation status vectors of DNA fragments FIG. 1A is a flow chart describing a process 100 for sequencing fragments of cell-free (cf)DNA to obtain a methylation state vector, according to an embodiment. To analyze DNA methylation, the analysis system first obtains a sample 110 from an individual that contains a plurality of cfDNA molecules. In general, the sample may be from a healthy individual, a subject known to have or suspected to have cancer, or a subject with no prior information known condition. The test sample may be a sample selected from the group consisting of blood, plasma, serum, urine, feces, and saliva samples. Alternatively, the test sample may include a sample selected from the group consisting of whole blood, blood fractions (e.g., white blood cells (WBC)), tissue biopsy, pleural fluid, pericardial fluid, cerebrospinal fluid, and peritoneal fluid. In additional embodiments, the process 100 may be applied to sequence other types of DNA molecules.

[0054] From the sample, the analysis system isolates each cfDNA molecule. The cfDNA molecule is treated 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, a commercially available kit such as EZ DNA Methylation™-Gold, EZ DNA Methylation™-Direct, or EZ DNA Methylation™-Lightning kit (available from Zymo Research Corp, Irvine, CA) is used for bisulfite conversion. In another embodiment, the conversion of unmethylated cytosine to uracil is achieved using an enzymatic reaction. For example, the conversion may use a commercially available kit for converting unmethylated cytosine to uracil, such as APOBEC-Seq (NEBiolabs, Ipswich, MA).

[0055] A sequencing library is prepared from the converted cfDNA molecules 130. Optionally, multiple hybridization probes may be used to enrich the sequencing library for cfDNA molecules or genomic regions that are informative regarding cancer status 135. Hybridization probes are short oligonucleotides that can hybridize to specifically designated cfDNA molecules or targeted regions and enrich those fragments or regions for subsequent sequencing and analysis. Hybridization probes can be used to perform targeted deep analysis of a set of designated CpG sites of interest to researchers. In one embodiment, the hybridization probes are designed to enrich DNA molecules that have been treated (e.g., using bisulfite) for conversion of unmethylated cytosines to uracils. Once prepared, the sequencing library or a portion thereof can be sequenced to obtain multiple sequence reads. The sequence reads may be in a computer-readable digital format for processing and interpretation by computer software.

[0056] From the sequence reads, the analysis system determines the location and methylation state of each CpG site based on the alignment to the reference genome 150. For each fragment, the analysis system generates a methylation state vector 160 that specifies the location of the fragment in the reference genome (e.g., specified by the location of the first CpG site in each fragment or another similar metric), the number of CpG sites in the fragment, and the methylation state of each CpG site in the fragment, whether methylated (e.g., denoted as M), unmethylated (e.g., denoted as U), or indeterminate (e.g., denoted as I). Observed states are methylated and unmethylated states, while unobserved states are indeterminate. Indeterminate methylation states may originate from sequencing errors and / or mismatches between methylation states of complementary strands of DNA fragments. The methylation state vector is stored in a transient or persistent computer memory for later use and processing. Additionally, the analysis system may remove duplicate reads or duplicate methylation state vectors that originate from a single sample. The analysis system may determine that certain fragments having one or more CpG sites have a number or percentage of uncertain methylation status above a threshold and may exclude such fragments or may construct a model that selectively includes such fragments but takes into account such uncertain methylation status, one such model is described below in conjunction with FIG. 4.

[0057] FIG. 1B is an illustration of the process 100 of FIG. 1A for sequencing a cfDNA molecule to obtain a methylation state vector, according to an embodiment. As an example, the analysis system receives a cfDNA molecule 112, which in this example includes three CpG sites. As shown, the first and third CpG sites of the cfDNA molecule 112 are methylated 114. During a processing step 120, the cfDNA molecule 112 is converted to generate a converted cfDNA molecule 122. During processing 120, the second CpG site, which was not methylated, has its cytosine converted to uracil. However, the first and third CpG sites were not converted.

[0058] After conversion, a sequencing library 130 is prepared and sequenced 140 to generate sequence reads 142. The analysis system aligns 150 the sequence reads 142 to a reference genome 144. The reference genome 144 provides context as to which location in the human genome the cfDNA fragments originate from. In this simplified example, the analysis system aligns 150 the sequence reads 142 such that three CpG sites correlate to CpG sites 23, 24, and 25 (arbitrary reference identifiers used for ease of illustration). The analysis system thus generates information regarding both the methylation status of all CpG sites on the cfDNA molecule 112 and the location in the human genome to which the CpG sites map. As shown, the CpG sites on the sequence reads 142 that are methylated are read as cytosines. In this example, cytosine appears at the first and third CpG sites only in sequence read 142, which can be interpreted as indicating that the first and third CpG sites in the original cfDNA molecule were methylated. Meanwhile, the second CpG site is read as a thymine (U is converted to T during the sequencing process), which can be interpreted as indicating that the second CpG site was not methylated in the original cfDNA molecule. Using these two pieces of information, methylation status and location, the analysis system generates 160 a methylation state vector 152 for the cfDNA fragment 112. In this example, the resulting methylation state vector 152 is: <M 23 , U 24 , M 25 >, where M corresponds to a methylated CpG site, U corresponds to an unmethylated CpG site, and the subscript numbers correspond to the position of the respective CpG site in the reference genome.

[0059] II.B. Identification of Anomalous Fragments The analysis system uses the methylation state vector of the sample to determine aberrant fragments of the sample. For each fragment in the sample, the analysis system uses the methylation state vector corresponding to the fragment to determine whether the fragment is an aberrant fragment. In one embodiment, the analysis system calculates a p-value score for each methylation state vector that describes the probability of observing that methylation state vector or other less likely methylation state vectors in a healthy control group. The process for calculating the p-value score is further discussed below in section II.Bi P-value filtering. The analysis system may determine as aberrant fragments those fragments that have a methylation state vector below a threshold p-value score. In another embodiment, the analysis system further labels as hypermethylated and hypomethylated fragments those fragments that have at least some number of CpG sites with methylation or unmethylation above some threshold percentage, respectively. Hypermethylated or hypomethylated fragments are also referred to as aberrant fragments with extreme methylation (UFXM). In other embodiments, the analysis system may implement various other probabilistic models for determining aberrant fragments. Examples of other probabilistic models include mixture models, deep probabilistic models, etc. In some embodiments, the analysis system may use any combination of the processes described below to identify aberrant fragments. With the identified aberrant fragments, the analysis system may filter a set of methylation state vectors of samples for use in other processes, for example for use in training and developing a cancer classifier.

[0060] II.BI P-value filtering In one embodiment, the analysis system calculates a p-value score for each methylation state vector in comparison with the methylation state vector obtained from the fragments in the healthy control group. The p-value score describes the probability of observing a methylation status in the healthy control group that matches the methylation state vector or other less likely methylation state vectors. To determine that a DNA fragment is abnormally methylated, the analysis system uses a healthy control group in which the majority of fragments are normally methylated. When performing this probabilistic analysis to determine abnormal fragments, the decision has weight in terms of comparison with a group of control subjects that constitute the healthy control group. To ensure robustness in the healthy control group, the analysis system may select some threshold number of healthy individuals to provide samples containing the DNA fragments. Figure 2A below describes how to generate a data structure for a healthy control group with which the analysis system calculates a p-value score. Figure 2B describes how to calculate a p-value score using the generated data structure.

[0061] 2A is a flow chart illustrating a process 200 for generating a healthy control data structure, according to an embodiment. To create the healthy control data structure, an analysis system receives a plurality of DNA fragments (e.g., cfDNA) from a plurality of healthy individuals. A methylation state vector is determined for each fragment, for example, by process 100.

[0062] Using the methylation state vector of each fragment, the analysis system subdivides 205 the methylation state vector into strings of CpG sites. In one embodiment, the analysis system subdivides 205 the methylation state vector such that the resulting strings are all shorter than a given length. For example, a methylation state vector of length 11 would be subdivided into strings of lengths 3 or less, resulting in 9 strings of length 3, 10 strings of length 2, and 11 strings of length 1. In another example, a methylation state vector of length 7 would be subdivided into strings of lengths 4 or less, resulting in 4 strings of length 4, 5 strings of length 3, 6 strings of length 2, and 7 strings of length 1. If the methylation state vector is shorter than or equal to the specified string length, the methylation state vector is converted into a single string that contains all of the CpG sites of the vector.

[0063] For each possible CpG site and methylation state possibility in the vector, the analysis system tallies the strings by counting the number of strings present in the control set that have the specified CpG site as the first CpG site in the string and have that methylation state possibility 210. For example, at a given CpG site, considering a string of length 3, there are 2^3 or 8 possible string configurations. At that given CpG site, for each of the 8 possible string configurations, the analysis system tallies the number of occurrences of each methylation state vector possibility in the control set 210. Continuing with the example, this may include tallying the following quantities: for each starting CpG site x in the reference genome, <M x , M x+1 , M x+2 >, <M x , M x+1 , U x+2 >,..., x , U x+1 , U x+2 ​The analysis system creates a data structure 215 that stores the tabulated counts for each start CpG site and possible string.

[0064] Setting an upper limit on the string length has several advantages. First, depending on the maximum string length, the size of the data structure created by the analysis system can increase dramatically in size. For example, a maximum string length of 4 means that each CpG site will count at least 2^4 for strings of length 4. Increasing the maximum string length to 5 means that every CpG site will count an additional 2^4 or 16, doubling the number of counts (and computer memory required) compared to previous string lengths. Reducing the string size helps to keep the creation and performance of the data structure reasonable in terms of computation and storage (e.g., for later access as described below). Second, a statistical consideration for limiting the maximum string length is to avoid overfitting of downstream models that use string counts. If long CpG site strings do not have a strong biological impact on the outcome (e.g., prediction of abnormality, which is predictive of the presence of cancer), calculating probabilities based on large CpG site strings is problematic because it requires a huge amount of data that may not be available and therefore becomes too sparse for the model to function properly. For example, to calculate the probability of abnormality / cancer conditioned on the previous 100 CpG sites, ideally, one would need counts of strings in a data structure of length 100, some of which exactly match the previous 100 methylation states. If only a sparse count of strings of length 100 is available, there would be insufficient data to determine whether a given string of length 100 is abnormal or not in a test sample.

[0065] 2B is a flow chart illustrating a process 220 for identifying aberrantly methylated fragments from an individual, according to an embodiment. In process 220, the analysis system generates methylation state vectors from the subject's cfDNA fragments 100. The analysis system treats each methylation state vector as follows:

[0066] For a given methylation state vector, the analysis system enumerates all possibilities for methylation state vectors that have the same starting CpG site and the same length (i.e., set of CpG sites) within the methylation state vector 230. Since each methylation state is generally either methylated or unmethylated, at each CpG site there are effectively two possible states, and thus the count of distinct possibilities for a methylation state vector depends on powers of 2, so that a methylation state vector of length n is the sum of the 2 possible states of the methylation state vector. n When using a methylation state vector that includes an uncertain state for one or more CpG sites, the analysis system may enumerate the possibilities of the methylation state vector considering only CpG sites that have an observed state 230.

[0067] The analysis system calculates 240 the probability of observing each possible methylation state vector for the identified starting CpG site and length of the methylation state vector by accessing a data structure of the healthy control group. In one embodiment, the calculation of the probability of observing a given possibility uses Markov chain probability modeling joint probability calculations. In other embodiments, calculation methods other than Markov chain probability are used to determine the probability of observing each possible methylation state vector.

[0068] The analysis system uses the calculated probability for each possibility to calculate 250 a p-value score for the methylation state vector. In one embodiment, this involves identifying a calculated probability that corresponds to the possibility of matching the methylation state vector in question. Specifically, this is the possibility of having the same set of CpG sites as the methylation state vector, or similarly the same starting CpG site and length. The analysis system sums the calculated probabilities of any possibilities that have a probability lower than or equal to the identified probabilities to generate a p-value score.

[0069] This p-value represents the probability of observing the methylation state vector of the fragment in a healthy control group or other methylation state vectors that are less likely to occur. A low p-value score therefore generally corresponds to a methylation state vector that is rare in healthy individuals, which causes the fragment to be labeled as abnormally methylated compared to the healthy control group. A high p-value score generally relates to a methylation state vector that is expected to be present in a healthy individual in a relative sense. If the healthy control group is a non-cancerous group, for example, a low p-value indicates that the fragment is abnormally methylated compared to the non-cancerous group, and therefore may indicate the presence of cancer in the test subject.

[0070] As described above, the analysis system calculates a p-value score for each of a plurality of methylation state vectors, each of which represents a cfDNA fragment in the test sample. To identify which of the fragments are abnormally methylated, the analysis system may filter the set of methylation state vectors based on their p-value scores 260. In one embodiment, the filtering is performed by comparing the p-value scores against a threshold value and keeping only the fragments that are below the threshold value. This threshold p-value score may be about 0.1, 0.01, 0.001, 0.0001, or the like.

