Methods for determining cancer

Abstract

The present invention relates to a computer-implemented method for determining a primary cancer of a malignancy of unknown origin (MUO) in a subject. The method comprises: i) providing sequence data of cell-free DNA (cfDNA) obtained from a liquid sample from the subject, the sequence data comprising regional methylation features (“sample data”); ii) providing the sample data as input to a machine learning model comprising a