[0071] Exemplary results from process 400 show that the analysis system yields a median (range) of 2,800 (1,500-12,000) fragments with aberrant methylation patterns for participants without cancer in the training and a median (range) of 3,000 (1,200-220,000) fragments with aberrant methylation patterns for participants with cancer in the training. These filtered sets of fragments with aberrant methylation patterns may be used for downstream analysis, as described in Section III below.

[0072] In one embodiment, the analysis system uses a sliding window to determine the probabilities of the methylation state vectors and to calculate p-values ​​255. Rather than enumerating possibilities and calculating p-values ​​for every methylation state vector, the analysis system enumerates possibilities and calculates p-values ​​only for a window of contiguous CpG sites, where the window is shorter in length (of CpG sites) than at least some of the fragments (otherwise the window would serve no purpose). The length of the window may be static, user-determined, dynamic, or otherwise selected.

[0073] When calculating p-values ​​for a methylation state vector larger than a window, the window identifies a set of contiguous CpG sites from the vector within the window starting from the first CpG site in the vector. The analysis system calculates a p-value score for the window that includes the first CpG site. The analysis system then "slides" the window to a second CpG site in the vector and calculates another p-value score for the second window. Thus, for a window size l and a methylation vector length m, each methylation state vector will generate m-l+1 p-value scores. After completing the p-value calculations for each portion of the vector, the lowest p-value score from all the sliding windows is taken as the overall p-value score for the methylation state vector. In another embodiment, the analysis system agglomerates the p-value scores for the methylation state vector to generate an overall p-value score.

[0074] The use of sliding windows helps to reduce the number of enumerated possibilities of the methylation state vector and their corresponding probability calculations that would otherwise have to be made. To take a realistic example, it is possible for a fragment to have more than 54 CpG sites. Instead of calculating probabilities for 2^54 (approximately 1.8×10^16) possibilities to generate a single p-score, the analysis system can use a window of size 5 (for example), resulting in 50 p-value calculations for each of the 50 windows of the methylation state vector of the fragment. Each of the 50 calculations enumerates 2^5 (32) possibilities of the methylation state vector, which in total results in 50×2^5 (1.6×10^3) ​​probability calculations. This results in a significant reduction in the calculations made, without a significant hit to the accurate identification of abnormal fragments.

[0075] In embodiments with uncertain states, the analysis system may calculate a p-value score by summing over the CpG sites with uncertain states in the methylation state vector of the fragment. The analysis system identifies all possibilities that are in consensus with all methylation states in the methylation state vector excluding the uncertain states. The analysis system may assign a probability to the methylation state vector as the sum of the probabilities of the identified possibilities. As an example, the analysis system may determine that the methylation state observed for CpG sites 1 and 3 is in consensus with the methylation states at CpG sites 1 and 3 of the fragment,<M1, I2, U3> The probability of the methylation state vector of<M1, M2, U3> and<M1, U2, U3> The probability of a methylation state vector having one or more uncertain states is calculated as the sum of the probabilities of the possibilities of the methylation state vector of CpG sites having one or more uncertain states. This method of summing CpG sites having an uncertain state uses a calculation of the probability of up to 2^i possibilities, where i refers to the number of uncertain states in the methylation state vector. In an additional embodiment, a dynamic programming algorithm may be implemented to calculate the probability of a methylation state vector having one or more uncertain states. Advantageously, the dynamic programming algorithm operates in linear computation time.

[0076] In one embodiment, the computational burden of calculating the probability and / or p-value scores may be further reduced by caching at least some of the calculations. For example, the analysis system may cache in transient or persistent memory the calculations of the probability of a methylation state vector (or a window thereof). If other fragments have the same CpG sites, caching the probability of the possibility allows efficient calculation of the p-score value without requiring recalculation of the underlying probability of the possibility. Equivalently, the analysis system may calculate a p-value score for each of the possibilities of the methylation state vector associated with a set of CpG sites from the vector (or a window thereof). The analysis system may cache the p-value score for use in determining the p-value score of other fragments containing the same CpG sites. In general, the p-value score of a possibility of a methylation state vector with the same CpG sites may be used to determine the p-value score of another of the possibilities derived from the same set of CpG sites.

[0077] II.B.II. Hypermethylated and hypomethylated fragments In another embodiment, the analysis system determines aberrant fragments as those fragments having a number of CpG sites above a threshold, with either a percentage of CpG sites above a threshold being methylated or a percentage of CpG sites above a threshold being unmethylated, and the analysis system identifies such fragments as hypermethylated or hypomethylated. Exemplary threshold fragment (or CpG site) length values ​​include greater than 3, greater than 4, greater than 5, greater than 6, greater than 7, greater than 8, greater than 9, greater than 10, etc. Exemplary percentage thresholds for methylation or unmethylation include greater than 80%, greater than 85%, greater than 90%, or greater than 95%, or any other percentage within the range of 50%-100%.

[0078] II.C. Exemplary Analytical System 4A is a flow chart of a device for sequencing a nucleic acid sample, according to one embodiment. This exemplary flow chart includes devices such as a sequencer 420 and an analysis system 400. The sequencer 420 and the analysis system 400 may operate in tandem to perform one or more steps in the processes of 100 of FIG. 1A, 200 of FIG. 2A, 220 of FIG. 2B, and other processes described herein.

[0079] In various embodiments, the sequencer 420 receives the enriched nucleic acid sample 410. As shown in FIG. 4A, the sequencer 420 may include a graphical user interface 425 that allows user interaction with specific tasks (e.g., start sequencing or end sequencing), as well as one further loading station 430 for loading a sequencing cartridge containing the enriched fragment sample and / or loading buffers required to perform a sequencing assay. Thus, once a user of the sequencer 420 provides the necessary reagents and sequencing cartridge to the loading station 430 of the sequencer 420, the user can start sequencing by interacting with the graphical user interface 425 of the sequencer 420. Once started, the sequencer 420 performs sequencing and outputs sequence reads of the enriched fragments derived from the nucleic acid sample 410.

[0080] In some embodiments, the sequencer 420 is communicatively coupled to the analysis system 400. The analysis system 400 includes several computing devices used to process sequence reads for various applications, such as evaluating methylation status at one or more CpG sites, variant calling, or quality control. The sequencer 420 may provide the sequence reads to the analysis system 400 in a BAM file format. The analysis system 400 may be communicatively coupled to the sequencer 420 through wireless, wired, or a combination of wireless and wired communication techniques. In general, the analysis system 400 is configured with a processor and a non-transitory computer-readable storage medium that stores computer instructions that, when executed by the processor, cause the processor to process sequence reads or perform one or more steps of any of the methods or processes disclosed herein.

[0081] In some embodiments, the sequence reads may be aligned to a reference genome using methods known in the art to determine alignment position information, for example, by step 140 of process 100 in FIG. 1A. The alignment position may generally describe 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. Corresponding to methylation sequencing, the alignment position information may be generalized to indicate the first and last CpG sites included in the sequence read according to the alignment to the reference genome. The alignment position information may further indicate the methylation status and positions of all CpG sites in a given sequence read. Regions in the reference genome may be associated with genes or segments of genes, so that the analysis system 400 may label the sequence read with one or more genes that align to the sequence read. In one embodiment, the length (or size) of the fragment is determined from the start and end positions.

[0082] In various embodiments, for example, when a paired-end sequencing process is used, the sequence read is composed of a read pair, denoted as R_1 and R_2. For example, the first read R_1 may be sequenced from a first end of a double-stranded DNA (dsDNA) molecule, while the second read R_2 may be sequenced from a second end of the double-stranded DNA (dsDNA). Thus, the nucleotide base pairs of the first read R_1 and the second read R_2 may be aligned consistently (e.g., in reverse) with the nucleotide bases of the reference genome. The alignment position information derived from the read pair R_1 and R_2 may include a start position in the reference genome corresponding to the end of the first read (e.g., R_1) and an end position in the reference genome corresponding to the end of the second read (e.g., R_2). In other words, the start position and the end position in the reference genome represent the possible positions in the reference genome to which the nucleic acid fragment corresponds. An output file having a SAM (sequence alignment map) format or a BAM (binary) format may be generated and output for further analysis.

[0083] Reference is now made to Figure 4B, which is a block diagram of an analytical system 400 for processing DNA samples, according to one embodiment. The analytical system implements one or more computing devices for use in analyzing DNA samples. The analytical system 400 includes a sequence processor 440, a sequence database 445, a model database 455, a model 450, a parameter database 465, and a score engine 460. In some embodiments, the analytical system 400 performs some or all of the processes 100 of Figure 1A and 200 of Figure 2.

[0084] The sequence processor 440 generates methylation state vectors for fragments obtained from the sample. For each CpG site on the fragment, the sequence processor 440 generates a methylation state vector for each fragment, by process 100 of FIG. 1A, that specifies the location of the fragment in the reference genome, the number of CpG sites in the fragment, and the methylation state of each CpG site in the fragment, which is methylated, unmethylated, or uncertain. The sequence processor 440 can store the methylation state vectors of the fragments in a sequence database 445. The data in the sequence database 445 may be organized such that the methylation state vectors obtained from the samples are related to each other.

[0085] Additionally, multiple different models 450 may be stored in a model database 455 and retrieved for use with a test sample. In one example, the model is a cancer classifier trained to determine a cancer prognosis for a test sample using feature vectors derived from abnormal fragments. The training and use of cancer classifiers is further discussed in conjunction with Section III. Cancer Classifiers for Determining Cancer. Analysis system 400 may train one or more models 450 and store various trained parameters in a parameter database 465. Analysis system 400 stores models 450 along with features in model database 455.

[0086] During estimation, the score engine 460 uses one or more models 450 and returns results. The score engine 460 accesses the models 450 in the model database 455 along with trained parameters from the parameter database 465. Depending on each model, the score engine receives inputs appropriate for that model and calculates outputs based on the received inputs, parameters, and functions of the respective model related to the inputs and outputs. In some use cases, the score engine 460 also calculates metrics that correlate to confidence in the outputs calculated from the models. In other use cases, the score engine 460 calculates other intermediate values ​​for use in the models.

[0087] III. Cancer classifier for determining cancer III.A. Overview The cancer classifier is trained to receive a feature vector of a test sample and determine whether the test sample is derived from a test subject having cancer, or more specifically, a particular cancer type. The cancer classifier includes a function that represents an association between a feature vector as an input and a cancer prediction as an output determined by the function operating on the input feature vector with the classification parameters. In one embodiment, the feature vector input to the cancer classifier is based on a set of abnormal fragments determined from the test sample. The abnormal fragments may be determined by process 220 in FIG. 2B, or more specifically, hypermethylated and hypomethylated fragments determined by step 270 of process 220, or abnormal fragments determined according to some other process. Prior to deployment of the cancer classifier, the analysis system trains the cancer classifier in process 300.

[0088] III.B. Training a Cancer Classifier 3A is a flow chart illustrating a process 300 for training a cancer classifier, according to an embodiment. The analysis system obtains 310 a plurality of training samples, each having a set of abnormal fragments and a cancer type label. The plurality of training samples includes any combination of samples obtained from healthy individuals with a general label of "non-cancer", samples obtained from subjects with a general or specific label of "cancer" (e.g., "breast cancer", "lung cancer", etc.). Training samples obtained from subjects for one cancer type may be referred to as a cohort for that cancer type or a cancer type cohort.

[0089] The analysis system determines 320 a feature vector for each training sample based on the set of abnormal fragments of the training sample. The analysis system calculates an anomaly score for each CpG site in the initial set of CpG sites. The initial set of CpG sites may be all CpG sites in the human genome or a portion thereof, which may be about 10 4 , 10 5 , 10 6 , 10 7 , 10 8 In one embodiment, the analysis system defines the anomaly score for the feature vector using a binary scoring based on the presence or absence of an aberrant fragment that includes a CpG site in the set of aberrant fragments. In another embodiment, the analysis system defines the anomaly score based on the count of aberrant fragments that overlap the CpG site. In one example, the analysis system may use a ternary scoring that assigns a first score for the absence of the presence of an aberrant fragment, a second score for the presence of several aberrant fragments, and a third score for the presence of more than several aberrant fragments. For example, the analysis system counts five aberrant fragments that overlap the CpG site in a sample and calculates the anomaly score based on the count of five.

[0090] Once all anomaly scores have been determined for the training sample, the analysis system determines a feature vector as a vector of elements, each element including one of the anomaly scores associated with one of the CpG sites in the initial set. The analysis system normalizes the anomaly scores of the feature vector based on the coverage of the sample, where coverage refers to the median or average sequencing depth across all CpG sites covered by the initial set of CpG sites used in the classifier, or based on the set of anomaly fragments of a given training sample.

[0091] By way of example, refer now to FIG. 3B, which illustrates a matrix for training feature vector 322. In this example, the analysis system identifies CpG sites [K] 326 to consider in generating the feature vector of the cancer classifier. The analysis system selects training sample [N] 324. For a first arbitrary CpG site [k1], the analysis system determines a first anomaly score 328 to use in the feature vector of the training sample [n1]. The analysis system examines each aberrant fragment in the set of aberrant fragments. When the analysis system identifies at least one aberrant fragment that includes the first CpG site, the analysis system determines the first aberration score 328 of the first CpG site to be 1, as shown in FIG. 3B. When considering a second arbitrary CpG site [k2], the analysis system similarly examines the set of aberrant fragments for at least one that includes the second CpG site [k2]. If the analysis system finds no such aberrant fragment containing the second CpG site, the analysis system determines a second aberration score 329 for the second CpG site [k2] to be 0, as shown in Figure 3B. Once the analysis system has determined all of the aberration scores for the initial set of CpG sites, the analysis system determines a feature vector for the first training sample [n1] including anomaly scores using a feature vector including the first aberration score 328 of 1 for the first CpG site [k1] and the second aberration score 329 of 0 for the second CpG site [k2], as well as the subsequent aberration scores, thereby forming a feature vector [1, 0, ...].

[0092] The analysis system may further restrict the CpG sites considered for use in the cancer classifier. For each CpG site in the initial set of CpG sites, the analysis system calculates 330 an information gain based on the feature vectors of the training samples. From step 320, each training sample has a feature vector that includes an anomaly score for all CpG sites in the initial set of CpG sites up to and including all CpG sites in the human genome. However, some CpG sites in the initial set of CpG sites may be less informative than others with respect to distinguishing between cancer types or may overlap with other CpG sites.

[0093] In one embodiment, the analysis system calculates 330 an information gain 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 a training sample with a given cancer type compared to all other samples. For example, two random variables are used: "abnormal fragment" ("AF") and "cancer type" ("CT"). In one embodiment, AF is a binary variable indicating whether there is an abnormal fragment that overlaps a given CpG site in a given sample, as determined with respect to the anomaly score / feature vector described above. CT is a random variable indicating whether the cancer is of a particular type. The analysis system calculates mutual information for CT in light of AF, i.e., the number of information bits about the cancer type that are gained when it is known whether there is an abnormal fragment that overlaps a particular CpG site. In practice, for a first cancer type, the analysis system calculates pairwise mutual information gain for each other cancer type and sums the mutual information gains over all other cancer types.

[0094] For a given cancer type, the analysis system uses this information to rank CpG sites based on how cancer-specific they are. This procedure is repeated for all cancer types under consideration. If a particular region is commonly aberrantly methylated in training samples of a given cancer, but not aberrantly methylated in training samples of other cancer types or in healthy training samples, the CpG sites where those aberrant fragments overlap will tend to have high information gain for the given cancer type. The ranked CpG sites for each cancer type are greedily added (selected) 340 to a set of selected CpG sites based on their ranking for use in the cancer classifier.

[0095] In additional embodiments, the analysis system may consider other selection criteria to select informative CpG sites to be used in the cancer classifier. One selection criterion may be that the selected CpG sites are separated from other selected CpG sites by more than a threshold number of base pairs (e.g., 100 base pairs) from any other selected CpG sites, so that any CpG sites that are separated within the threshold are not selected for consideration in the cancer classifier.

[0096] In one embodiment, the analysis system may optionally modify the feature vectors of the training samples according to the set of CpG sites selected from the initial set 350. For example, the analysis system may trim the feature vectors to remove anomaly scores corresponding to CpG sites that are not included in the set of selected CpG sites.

[0097] Using the feature vector of the training samples, the analysis system can train a cancer classifier in any number of ways. The feature vector may correspond to the initial set of CpG sites obtained from step 320 or the set of selected CpG sites obtained from step 350. In one embodiment, the analysis system trains 360 a binary cancer classifier to distinguish between cancer and non-cancer based on the feature vector of the training samples. In this manner, the analysis system uses training samples that include both non-cancer samples from healthy individuals and cancer samples from the subject. Each training sample has 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.

[0098] In another embodiment, the analysis system trains 450 a multi-class cancer classifier to distinguish between multiple cancer types (also referred to as tissue of origin (TOO) labels). The cancer types include one or more cancers and may include non-cancer types (including any additional other diseases or genetic disorders, etc.). To do so, the analysis system uses a cancer type cohort, which may or may not also 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 includes a predictive value for each of the cancer types being classified. The predictive value may correspond to the likelihood that a given training sample (or during estimation, a test sample) has each of the cancer types. In one implementation, the predictive value is scored from 0 to 100, with the cumulative predictive values ​​adding up to 100. For example, the cancer classifier returns a cancer prediction that includes predictive values ​​for breast cancer, lung cancer, and non-cancer. For example, the classifier may return a cancer prediction that the test sample has a 65% likelihood of breast cancer, a 25% likelihood of lung cancer, and a 10% likelihood of non-cancer. The analysis system may further evaluate the predictive values ​​to generate a prediction of the presence of one or more cancers in the sample, which may also be referred to as a TOO prediction indicating one or more TOO labels, e.g., a first TOO label having the highest predictive value, a second TOO label having the second highest predictive value, etc. Continuing with the above example and with the given percentages, in this example, the system may determine that the sample has breast cancer given that breast cancer has the highest likelihood.

[0099] In either embodiment, the analysis system trains the 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 classifier's function accurately relates the training feature vectors to their corresponding labels. The analysis system may group the training samples into one or more sets of training samples for iterative batch training of the cancer classifier. After inputting all sets of training samples, including their training feature vectors, and adjusting the classification parameters, the cancer classifier is sufficiently trained to label test samples according to their feature vectors within some margin of error. The analysis system may train the cancer classifier according to any one of several methods. As an example, the binary cancer classifier may be an L2 regularized logistic regression classifier trained using a log loss function. As another example, the multi-cancer classifier may be a multinomial logistic regression. In practice, either type of cancer classifier may be trained using other techniques. These techniques are numerous, including the potential use of kernel methods, random forest classifiers, mixture models, autoencoder models, machine learning algorithms such as multi-layer neural networks, and the like.

[0100] III.C. Tuning the Cancer Classifier During the use of the cancer classifier, the analysis system may perform an operation to adjust the predictive ability of the cancer classifier. In general, the sample distribution may include one or more non-cancer samples with high tissue signals. Some of these high tissue signal non-cancer samples may even be pre-stage cancer, early stage cancer, or undiagnosed cancer. Therefore, the non-cancer samples with high tissue signals will disrupt the predictive ability of the cancer classifier.

[0101] As used herein, "high tissue signal" refers to samples that have tissue signals above some threshold for any tissue type or specific cancer type, generally, for example, also referred to as TOO label. Tissue signals may be determined by a multi-class cancer classifier or other approach, compared to a healthy distribution. Non-cancer samples with high tissue signals are outliers in the non-cancer distribution. Some of these non-cancer samples may be pre-stage cancer, early stage cancer, or undiagnosed cancer. The analysis system may identify non-cancer samples that have high tissue signals in at least one TOO label. In one approach to determine high tissue signals, the predicted value of the TOO label output by the multi-class cancer classifier is compared against a tissue signal threshold. Samples with predicted values ​​above the tissue signal threshold are considered to have high tissue signals for that TOO label, while samples with predicted values ​​below the tissue signal threshold are considered to have no high tissue signals (or have low tissue signals) for that TOO label. In another approach, one or more top predictions in the TOO prediction are considered. For example, the TOO prediction of a sample has a first prediction of a colorectal TOO label, a second prediction of a breast TOO label, and a third prediction of a head / neck TOO label. When considering the top predictions, the sample is deemed to have a high tissue signal for the TOO label in the first prediction, which in this example is the colorectal TOO label. When the top two predictions are considered, there is a high tissue signal in both the colorectal TOO label and the breast TOO label. Other approaches to determining tissue signals may include other models trained to determine tissue signals for one or more TOO labels. Such models may include classifiers trained to determine tissue signals for a subset of TOO labels. For example, a hematology-specific classifier may be trained and used to determine tissue signals for one or more hematological subtypes. Two such exemplary implementations are described in section V. Exemplary Results of Cancer Classifiers.Other models include deconvolution models that can deconvolute tissue signals from methylation sequencing data (and / or other types of sequencing data).

[0102] In one embodiment of tuning the classifier, the sensitivity of the classifier in predicting cancer is improved by excluding non-cancer samples with high tissue signals. Using the excluded sample distribution, a binary threshold cutoff can be determined according to a minimum specificity, and this binary threshold cutoff is used to predict the presence or absence of cancer in the test sample. This method is further detailed in the section III.Ci Removal of non-cancer samples with high signals.

[0103] In another embodiment of training the cancer classifier, the sample distribution may be stratified according to the TOO signal. The analysis system determines a binary threshold cutoff for each resulting stratum into which the samples are stratified. Using a test sample, the analysis system places the test sample into a stratum according to the TOO signal and predicts the presence or absence of cancer in the test sample using the binary threshold cutoff of the stratum. This method is further detailed in section III.C.ii. Stratification of sample distribution according to TOO signal.

[0104] Referring to FIG. 8, FIG. 8 illustrates a graph of the cancer type likelihood of non-cancer samples with a specificity of more than 95%. A cancer score was calculated for each non-cancer sample of a plurality of non-cancer samples, i.e., samples obtained from healthy individuals not currently diagnosed with cancer. The cancer score can be determined by a binary classifier as the likelihood that the sample has cancer given the methylation sequencing data of the sample. In other embodiments, the cancer score can be calculated according to other methods that input at least sequencing data (e.g., methylation, single nucleotide polymorphism (SNP), DNA, RNA, etc.) and output the likelihood that the sample has cancer based on the input sequencing data. One example of a classifier is a mixed model classifier. A distribution of non-cancer samples can be generated according to the cancer scores of the non-cancer samples. A binary threshold cutoff can be set to ensure some level of specificity of the binary classification, e.g., true negative rate. Typically, a high specificity cutoff is used when classifying cancer, e.g., 99.4% specificity or higher. However, many non-cancer samples used to train the cancer classifier that are just below the specificity cutoff may have high tissue signal, thereby positively biasing the binary threshold cutoff.

[0105] In some embodiments, certain cancer types are further divided into cancer subtypes.For example, blood cancer types can be further divided into, for example, circulating lymphoid subtypes, non-Hodgkin's lymphoma (NHL) indolent subtypes, NHL aggressive subtypes, Hodgkin's lymphoma (HL) subtypes, myeloid subtypes, and plasma cell subtypes, all of which belong to the class of lymphoid neoplasms.

[0106] To demonstrate, non-cancer samples with greater than 95% specificity were selected and then input into a multi-class cancer classifier to determine the probability of each cancer type or tissue of origin (TOO). The cancer types or TOO labels used in this embodiment of the multi-class cancer classifier include circulating lymphoid, myeloid, NHL indolent, colorectal, NHL aggressive, lung, uterus, breast, prostate, pancreas and gallbladder, upper gastrointestinal, bladder and urothelium, plasma cell, head and neck, kidney, ovary, sarcoma, liver and bile duct, cervix, other tissues, HL, anorectal, melanoma, and thyroid. The graph in Figure 8 shows a number of non-cancer samples with high tissue signals from at least one tissue type. Each point in the tissue type column corresponds to the likelihood of the tissue of origin of the non-cancer sample exceeding the 95% specificity threshold. Notably, many tissue types have multiple non-cancer sample outliers with significant tissue contributions that are not typical for non-cancer samples. This occurs when such non-cancer samples have cfDNA signals that are driven by cancer-like methylation, clonal fraction, and / or growth / turnover rates. In any case, these non-cancer samples with significant tissue contributions shift the binary classification threshold cutoff upwards, especially for samples with significant tissue signals that are slightly below the pre-set binary classification threshold cutoff, thereby decreasing the sensitivity of the cancer classification. In fact, such signals (e.g., corresponding to circulating lymphoid, myeloid, and indolent NHL) are major attractors of false positive determinations. Of note, circulating lymphoid, myeloid, indolent NHL, colorectal, aggressive NHL, lung, uterus, breast, prostate, pancreas and gallbladder, upper gastrointestinal tract, plasma cell, head and neck, cervix, HL had at least one non-cancer sample with a tissue of origin probability above 0.1. Specifically, circulating lymphoid, myeloid, indolent NHL, and aggressive NHL (all hematological subtypes) had two or more cancer samples with a tissue of origin probability greater than 0.5.

[0107] Referring to Figures 9A and 9B, 9A and 9B illustrate graphs of hematological subtypes separated according to methylation sequencing data. The graphs in Figures 9A and 9B show the ability to model hematological subtypes. This can prove beneficial in providing additional granularity to multi-class cancer classification (e.g., further classifying with hematological subtype labels) or as a way to refine cancer classification through excluding non-cancer samples with high hematological subtype signals before training a cancer classifier. As mentioned above, methylation signals can cover multiple CpG sites, thereby creating a high-dimensional vector space. The hematological subtypes shown may include circulating lymphoid, solid lymphoid, plasma cell, and myeloid. The solid lymphoid subtype may be further divided into HL, NHL indolent, and NHL aggressive.

[0108] In FIG. 9A, the analysis system executes a t-distributed stochastic neighbor embedding method. The t-distributed stochastic neighbor embedding method identifies a vector space (containing the methylation sequencing data) of reduced dimensionality with fewer embeddings. The embeddings are in order of variance in the methylation signals between samples. The first principle embedding, denoted as "V1" on the horizontal axis of the graph, has the highest variance. The third principle embedding, denoted as "V2" on the vertical axis of the graph, has the third highest variance. The clusters of each hematological subtype and non-cancer samples are annotated on the graph.

[0109] In FIG. 9B, the analysis system performs UMAP embedding. UMAP embedding also reduces the dimensionality of the vector space to a smaller number of embeddings. The embeddings are in order of variance in the methylation signals between samples. The first principle embedding, denoted "Embedding 1" on the horizontal axis of the graph, has the highest variance. The second principle embedding, denoted "Embedding 2" on the vertical axis of the graph, has the second highest variance. Non-cancer samples are shown using contour density.

[0110] The graphs show the ability to classify according to hematological subtypes, either for the addition of hematological subtypes in the multi-class cancer classification, or for modeling each of the hematological subtypes for tuning the cancer classifier.

[0111] III.CI Removal of high signal non-cancer samples In one or more embodiments, the analysis system refines the trained cancer classifier by excluding non-cancer samples that were used to train the cancer classifier. Specifically, the analysis system may seek to exclude non-cancer samples with high tissue signals that dilute the sensitivity of the cancer classifier in predicting cancer.

[0112] FIG. 10A illustrates a flow chart describing a process 1000 for determining a binary threshold cutoff for binary cancer classification according to one or more embodiments. Binary classification for making a prediction between cancer and non-cancer evaluates a sample's cancer score against a determined binary threshold cutoff, where a sample with a cancer score below the binary threshold cutoff is determined as non-cancer and a sample with a cancer score at or above the binary threshold cutoff is determined as cancer. A trained multi-class cancer classifier evaluates the sample's methylation signal (and / or other sequencing data) to determine a probability for several TOO labels classified by the multi-class cancer classifier. The TOO labels used in the multi-class cancer classifier may be cancer tissue types or cancer tissue subtypes (e.g., hematological subtypes as described above). Process 1000 may be performed or accomplished by an analysis system.

[0113] The analysis system receives 1010 sequencing data of a plurality of biological samples containing cfDNA fragments, the biological samples including cancer samples and non-cancer samples. The sequencing data can be methylation sequencing data, SNP sequencing data, other DNA sequencing data, RNA sequencing data, etc.

[0114] For each non-cancer sample, the analysis system classifies 1020 the non-cancer sample based on the sequencing-derived features using a multi-class cancer classifier, where the multi-class cancer classifier predicts a probability for each of a plurality of TOO labels. The analysis system can generate a feature vector for the non-cancer sample according to step 320 of FIG. 3A, i.e., assign an anomaly score for each of its CpG sites based on at least one aberrantly methylated cfDNA fragment that overlaps the considered CpG site.

[0115] For each non-cancer sample, the analysis system determines whether the predicted probability likelihood for one or more TOO labels is above a TOO threshold 1030. Determining the TOO threshold is further described below in FIG. 10B.

[0116] The analysis system determines 1040 a binary threshold cutoff for predicting the presence of cancer, where the binary threshold cutoff is determined based on the distribution of non-cancer samples excluding one or more non-cancer samples identified as having at least one probability likelihood above the TOO threshold. For a TOO label, non-cancer samples having at least one probability likelihood above the TOO threshold corresponding to that TOO label are excluded. The analysis system then calculates the distribution of non-cancer samples according to the cancer score of each non-cancer sample, and then determines a binary threshold cutoff from the distribution at a desired specificity level (e.g., 99.4-99.9% specificity). It is noted that each cancer score may be determined according to sequencing data, for example, the cancer score may be output by a binary cancer classifier that predicts the likelihood of cancer based on methylation sequencing data, as described in FIG. 3A. In other embodiments, the cancer score can be calculated according to other methods that input at least sequencing data (e.g., methylation, single nucleotide polymorphisms (SNPs), DNA, RNA, etc.) and output the likelihood that the sample has cancer based on the input sequencing data.

[0117] FIG. 10B illustrates a flow chart describing a process 1005 of thresholding TOO labels to determine a binary threshold cutoff for binary cancer classification, according to one or more embodiments. This process 1005 may be an embodiment of process 1000. Binary classification to make a prediction between cancer and non-cancer evaluates the cancer score of a sample against a determined binary threshold cutoff, where a sample with a cancer score below the binary threshold cutoff is determined as non-cancer and a sample with a cancer score at or above the binary threshold cutoff is determined as cancer. A trained multi-class cancer classifier evaluates the methylation signal (and / or other sequencing data) of the sample to determine a probability for some TOO labels classified by the multi-class cancer classifier. The TOO labels may be cancer tissue types, or more specifically cancer tissue subtypes (e.g., hematological subtypes as described above). Process 1005 may be performed or accomplished by an analysis system.

[0118] The analysis system obtains 1015 a training set including a plurality of samples with a cancer or non-cancer label, i.e., either cancer samples or non-cancer samples, respectively, and a holdout set including a plurality of samples with a cancer or non-cancer label, i.e., either cancer samples or non-cancer samples, respectively. Each sample in the training set includes methylation sequencing data, for example, generated according to process 100 of FIG. 1A. In other embodiments, each training sample has other sequencing data used in tandem with or as an alternative to the methylation sequencing data. Additionally, each sample from the training set and the holdout set has a cancer score. As described above, the cancer score can be determined by a binary classifier as the likelihood that the sample has cancer given the methylation sequencing data of the sample. In other embodiments, the cancer score is calculated according to other methods that input at least sequencing data (e.g., methylation, single nucleotide polymorphism (SNP), DNA, RNA, etc.) and output the likelihood that the sample has cancer according to the input sequencing data.

[0119] The analysis system determines 1025 a feature vector for each non-cancer training sample based on the methylation sequencing data. The analysis system may determine a feature vector for each non-cancer training sample, for example, in a manner similar to step 320 of FIG. 3A , which describes determining an anomaly score for each CpG site in the set of CpG sites considered. In some embodiments, the analysis system defines the anomaly score of the feature vector using a binary score based on whether an anomaly fragment containing the CpG site is present in the set of anomaly fragments. Once all anomaly scores have been determined for a sample, the analysis system determines the feature vector as a vector of anomaly scores associated with each CpG site considered. The analysis system may further normalize the anomaly scores of the feature vector based on the coverage of the sample.

[0120] The analysis system inputs the feature vector of each non-cancer training sample into a multi-class cancer classifier to generate a TOO prediction 1035. The multi-class cancer classifier is trained with multiple TOO labels, including cancer type, cancer subtype, non-cancer, or any combination thereof. The multi-class cancer classifier may be trained according to process 300 of FIG. 3A. The trained multi-class classifier determines multiple probabilities for the TOO labels as cancer predictions, where the probabilities for the TOO labels indicate the likelihood of having cancer corresponding to the TOO labels.

[0121] In some examples, the analysis system sweeps 1045 or iterates through a range of probabilities for the TOO label as a candidate TOO threshold over which it calculates specificity and sensitivity rates. The analysis system may sweep the range of probabilities in increments of, for example, 0.01, 0.02, 0.03, 0.04, 0.05, etc. As the analysis system sweeps through the range of probabilities, the analysis system filters non-cancer training samples that have a probability of the TOO label that is at or above the candidate TOO threshold according to the output of the multi-class cancer classifier. As a numerical example, the analysis system considers a candidate TOO threshold of 0.35. Non-cancer training samples that have a probability of the TOO label at or above 0.35 are filtered out of the training set. The analysis system determines an adjusted binary threshold cutoff based on the filtered training set. The analysis system calculates a specificity rate of the prediction using the adjusted binary threshold cutoff for the holdout set. Specificity refers to the accuracy rate of identifying non-cancer samples as non-cancer labels. The analysis system also calculates the sensitivity rate of prediction using a binary threshold cutoff adjusted for the holdout set. Sensitivity refers to the accuracy rate of identifying cancer samples as cancer labels. In practice, the specificity rate and / or sensitivity rate may be defined according to a true positive rate, a false positive rate, a true negative rate, a false negative rate, another statistical calculation, etc.

[0122] The analysis system determines 1055 a TOO threshold value for the TOO label. The analysis system selects a TOO threshold value from the candidate TOO threshold values ​​by optimizing the specificity rate and / or sensitivity rate calculated over a range of candidate TOO threshold values. In some examples, a TOO threshold value is determined or otherwise applied for a particular TOO tissue type or subtype class, e.g., hematological class. By way of example only, an algorithm for calculating and applying a TOO-specific probability threshold can be used to remove non-cancer samples with excess signals for hematological disorders. The algorithm may include, for each pre-specified TOO label, first searching through a grid of probability values, and for all values, evaluating the clinical specificity and clinical sensitivity of the holdout set using a binary detection threshold calculated after removing non-cancer samples with a probability equal to or greater than the specified TOO label. By iterating through the probability grid, the algorithm will identify a combination of values ​​of the TOO threshold value that optimizes the tradeoff between clinical specificity and clinical sensitivity of the holdout set for the pre-specified TOO label. The final optimized TOO probability threshold value will be used to filter out non-cancer samples that are above any of the values ​​of the given TOO label. The cleaned set of non-cancer samples will be used to calculate the cancer-non-cancer detection threshold. Still, in some examples, the TOO specific thresholding may be manually set to any boundary point, e.g., a desired specificity level (e.g., 99.4-99.9% specificity).

[0123] The analysis system adjusts the binary cancer classification by filtering out non-cancer training samples that exceed the TOO thresholding before determining the binary threshold cutoff 1065. The analysis system filters out the non-cancer training samples from the training set according to the TOO threshold determined for the TOO labels. The analysis system sets the binary threshold cutoff according to the filtered training set. For example, the analysis system determines a new binary threshold cutoff based on the filtered score distribution. In additional embodiments, the analysis system can determine a TOO threshold for any of the TOO labels according to steps 1010, 1020, 1030, and 1040 to adjust the binary cancer classification.

[0124] III.C.II. Stratification of sample distribution according to TOO signal In one or more embodiments, the analysis system adjusts the cancer classifier by stratifying the sample distribution according to the TOO signal and determining a binary threshold cutoff for each stratum. The analysis system can stratify the sample distribution according to a signal of one or more TOO labels determined according to the TOO predictions output by the multi-class cancer classifier.

[0125] Reference is now made to Figure 13A, which illustrates a process for stratifying hematological signals into two layers, according to one or more embodiments. Although the following description describes stratification using hematological signals, the principles can be readily applied to other TOO signals.

[0126] The analysis system stratifies 1300A a holdout set of cancer and non-cancer samples into a low signal tier 1310 and a high signal tier 1320 according to the hematological signal. Each sample in the holdout set has a cancer score determined by a binary cancer classifier and a TOO prediction determined by a multi-class cancer classifier. In one embodiment, the hematological signal of the sample is determined according to the TOO prediction output by the multi-class cancer classifier. A high tissue signal may be determined as described in section III.C. Tuning the Cancer Classifier. In one embodiment, when considering one or more top predictions (e.g., top 1, top 2, etc.), a high hematological signal is determined if at least one of the top predictions being considered is of a hematological subtype (e.g., lymphoid neoplasm subtype and myeloid neoplasm subtype). Other hematological subtypes may be included. Thus, if a sample has a TOO prediction in which at least one of the top predictions is considered to be a lymphoid or myeloid neoplasm subtype, the sample is determined to have a high hematological signal. Otherwise, the sample is determined not to have a high hematological signal.

[0127] The analysis system determines a binary threshold cutoff for each stratum to predict the presence or absence of cancer in the samples. The samples in the low signal stratum 1310 are used by the analysis system to determine 1305 a binary threshold cutoff for predicting the absence or presence of cancer in the samples in the low signal stratum 1310. The binary threshold cutoff is determined 1305 according to a false positive budget set for the low signal stratum 1310. Using the cancer scores of the samples in the low signal stratum 1310, the analysis system sweeps through a range of candidate binary threshold cutoffs and evaluates the true positive rate (also referred to as sensitivity) and false positive rate at each candidate binary threshold cutoff. The candidate binary threshold cutoff with the closest false positive rate within the false positive budget is determined to be the candidate binary threshold cutoff. The analysis system performs similar operations to determine 1315 a binary threshold cutoff for the high signal stratum 1320. The false positive budget of the low signal stratum 1310 and the false positive budget of the high signal stratum 1320 may be set according to the ratio of the statistical true positive rates of the stratum, which aims to suppress the false positive rate in the high signal stratum 1320.

[0128] For a test sample, the analytical system places the test sample in either a low signal tier 1310 or a high signal tier 1320 according to the hematological signal. If the test sample is placed in the low signal tier 1310, the analytical system applies the binary threshold cutoff 1315 of the low signal tier 1310 to the cancer score of the test sample. If the cancer score is above or equal to the binary threshold cutoff of the low signal tier 1310, the analytical system returns a prediction of the presence of cancer in the test sample, otherwise it returns a prediction of the absence of cancer. If the test sample is placed in the high signal tier 1320, the binary threshold cutoff of the low signal tier 1320 is applied 1325 to the cancer score of the test sample. If the cancer score is above or equal to the binary threshold cutoff of the high signal tier 1320, the analytical system returns a prediction of the presence of cancer in the test sample, otherwise it returns a prediction of the absence of cancer.

[0129] 13B illustrates a process for stratifying hematological signals into three layers according to one or more embodiments. Although the following description describes stratification using hematological signals, the principles can be easily applied to other TOO signals. The principles can also be easily extended to stratification into more than three layers.

[0130] The analysis system stratifies the holdout set of cancer and non-cancer samples according to hematological signals into three tiers: low signal tier 1330, moderate signal tier 1340, and high signal tier 1350. Each sample in the holdout set has a cancer score determined by a binary cancer classifier and a TOO prediction determined by a multi-class cancer classifier. In one or more embodiments, the hematological TOO label includes multiple hematological subtypes. Any sample in the holdout set that has a high tissue signal in one or more aggressive hematological subtypes is placed in the high signal tier 1350. Any sample in the holdout set that has a high tissue signal in one or more indolent hematological subtypes (not already classified in the high signal tier 1350) is placed in the moderate signal tier 1340. Any sample that is not classified in either the high signal tier 1350 or the moderate signal tier 1340 is placed in the low signal tier 1330. Following the principles described in Figure 13A, the analysis system determines a binary threshold cutoff for each stratum based on the false positive budget for each stratum, where a binary threshold cutoff for low signal stratum 1330 is determined 1335, a binary threshold cutoff for moderate signal stratum 1340 is determined 1345, and a binary threshold cutoff for high signal stratum 1350 is determined 1355. Similar to Figure 13A, for a test sample, the analysis system identifies the stratum in which the test sample is located and applies the binary threshold cutoff for that stratum to predict the presence or absence of cancer in the test sample.

[0131] 13C illustrates a process for stratifying first the hematological signal, followed by the colorectal signal, according to one or more embodiments. Although the following description describes stratification using the hematological signal followed by the colorectal signal, the principles can be readily applied to other combinations of TOO signals.

[0132] The analysis system stratifies the holdout set of cancer and non-cancer samples according to the hematological signal 1300C and then according to the colorectal signal 1370. Each sample in the holdout set has a cancer score determined by a binary cancer classifier and a TOO prediction determined by a multi-class cancer classifier. Similar to the principle described above in FIG. 13A, any sample in the holdout set that has a high hematological signal is placed in the high signal tier 1360. The remaining samples are then stratified according to the colorectal signal 1370. Similar to the hematological stratification, any sample that has a high colorectal signal is placed in the high colorectal signal tier 1380. Samples that are not placed in either the high hematological signal tier 1360 or the high colorectal signal tier 1380 are grouped into the low signal tier 1390. In this embodiment, the hematological signal is of higher priority than the colorectal signal. In one or more embodiments, the multiple TOO signals may be evaluated sequentially in order of priority. Thus, a sample with both high hematology and high colorectal signals would be placed in the high hematology tier 1360 instead of the high colorectal tier 1380, since the hematology signal is of higher priority than the colorectal signal. In accordance with the principles described in FIG. 13A, the analytical system determines a binary threshold cutoff for each tier based on the false positive budget of each tier. A binary threshold cutoff for the high hematology signal tier 1360 is determined 1365, a binary threshold cutoff for the high colorectal signal tier 1380 is determined 1385, and a binary threshold cutoff for the low signal tier 1390 is determined 1395. Similar to FIG. 13A, for a test sample, the analytical system identifies a tier in which to place the test sample and applies the binary threshold cutoff for that tier to predict the presence or absence of cancer in the test sample.

[0133] 14 illustrates a process 1400 for determining a binary threshold cutoff for TOO stratification, according to one or more embodiments. Although process 1400 is described as being performed by an analysis system, process 1400 may more generally be performed by any computing system.

[0134] The analysis system obtains a holdout set including a plurality of samples classified as cancer or non-cancer or having a cancer or non-cancer label 1410. Each sample in the holdout set is associated with a cancer score (e.g., determined by a binary cancer classifier), e.g., representing the likelihood that the sample corresponds to cancer, and a TOO prediction (e.g., determined by a multi-class cancer classifier), e.g., representing the likelihood that the sample corresponds to cancer of a particular tissue type.

[0135] The analysis system stratifies the holdout set into a first tier of high signal and a second tier of low signal for the first TOO label based on the TOO prediction 1420. In one example, the stratification uses a prediction value threshold. Any sample whose prediction value for the first TOO label in the TOO prediction is at or above the prediction value threshold is classified as a high signal for the first TOO label. Otherwise, the sample is classified as a low signal for the first TOO label. In another example, the analysis system considers one or more top predictions in the TOO prediction of each sample. Any sample whose first TOO label is in at least one of the top predictions considered is classified as a high signal for the first TOO label. Otherwise, the sample is classified as a low signal for the first TOO label.

[0136] In one or more embodiments, the analysis system further stratifies into a third tier of moderate signals for the first TOO label. In embodiments in which the signals are stratified according to predictive value, the range of predictive values ​​may be segmented into three parts that determine high, moderate, and low signals.

[0137] In one or more embodiments, the analysis system further stratifies the one or more layers into additional layers according to tissue signals for one or more additional TOO labels, which may have a lower priority in the stratification than the first TOO label.

[0138] For each stratum, the analysis system sweeps 1440 across the domain of cancer scores at multiple candidate binary threshold cutoffs and calculates the true positive rate and false positive rate for each candidate binary threshold cutoff. The true positive rate can be plotted against the false positive rate to generate a receiver operator characteristic (ROC) curve.

[0139] The analysis system determines a binary threshold cutoff for each stratum based on the false positive budget 1440. The false positive budget may be assigned to each stratum according to the ratio of the stratum's statistical true positive rates.

[0140] 15 illustrates a flow chart describing a process 1500 for predicting the presence or absence of cancer for a test sample using a binary threshold cutoff determined by TOO stratification, according to one or more embodiments. Although process 1500 is described as being performed by an analytical system, process 1500 may be accomplished more generally by any computing system.

[0141] The analysis system obtains 1510 a test sample with unknown presence of cancer. The test sample is accompanied by a cancer score, e.g., determined by a binary cancer classifier, and a TOO prediction, e.g., determined by a multi-class cancer classifier.

[0142] The analytical system stratifies the test sample into a first tier of high signal or a second tier of low signal for the first TOO label based on the TOO prediction 1520. The placement (or classification) is described above (e.g., with respect to the stratification in step 1420 of process 1400).

[0143] The analysis system predicts 1530 whether the test sample has the presence or absence of cancer by comparing the cancer score against the binary threshold cutoff of the tier in which the test sample is placed. For example, if the test sample has a high signal for the first TOO label and is placed in the first tier of high signals, the analysis system applies the binary threshold cutoff determined for the first tier of high signals to the cancer score of the test sample. Alternatively, if the test sample is placed in the second tier of low signals, the binary threshold cutoff determined for the second tier is used. If the cancer score of the test sample is at or above the binary threshold cutoff used, the test sample is predicted to have the presence of cancer. Otherwise, the test sample is predicted to have the absence of cancer.

[0144] III.D. Deployment of Cancer Classifiers During use of the cancer classifier, the analysis system obtains a test sample obtained from a subject of unknown cancer type. The analysis system processes the test sample composed of DNA molecules using any combination of processes 100, 200, and 220 to achieve a set of abnormal fragments. The analysis system determines a test feature vector for use by the cancer classifier according to similar principles discussed in process 300. The analysis system calculates an anomaly score for each CpG site within the plurality of CpG sites used by the cancer classifier. For example, the cancer classifier receives as input a feature vector including an anomaly score for 1,000 selected CpG sites. The analysis system thus determines a test feature vector including an anomaly score for the 1,000 selected CpG sites based on the set of abnormal fragments. The analysis system calculates the anomaly score in the same manner as the training sample. In one embodiment, the analysis system defines the anomaly score as a binary score based on whether there is a hypermethylated or hypomethylated fragment that includes the CpG site within the set of abnormal fragments.

[0145] The analysis system then inputs the test feature vector into a cancer classifier. A function of the cancer classifier then generates a cancer prediction based on the classification parameters trained in process 300 and the test feature vector. In a first mode, the cancer prediction is binary and selected from the group consisting of "cancer" or "non-cancer", and in a second mode, the cancer prediction is selected from the group of multiple cancer types and "non-cancer". In an additional embodiment, the cancer prediction has a predictive value for each of multiple cancer types. Furthermore, the analysis system may determine that the test sample is likely to be one of the cancer types. Continuing with the above example of a test sample having cancer predictions as 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, if the cancer prediction is binary as a 60% likelihood of non-cancer and a 40% likelihood of cancer, the analysis system determines that the test sample is most likely to not have cancer. In additional embodiments, the cancer prediction with the highest likelihood may still be compared to a threshold (e.g., 40%, 50%, 60%, 70%) to call the test subject having that cancer type. If the cancer prediction with the highest likelihood does not exceed that threshold, the analysis system may return an inconclusive result.

[0146] In an additional embodiment, the analysis system chains the cancer classifier trained in step 360 of process 300 with another cancer classifier trained in step 370 or process 300. The analysis system inputs the feature vector into the cancer classifier trained as a binary classifier in step 360 of process 300. The analysis system receives the output of the cancer prediction. The cancer prediction may be binary with respect to whether the test subject is likely to have cancer or likely not to have cancer. In other implementations, the cancer prediction includes a predictive value that describes the likelihood of cancer and the likelihood of non-cancer. For example, the cancer prediction has a predictive value of 85% for cancer and a predictive value of 15% for non-cancer. 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 between different cancer types. The multi-class cancer classifier receives the test feature vector and returns a cancer prediction for 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 implementation, the multi-class cancer classifier provides a prediction value for each cancer type of the multiple cancer types. For example, the cancer predictions may include a 40% prediction value for breast cancer type, a 15% prediction value for colorectal cancer type, and a 45% prediction value for liver cancer type.

[0147] According to a generalized embodiment of binary cancer classification, the analysis system determines a cancer score for the test sample based on the 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 compares the cancer score of the test sample against a binary threshold cutoff to predict whether the test sample is likely to have cancer. The binary threshold cutoff may be adjusted using TOO thresholding based on one or more TOO subtype classes. The analysis system may further generate a feature vector of the test sample for use in a multi-class cancer classifier to determine a cancer prediction indicative of one or more possible cancer types.

[0148] IV. Application In some embodiments, the methods, analysis systems, and / or classifiers of the present invention may be used to detect the presence of cancer, monitor the progression or recurrence of cancer, monitor the response or effectiveness of treatment, determine the presence or monitoring of minimal residual disease (MRD), or any combination thereof. For example, as described herein, the classifiers may be used to generate a probability score (e.g., from 0 to 100) that describes the likelihood that a test feature vector is derived 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 may be evaluated at multiple different time points (e.g., before or after treatment) to monitor disease progression or to monitor the effectiveness of a treatment (e.g., therapeutic effect). In still other embodiments, the likelihood or probability score may be used to make or affect a clinical decision (e.g., diagnosing cancer, selecting a treatment, evaluating the effectiveness of a treatment, etc.). For example, in one embodiment, if the probability score is above a threshold, a physician may prescribe an appropriate treatment.

[0149] IV.A. Early Cancer Detection 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 above in Section III and exemplified in Section V) can be used to determine a cancer prediction that describes the likelihood that a test feature vector is derived from a subject with cancer.

[0150] In one embodiment, the cancer prediction is a likelihood (e.g., a score of 0 to 100) that the test sample has or does not have cancer (i.e., a binary classification). Thus, the analysis system may determine a threshold value for determining whether the test subject has cancer. For example, a cancer prediction of 60 or above can indicate that the subject has cancer. In yet other embodiments, a cancer prediction of 65 or above, 70 or above, 75 or above, 80 or above, 85 or above, 90 or above, or 95 or above indicates that the subject has cancer. In other embodiments, the cancer prediction can indicate disease severity. For example, a cancer prediction of 80 can indicate a more severe form or a later stage of cancer compared to a cancer prediction below 80 (e.g., a probability score of 70). Similarly, an increase in the cancer prediction over time (e.g., as determined by classifying test feature vectors derived from multiple samples obtained from the same subject taken at two or more time points) can indicate disease progression, or a decrease in the cancer prediction over time can indicate treatment success.

[0151] In another embodiment, the cancer prediction includes multiple predictive values, where each of a plurality of classified (i.e., multi-class classification) cancer types has a predictive value (e.g., scored from 0 to 100). The predictive values ​​may correspond to the likelihood that a given training sample (and during estimation, the training sample) has each of the cancer types. The analysis system may identify the cancer type with the highest predictive value and indicate that the test subject is likely to have that cancer type. In other embodiments, the analysis system may further compare the highest predictive 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 predictive value may also indicate the severity of the disease. For example, a predictive value above 80 may indicate a more severe form or a later stage of cancer compared to a predictive value of 60. Similarly, an increase in predictive value over time (e.g., as determined by classifying test feature vectors derived from multiple samples obtained from the same subject taken at two or more time points) can indicate disease progression, or a decrease in predictive value over time can indicate successful treatment.

[0152] According to an embodiment of the present invention, the method and system of the present invention can be trained to detect or classify multiple cancer indications.For example, the method, system and classifier of the present invention can be used to detect the presence of one or more, two or more, three or more, five or more, ten or more, fifteen or more, or twenty or more different types of cancer.

[0153] 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 squamous cell carcinoma (e.g., epithelial squamous cell carcinoma), skin cancer, melanoma, small cell lung cancer, non-small cell lung cancer ("NSCLC"), lung adenocarcinoma, and lung squamous cell carcinoma, lung cancer, peritoneal cancer, gastric or stomach cancer, including gastrointestinal cancer. 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 cancer, 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 or uterine cancer, salivary gland cancer, kidney or renal cancer (e.g., renal cell carcinoma, nephroblastoma, or Wilms tumor), prostate cancer, vulvar cancer, thyroid cancer, anal cancer, penile cancer, head and neck cancer, esophageal cancer, and nasopharyngeal carcinoma (NPC). Additional examples of cancer include, without limitation, retinoblastoma, thecoma, androgenitoma, non-Hodgkin's lymphoma (NHL), multiple myeloma, and hematological malignancies, including but not limited to acute hematological malignancies, endometriosis, fibrosarcoma, choriomas, laryngeal carcinoma, Kaposi's sarcoma, schwannoma, oligodendroglioma, neuroblastoma, rhabdomyosarcoma, osteogenic sarcoma, leiomyosarcoma, and urinary tract cancer.

[0154] In some embodiments, the cancer is one or more of anorectal cancer, bladder cancer, breast cancer, cervical cancer, colorectal cancer, esophageal cancer, gastric cancer, head and neck cancer, hepatobiliary cancer, leukemia, lung cancer, lymphoma, melanoma, multiple myeloma, ovarian cancer, pancreatic cancer, prostate cancer, renal cancer, thyroid cancer, uterine cancer, or any combination thereof.

[0155] In some embodiments, the one or more cancers can be "high signal" cancers (defined as cancers with a 5-year cancer-specific mortality rate of greater than 50%), such as anorectal, colorectal, esophageal, head and neck, hepatobiliary, lung, ovarian, and pancreatic cancers, as well as lymphoma and multiple myeloma. High signal cancers tend to be more aggressive and typically have higher than average cell-free nucleic acid concentrations in test samples obtained from patients.

[0156] IV.B. Cancer and Treatment Monitoring In some embodiments, the cancer prognosis may be evaluated at multiple different time points (e.g., before or after treatment) to monitor disease progression or to monitor the effectiveness of a treatment (e.g., therapeutic effect). For example, the present invention includes a method comprising 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 prognosis therefrom (as described herein), obtaining a second test sample (e.g., a second plasma cfDNA sample) from a cancer patient at a second time point and determining a second cancer prognosis therefrom (as described herein).

[0157] In certain embodiments, the first time point is before cancer treatment (e.g., before resection surgery or therapeutic intervention), and the second time point is after cancer treatment (e.g., after resection surgery or therapeutic intervention), and the classifier is utilized to monitor the effectiveness of the treatment. For example, if the second cancer prediction is decreased compared to the first cancer prediction, the treatment is considered successful. However, if the second cancer prediction is increased compared to the first cancer prediction, the treatment is considered unsuccessful. In other embodiments, both the first and second time points are before cancer treatment (e.g., before resection surgery or therapeutic intervention). In yet other embodiments, both the first and second time points are after cancer treatment (e.g., after resection surgery or therapeutic intervention). In still other embodiments, cfDNA samples may be obtained from a cancer patient at a first and second time point and analyzed, e.g., to monitor the progression of the cancer, to determine if the cancer is in remission (e.g., following treatment), to monitor or detect residual disease or disease recurrence, or to monitor the effectiveness of a treatment (e.g., therapy).

[0158] One of skill in the art will readily appreciate that test samples may be obtained from a cancer patient over any desired set of time points and analyzed according to the methods of the present invention to monitor the status of the cancer in the patient. In some embodiments, the first and second time points range from about 15 minutes up to about 30 years, e.g., about 30 minutes, e.g., about 1 hour, 2 hours, 3 hours, 4 hours, 5 hours, 6 hours, 7 hours, 8 hours, 9 hours, 10 hours, 11 hours, 12 hours, 13 hours, 14 hours, 15 hours, 16 hours, 17 hours, 18 hours, 19 hours, 20 hours, 21 hours, 22 hours, 23 hours, or about 24 hours, e.g., about 1 day, 2 days, 3 days, 4 days, 5 days, 10 days, 15 days, 20 days, 25 days, or about 30 days, or e.g., about 1 month, 2 months, 3 months, 4 months, 5 months, 6 months, 7 months, 8 months, 9 months, 10 months, 11 months, or 12 months, or e.g., about 1 year, 1.5 years, 2 years, 2.5 years. 3 years, 3.5 years, 4 years, 4.5 years, 5 years, 5.5 years, 6 years, 6.5 years, 7 years, 7.5 years, 8 years, 8.5 years, 9 years, 9.5 years, 10 years, 10.5 years, 11 years, 11.5 years, 12 years, 12.5 years, 13 years, 13.5 years, 14 years, 14.5 years, 15 years, 15.5 years, 16 years, 16.5 years, 17 years, 17.5 The test samples may be separated by 1 year, 18 years, 18.5 years, 19 years, 19.5 years, 20 years, 20.5 years, 21 years, 21.5 years, 22 years, 22.5 years, 23 years, 23.5 years, 24 years, 24.5 years, 25 years, 25.5 years, 26 years, 26.5 years, 27 years, 27.5 years, 28 years, 28.5 years, 29 years, 29.5 years, or about 30 years. In other embodiments, test samples may be obtained from a patient at least once every 3 months, at least once every 6 months, at least once every 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.

[0159] IV.C. Treatment In yet another embodiment, the cancer prediction may be used to make or influence clinical decisions (e.g., diagnosing cancer, selecting a treatment, evaluating the effectiveness of a treatment, etc.) For example, in one embodiment, if the cancer prediction (e.g., for a cancer, or for a particular cancer type) is above a threshold, a physician can prescribe an appropriate treatment (e.g., resective surgery, radiation therapy, chemotherapy, and / or immunotherapy).

[0160] A classifier (described herein) can be used to determine a cancer prediction where the feature vector of the sample is derived from a subject with cancer. In one embodiment, an appropriate treatment (e.g., resection surgery or therapy) is prescribed if the cancer prediction is above a threshold. For example, in one embodiment, if the cancer prediction is 60 or above, one or more appropriate treatments are prescribed. In another embodiment, if the cancer prediction is 65 or above, 70 or above, 75 or above, 80 or above, 85 or above, 90 or above, or 95 or above, one or more appropriate treatments are prescribed. In other embodiments, the cancer prediction may indicate the severity of the disease. An appropriate treatment that matches the severity of the disease may then be prescribed.

[0161] 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, hormonal therapeutic agents, and immunotherapeutic agents. For example, the treatment may be one or more chemotherapeutic agents selected from the group consisting of alkylating agents, antimetabolites, anthracyclines, antitumor antibiotics, cytoskeleton disrupting agents (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, proteosome inhibitors, angiogenesis inhibitors, and monoclonal antibody conjugates. In some embodiments, the treatment is one or more differentiation therapeutic agents, including retinoids, e.g., tretinoin, alitretinoin, and bexarotene. In some embodiments, the treatment is one or more hormonal therapy agents selected from the group consisting of antiestrogens, aromatase inhibitors, progestins, estrogens, antiandrogens, and GnRH agonists or analogs. In one embodiment, the treatment is one or more immunotherapy agents selected from the group including monoclonal antibody therapy, such as rituximab (RITUXAN) and alemtuzumab (CAMPATH), non-specific immunotherapeutic agents and adjuvants, such as BCG, interleukin-2 (IL-2), and interferon-alpha, immunomodulatory agents, such as thalidomide and lenalidomide (REVLIMID). It is within the capabilities of a skilled physician or oncologist to select the appropriate cancer therapy agent based on characteristics, such as tumor type, stage of cancer, previous exposure to cancer treatments or therapeutic agents, and other characteristics of the cancer.

[0162] V. Exemplary Results of Cancer Classifier VA sample collection and processing Study design and samples: CCGA (NCT02889978) is a prospective, multicenter, case-control, observational study with longitudinal follow-up. De-identified biospecimens were collected from approximately 15,000 participants from 142 sites. Samples were split into training (1,785) and testing (1,015) sets, samples were selected to ensure a pre-specified distribution of cancer types and non-cancer across sites in each cohort, and cancer and non-cancer samples were often age-matched by sex.

[0163] Whole-genome bisulfite sequencing: cfDNA was isolated from plasma and whole-genome bisulfite sequencing (WGBS, 30x depth) was utilized for analysis of cfDNA. cfDNA was extracted from two tubes of plasma per patient (maximum, 10ml total volume) using a modified QIAamp Circulating Nucleic Acid kit (Qiagen, Germantown, MD). Up to 75ng of plasma cfDNA was subjected to bisulfite conversion using the EZ-96 DNA methylation kit (Zymo Research, D5003). The converted cfDNA was used to prepare dual-indexed sequencing libraries using the Accel-NGS Methyl-Seq DNA Library Preparation Kit (Swift BioSciences, Ann Arbor, MI), and constructed libraries were quantified using the KAPA Library Quantification Kit for Illumina Platforms (Kapa Biosystems, Wilmington, MA). The four libraries were pooled with a 10% PhiX v3 library (Illumina, FC-110-3001) and clustered on an Illumina NovaSeq 6000 S2 flow cell, followed by 150 bp paired-end sequencing (30×).

[0164] For each sample, the WGBS fragment set was reduced to a smaller subset of fragments with abnormal methylation patterns. In addition, cfDNA fragments that were hyper- or hypomethylated were selected. cfDNA fragments that had abnormal methylation patterns and were hyper- or hypomethylated, i.e., UFXM, were selected. Fragments that occur frequently in individuals without cancer, or those with unstable methylation, are unlikely to provide highly distinguishable features for classification of cancer status. We therefore used an independent reference set (i.e., reference genome) of 108 cancer-free, non-smoker participants (age: 58±14 years, 79 [73%] women) from the CCGA study to generate statistical models and data structures of representative fragments. These samples were used to train a Markov chain model (order 3) that estimates the likelihood of the CpG methylation status of a given sequence within a fragment, as described above in section II.B. This model has been shown to be calibrated within the range of normal fragments (p-value > 0.001) and was used to reject fragments with p-values ​​≥ 0.001 obtained from the Markov model as not being sufficiently abnormal.

[0165] A further data reduction step, as described above, selected only fragments that covered at least five CpGs and had a mean methylation above 0.9 (hypermethylated) or below 0.1 (undermethylated). This procedure yielded a median (range) of 2,800 (1,500-12,000) UFXM fragments for training participants without cancer and a median (range) of 3,000 (1,200-220,000) UFXM fragments for training participants with cancer. Because this data reduction procedure only used data from the reference set, this stage only required a single application to each sample.

[0166] VB tissue of origin cancer classification Figures 5-7, 11, 12A, 12B, 16A, 16B, 17, and 18 illustrate a number of graphs showing the accuracy of cancer predictions for various trained cancer classifiers, according to embodiments. The cancer classifiers used to obtain the results shown in Figures 5-7, 11, 12A, 12B, 16A, 16B, 17, and 18 were trained according to the exemplary implementation of process 300 described above in Figure 3A.

[0167] The analysis system selects the CpG sites to be considered in the cancer classifier. The information gain is calculated for a training sample with a given cancer type compared to all other samples. For example, two random variables are used: "abnormal fragment" ("AF") and "cancer type" ("CT"). CT is a random variable that indicates whether the cancer is of a particular type. The analysis system calculates the mutual information on CT given AF, i.e., the number of information bits about the cancer type that are gained when it is known whether there is an abnormal fragment that overlaps a particular CpG site. For a given cancer type, the analysis system uses this information to rank CpG sites based on how cancer-specific they are. This procedure is repeated for all cancer types under consideration. The ranked CpG sites for each cancer type are greedily added for use in the cancer classifier (e.g., to achieve approximately 3,000 CpG sites).

[0168] For sample characterization, the analysis system identifies in each sample fragments with abnormal methylation patterns, and also UFXM fragments. For a sample, the analysis system calculates an abnormality score for each considered CpG site (approximately 3,000). The analysis system defines the abnormality score using a binary scoring based on whether the sample has a UFXM fragment encompassing the CpG site.

[0169] FIG. 5 illustrates a number of graphs showing the cancer prediction accuracy rate of a multi-class cancer classifier for various cancer types according to an exemplary implementation. In this illustrated example, the multi-class cancer classifier is trained to distinguish feature vectors according to eleven cancer types: breast cancer type, colorectal cancer type, esophageal cancer type, head / neck cancer type, hepatobiliary cancer type, lung cancer type, lymphoma cancer type, ovarian cancer type, pancreatic cancer type, non-cancer type, and other cancer type. The samples used in this example were derived from subjects known to have each of the cancer types. For example, a cohort of breast cancer type samples was used to validate the accuracy rate of the cancer classifier in calling breast cancer type. Furthermore, the samples used are derived from subjects at various stages of cancer.

[0170] For the breast, colorectal, and lung cancer cohorts, the cancer classifier became progressively more accurate in correctly predicting cancer type at subsequent cancer stages. For the head / neck, ovarian, and pancreatic cohorts, the cancer classifier had an increasing accuracy rate at later stages, i.e., stage III and / or stage IV. For the esophageal and hepatobiliary cohorts, the cancer classifier also had an accuracy rate at later stages, i.e., stage III and stage IV. For the non-cancer cohorts, the cancer classifier was completely accurate in predicting non-cancer samples as likely to not have cancer. Finally, the lymphoma cohort had success throughout the various stages, with the peak success being in correctly predicting stage II cancer samples.

[0171] FIG. 6 illustrates a number of graphs showing the accuracy rate of cancer prediction of a multi-class cancer classifier for various cancer types after first using a binary cancer classifier according to an exemplary implementation. In this example, the analysis system first inputs samples from multiple cancer type cohorts into a binary cancer classifier to determine whether the sample is likely to have cancer or not. The analysis system then inputs samples determined to be likely to have cancer into a multi-class cancer classifier to predict the cancer type of the sample. The cancer types considered include breast cancer type, colorectal cancer type, esophageal cancer type, head / neck cancer type, hepatobiliary cancer type, lung cancer type, lymphoma cancer type, ovarian cancer type, pancreatic cancer type, and other cancer types.

[0172] Compared to the example in Figure 5, the analysis system shows an increase in accuracy when first using a binary cancer classifier and then using a multi-class cancer classifier. Among the breast cancer cohort, colorectal cancer cohort, lung cancer cohort, and lymphoma cancer cohort, the analysis system had an overall increase in accuracy. Specifically, the analysis system had a notable increase in prediction accuracy for each of the cancer types in early stage cancers, i.e., stage I, stage II, and even stage III.

[0173] FIG. 7 illustrates a confusion matrix showing the performance of a trained cancer classifier according to an exemplary implementation. In one example of training according to process 300, a multi-class kernel logistic regression (KLR) classifier with ridge regression penalty was trained on the derived feature vector with a penalty on the weights and a fixed penalty on the bias term for each cancer type. The ridge regression penalty was optimized (using log loss) on a portion of the training data that was not used to select high relevance locations, and after the optimal parameters were found, the logistic classifier was retrained on the entire set of local training folds. The selected high relevance sites and classifier weights were then applied to new data. Within the CCGA training set, one fold was iteratively held out, and relevant sites were selected in 8 of 9 folds, the hyperparameters of the KLR classifier were optimized on the 9th set, and the KLR was retrained on 9 of 10 folds and applied to the held out folds. This was repeated 10 times to estimate TOO within the CCGA training set. For the CCGA test set, relevant sites were selected in 9 out of 10 folds of the CCGA training, hyperparameters were optimized in the 10th fold, the KLR classifier was retrained on all CCGA training data, and the selected sites and the KLR classifier were applied to the test set. The cancer types considered include multiple myeloma cancer type, colorectal cancer type, lymphoma cancer type, ovarian cancer type, lung, head / neck cancer type, pancreatic cancer type, breast cancer type, hepatobiliary cancer type, esophageal cancer type, and other cancer types. Other cancer types include cancers with less than 5 samples taken in CCGA, such as anorectal, bladder, cancer of unknown primary TOO, cervical, stomach, leukemia, melanoma, prostate, kidney, thyroid, uterus, and other additional cancers.

[0174] The confusion matrix shows the agreement between cancer types with samples having known cancer TOO (along the x-axis) and predicted cancer TOO (along the y-axis). To validate the performance of the trained KLR classifier, a cohort of samples for each cancer type (shown in brackets along the y-axis for each cancer type) was classified with the KLR classifier. The x-axis shows the number of samples from each cohort that were classified under each cancer type. For example, in a lung cancer cohort with 25 samples with known lung cancer, the KLR classifier predicted that one sample had ovarian cancer, 19 samples had lung cancer, two samples had head / neck cancer, one sample had pancreatic cancer, one sample had breast cancer, and one sample was labeled as the other cancer type. Notably, for all cancer types except for the other cancer types, the KLR classifier correctly predicted more than half of the respective cohorts, with particularly high accuracy rates for multiple myeloma (2 / 2 or 100%), colorectal (18 / 20 or 90%), lymphoma (8 / 9 or 88.8%), ovarian (4 / 5 or 80%), lung (19 / 25 or 76%), and head / neck (3 / 4 or 75%) cancer types. These results indicate the prediction accuracy rate of the KLR classifier.

[0175] FIG. 11 illustrates a confusion matrix showing the performance of a cancer classifier trained with additional hematological cancer subtypes in an exemplary implementation. The cancer classifier may be trained according to the principles described above, for example, with respect to the exemplary results of the cancer classifier in FIG. 7. TOO labels for hematological subtypes include Hodgkin's lymphoma (HL), NHL aggressive, NHL indolent, myeloid, circulating lymphoma (or lymphoid), and plasma cell. Note that the accuracy of the classification is 87.5% for 1,076.

[0176] 12A and 12B illustrate graphs showing cancer prediction accuracy rates for multiple cancer types across cancer stages in an exemplary implementation. In this example, a cancer classifier is trained after filtering out non-cancer samples according to the process 1000 described above. The analysis system determined multiple TOO thresholds for hematological subtypes. The analysis system filtered out non-cancer samples with at least one TOO probability at or above the corresponding TOO threshold for the hematological subtype. The graphs shown show the sensitivity of classification across various stages of cancer for anorectal, bladder and urothelial, breast, cervical, colorectal, head and neck, liver and bile duct, lung, melanoma, ovarian, pancreatic and gallbladder, prostate, kidney, sarcoma, thyroid, upper gastrointestinal tract, and uterine cancer types. The graphs for each cancer type show the prediction sensitivity across each stage of the cancer type using the first cancer classifier without TOO thresholding, labeled "locked_v1_orgi", and the second cancer classifier with TOO thresholding, labeled "v2_custom". Notably, for many cancer types, the second cancer classifier has a higher prediction accuracy rate while maintaining a narrow confidence interval, given that more samples are available for validation. Of particular note, there is a high prediction accuracy rate for many cancer types at the stage I and II levels, indicating the improvement of predictive ability with TOO thresholding in early stage cancers.

[0177] Exemplary Hematology-Specific Classifiers Figure 16A, Figure 16B, Figure 17, and Figure 18 illustrate graphs showing the cancer prediction accuracy rate of the hematology-specific cancer classifier according to the first exemplary implementation. First, cfDNA samples were accessed from the second pre-specified sub-study of CCGA, which was designed for targeted methylation assay validation. In this example, only samples from the training set were used, and tumor tissue samples obtained from an in-house tissue biopsy reference database were included in the training of the classification model. Specifically, the samples used to train the custom classification model of hematological malignancies were obtained from participants recruited with a hematological cancer diagnosis (cancer cases) and participants recruited without a cancer diagnosis (non-cancer controls). 154 blood cell samples or tissue FFPE samples of hematological malignancies were also included. In total, cfDNA samples from 185 participants with hematological cancer and 1,998 non-cancer controls confirmed to have no cancer diagnosis at one-year follow-up were included in the performance evaluation.

[0178] In this first exemplary implementation, a cancer classifier is trained to distinguish between five hematological subtypes and the absence of cancer ("non-cancer"). The five hematological subtypes are myeloid neoplasm, non-Hodgkin's lymphoma (NHL), circulating lymphoma, plasma cell neoplasm, and Hodgkin's lymphoma (HL). A cross-validated mutual information based algorithm was used to identify features that distinguish between the five hematological subtypes and a control class. A multinomial classifier was then trained to detect the presence or absence of cancer and predict tissue of origin among the five hematological cancers and non-cancer using 6-fold cross-validation.

[0179] The trained hematology-specific classifier achieved an overall specificity of greater than 99.4% [95% confidence interval (CI): 99.0-99.7%] in controls and a sensitivity of 74.3% [95% CI: 67.4-80.5%] for hematological cancers. Figure 16A illustrates a graph showing the sensitivity of the classifier at the 99.5% specificity level across hematological subtypes. The sensitivity of the hematology-specific cancer classifier for each hematological subtype is arranged in ascending order, the numbers in the class labels indicate the number of samples, and the error bars indicate the 95% confidence interval. Myeloid neoplasms, for which four samples were classified, had a sensitivity just below 50% with a wide 95% confidence interval of about 10% to about 90%. This low sensitivity is due to the limited samples used for training. In the embodiment of Figure 16A, NHL, circulating lymphomas, plasma cell neoplasms, and HL have a sensitivity of approximately 70% to 87%, which is better than myeloid neoplasms. Specifically, the sensitivity by hematological subtype was 45.8% [95% CI: 5.3-91.6%] for myeloid neoplasms, 76.5% [95% CI: 61.3-88.0%] for circulating lymphomas, 86.1% [95% CI: 54.7-98.7%] for Hodgkin's lymphomas, 71.3% [95% CI: 60.8-80.3%] for other non-Hodgkin's lymphomas, and 78.9% [95% CI: 61.6-91.0%] for plasma cell neoplasms.

[0180] FIG. 16B illustrates a graph showing the sensitivity of the classifier at 95% specificity across stages of Hodgkin's and non-Hodgkin's lymphoma. Stage I sensitivity (in 15 samples) is about 25%. Stage II sensitivity (in 27 samples) is about 85%. Stage III sensitivity (in 27 samples) is about 75%. Stage IV sensitivity (in 32 samples) is about 85%. The graph shows the dramatic increase in sensitivity of the hematology-specific cancer classifier between stage I and stage II (and beyond). Specifically, for lymphomas (HL and NHL), the stage-specific sensitivity was 25.6% [95% CI: 7.2-54.0%] for stage I, 84.6% [95% CI: 65.5-95.5%] for stage II, 72.8% [95% CI: 52.4-88.0%] for stage III, and 83.9% [95% CI: 66.6-94.4%] for stage IV.

[0181] FIG. 17 illustrates a confusion matrix showing the cancer prediction accuracy rate of the hematology-specific cancer classifier in the first exemplary implementation. The numbers in each box represent the total number of predicted samples. In the confusion matrix, the coloring / shading corresponds to the proportion of predicted hematological subtypes, as shown on the right side of the plot. Similarly, the percentage of correct predictions is shown on the right side of the graph. The tissue of origin localization was evaluated in cancer cases correctly detected as hematological cancer by the TOO multi-class classifier. As shown in FIG. 17, the hematology-specific classifier achieved an overall TOO prediction accuracy rate of 87.7%, with the highest prediction accuracy rate (100%) for Hodgkin's lymphoma and myeloid neoplasms, followed by plasma cell neoplasms (96.4%), non-Hodgkin's lymphomas (85.9%), and circulating lymphomas (80%). Of the 11 non-cancer controls classified as having hematologic cancer (0.55% of non-cancer controls), five were predicted as circulating lymphoma and six were predicted as other non-Hodgkin lymphoma (false-positive rate <1%), with most showing reliable TOO signals localized to the predicted hematologic class (>50% of the total probability volume).

[0182] In some instances, a low-dimensional representation of methylation features useful for the final classifier can be generated using the UMAP method, which preserves the topology of the high-dimensional data. For example, in Figure 9B discussed above, the UMAP embedding shows that the majority of hematological malignancies were separated into five major clusters reflecting developmental lineage and disease ontology. The majority of non-cancer controls (shown using contour density in Figure 9B) were clustered separately from the hematological cancers.

[0183] Scatter plots were then used to visualize the relationship between the spatial distribution of feature embeddings and the class probabilities of the model. For example, Figure 18 illustrates a series of graphs plotting cancer score against distance from the centroid in the UMAP embedding in an exemplary implementation. The UMAP embedding is the same as the UMAP embedding in Figure 9B. The x-axis plots the logit-transformed probability that a sample has cancer, i.e., the logit of the cancer score. The logit function (also called log-odds) is the logit function of the odds

[0184]

number

[0185] , where p is the probability. The y-axis plots the Euclidean distance from the centroid of the UMAP embedding. These plots depict the correlation between cancer scores and localization in the UMAP embedding for various hematological subtypes. Graph 1810 represents the correlation in myeloid neoplasm subtypes. Graph 1820 represents the correlation in NHL subtypes. Graph 1830 represents the correlation in circulating lymphoma subtypes. Graph 1840 represents the correlation in HL subtypes. Graph 1850 represents the correlation in plasma cell neoplasm subtypes. Graph 1860 represents minimal correlation in non-cancer samples. As shown in FIG. 18, there was a strong positive correlation between their UMAP embedding localization and classification scores.

[0186] As shown by the exemplary hematology-specific classifier, a custom classifier for hematological malignancies may provide a convenient means for simultaneously detecting and differentiating the five major hematological malignancies, thereby facilitating clinical diagnosis and treatment selection. In this manner, the custom classifier may achieve more sensitive detection of multiple cancers, which may be used to refine the cancer detection and TOO prediction accuracy rates.

[0187] Figure 19, Figure 20, and Figure 21 illustrate graphs showing the cancer prediction accuracy rate of the hematology-specific cancer classifier according to the second exemplary implementation. First, cfDNA samples were accessed from the second pre-specified sub-study of CCGA, designed for targeted methylation assay validation. In this example, only samples from the training set were used, and tumor tissue samples obtained from an in-house tissue biopsy reference database were included in the training of the classification model. Specifically, the samples used to train the custom classification model for hematological malignancies were obtained from participants recruited with a hematological cancer diagnosis (cancer cases) and participants recruited without a cancer diagnosis (non-cancer controls). In total, cfDNA samples obtained from 534 participants with hematological cancer were included in the performance evaluation.

[0188] In this second exemplary implementation, a cancer classifier is trained to distinguish between seven hematological subtypes and the absence of cancer ("non-cancer"). The seven hematological subtypes are myeloid neoplasm, non-Hodgkin's lymphoma (NHL), circulating lymphoma, plasma cell neoplasm, Hodgkin's lymphoma (HL), heme_1, and heme_3. Subtypes heme_1 and heme_3 refer to two types of hematological precursor states that may develop into hematological cancers, such as other hematological subtypes. Hematological precursor states may include, but are not limited to, monoclonal gammopathy of undetermined significance or monoclonal B-cell lymphocytosis. A cross-validated mutual information based algorithm was used to identify features that distinguish between the seven hematological subtypes and the non-cancer class. A multinomial classifier was then trained to detect the presence or absence of cancer and predict tissue of origin among the five hematological cancers and non-cancer using 6-fold cross-validation.

[0189] FIG. 19 illustrates a graph plotting the anomaly scores of multiple training samples for hematology-specific cancer classification. With reference to the graph, "M" refers to myeloid neoplasm hematological subtype, "H3" refers to heme_3 hematological subtype, "HL" refers to Hodgkin's lymphoma hematological subtype, "nHL" refers to non-Hodgkin's lymphoma subtype, "CL" refers to circulating lymphoma hematological subtype, "H1" refers to heme_1 hematological subtype, and "P" refers to plasma cell neoplasm hematological subtype. On the y-axis are a number of top features that are selected for use according to mutual information gain. Thus, the first column shows each hematological subtype compared against each of the other hematological subtypes shown in the second column. On the x-axis are training samples grouped by known hematological subtype. For example, under the column "nHL" are training samples that are known to be labeled as non-Hodgkin's lymphoma hematological subtype. The analysis system determines an anomaly score for each selected feature. In this implementation, the anomaly score is a binary score based on the presence (shown in white) or absence (shown in grey) of aberrantly methylated fragments covering the feature. Of note, the white area along the main diagonal (top left to bottom right) provides an indication for the discriminatory power in classifying hematological subtypes. If within a row, a feature is spread in white across different samples from different hematological subtypes, there is an indication that the feature has low discriminatory power and is noisy.

[0190] FIG. 20 illustrates a graph showing the sensitivity of the hematology-specific cancer classifier at 99.5% specificity. The set of data on the left for each hematology subtype is for the training set used to train the hematology-specific cancer classifier, while the set of data on the right is for the holdout set. The number of samples present in the training set and the holdout set, respectively, is noted on the lower x-axis after the label of each hematology subtype. 95% confidence intervals are shown for the sensitivity measured in the training set and the holdout set for each subtype. The heme_1 subtype had low sensitivity for both the training set and the holdout set. The heme_3 subtype had a sensitivity of about 25% for both the training set and the holdout set. The myeloid neoplasm subtype had a sensitivity of 50% for the training set (correctly predicted one of two training samples) and 100% for the holdout set (correctly predicted one of one holdout sample). The circulating lymphoma subtype had a sensitivity of about 70% for both sets. The non-Hodgkin's lymphoma subtype had a sensitivity of about 70% for the training set and about 75% for the holdout set. The plasma cell neoplasm subtype had a sensitivity of about 75% for both sets. The Hodgkin's lymphoma subtype had a sensitivity of about 80% for the training set and about 70% for the holdout set.

[0191] FIG. 21 illustrates a confusion matrix showing the cancer prediction accuracy rate of the hematology-specific cancer classifier in the second exemplary implementation. The numbers in each box represent the total number of predicted samples. In the confusion matrix, the coloring / shading corresponds to the proportion of predicted hematological subtypes, as shown on the right side of the plot. Similarly, the percentage of correct predictions is shown on the right side of the graph. The tissue of origin localization was evaluated in cancer cases correctly detected as hematological cancer by the TOO multi-class classifier. As shown in FIG. 21, the hematology-specific classifier achieved an overall TOO prediction accuracy rate of about 75%. The plasma cell neoplasm subtype correctly predicted 17 out of 17 known samples, with a prediction accuracy rate of 100%. The Heme_1 subtype correctly predicted 1 out of 4 known samples, with a prediction accuracy rate of 25%. The circulating lymphoma subtype correctly predicted 25 of 27 known samples with a predictive accuracy of 92.6%. The non-Hodgkin's lymphoma subtype correctly predicted 48 of 55 known samples with a predictive accuracy of 87.3%. The Hodgkin's lymphoma subtype correctly predicted 8 of 8 known samples with a predictive accuracy of 100%. The Heme_3 subtype correctly predicted 19 of 20 known samples with a predictive accuracy of 95%. The myeloid neoplasm subtype correctly predicted 1 of 1 known samples with a predictive accuracy of 100%.

[0192] As shown by the second exemplary hematology-specific classifier, the custom classifier for hematological malignancies and hematological precursors can also identify such precursors that eventually transform into hematological malignancies.This precursor classification ability is shown to be useful for identifying individuals who will develop hematological malignancies in the future, which can lead to earlier clinical diagnosis and treatment selection.In this manner, the custom classifier can achieve more sensitive detection of multiple cancers, which can be used to refine the accuracy rate of cancer detection and TOO prediction.

[0193] VI. Additional Considerations The detailed description of the foregoing embodiments refers to the accompanying drawings which illustrate certain embodiments of the present disclosure. Other embodiments having different structures and operations do not depart from the scope of the present disclosure. Terms such as "the present invention" are used with reference to certain specific examples of the many alternative aspects or embodiments of Applicants' invention described herein, and neither their use nor their absence is intended to limit the scope of Applicants' invention or the claims.

[0194] An embodiment of the present invention also relates to an apparatus for performing the operations herein. This apparatus may be specially constructed for the required purposes and / or it may include a general-purpose computing device selectively activated or reconfigured by a computer program stored in the computer. Such a computer program may be stored in a non-transitory tangible computer-readable storage medium or any type of medium suitable for storing electronic instructions, the medium may be coupled to a computer system bus. Furthermore, any computing system referred to herein may include a single processor or may be an architecture using a multiple processor design for increased computing power.

[0195] Any of the steps, operations, or processes described herein as being performed by the analysis system may be performed or implemented using one or more hardware or software modules of the apparatus, alone or in combination with other computing devices. In one embodiment, the software modules are implemented using a computer program product that includes a computer readable medium containing computer program code that may be executed by a computer processor to perform any or all of the steps, operations, or processes described.

Claims

1. 1. A method for predicting the presence or absence of cancer in a test sample, comprising: accessing the test sample having a tissue signal related to a cancer score and a first tissue label; selecting one of a plurality of tiers based on the tissue signal for the first tissue label, the plurality of tiers including a tier of high signal for the first tissue label and a tier of low signal for the first tissue label; predicting whether the test sample is associated with the presence or absence of cancer by comparing the cancer score against a binary threshold cutoff for the selected stratum; The method according to claim 1, further comprising:

2. 2. The method of claim 1, wherein the test sample comprises a test feature vector determined according to methylation sequencing data of the test sample.

3. 3. The method of claim 2, wherein the cancer score is determined by applying a binary cancer classifier to the test feature vector.

4. 3. The method of claim 2, wherein the tissue signal is a tissue of origin (TOO) prediction determined by applying a multi-class cancer classifier to the test feature vector.

5. 5. The method of claim 4, wherein the TOO predictions include a predicted value for each of a plurality of tissue labels, each predicted value indicating the likelihood that the test sample corresponds to a cancer type associated with the tissue label.

6. The step of selecting one of a plurality of layers based on the tissue signal for the first tissue label comprises: determining whether the tissue signal for the first tissue label is at or above a predictive value threshold; selecting the high signal tier in response to determining that the tissue signal for the first tissue label is at or above the predicted value threshold; selecting the low signal stratum in response to determining that the tissue signal for the first tissue label is below the predictor threshold; and 6. The method of claim 5, comprising:

7. 5. The method of claim 4, wherein the TOO prediction indicates one or more top predictions for one or more tissue labels of a plurality of tissue labels, a top prediction of a tissue label indicating that the test sample is predicted to have a cancer type associated with the tissue label of the top prediction.

8. The step of selecting one of the plurality of layers comprises: determining whether the first tissue label is a top prediction; selecting the high signal tier in response to determining that the first tissue label is the top predictive; and selecting the low signal tier in response to determining that the first tissue label is not the top predictive; and 8. The method of claim 7, comprising:

9. The step of selecting one of the plurality of layers includes: determining whether the first tissue label is a second top prediction; selecting the high signal tier in response to determining that the first tissue label is the second top predictive; and selecting the low signal tier in response to determining that the first tissue label is not the second top predictive; and 9. The method of claim 8, comprising:

10. The method of claim 1 , wherein the plurality of layers includes a medium signal layer for a medium tissue signal.

11. 2. The method of claim 1, wherein the test sample has a tissue signal for a second tissue class, and selecting one of the plurality of layers is further based on the tissue signal for a second tissue label.

12. The binary threshold cutoff for each stratum is: obtaining a holdout set of samples, each sample having a cancer score and a tissue signal for said first tissue label; stratifying the holdout set of samples into the plurality of strata based on the tissue signal for the first tissue label of the holdout set of samples; For each of the plurality of layers, sweeping across a domain of cancer scores at a plurality of candidate binary threshold cutoffs by calculating a true positive rate and a false positive rate for each binary threshold cutoff based on the cancer scores of the samples in the stratum; selecting a binary threshold cutoff from the plurality of candidate binary threshold cutoffs for the tier based on a false positive budget for the tier and the calculated false positive rate; 2. The method of claim 1, wherein the value is determined by:

13. A system comprising: a hardware processor; and a non-transitory computer-readable storage medium storing executable instructions that, when executed by the hardware processor, cause the hardware processor to perform steps comprising a method according to any one of claims 1 to 12.

Images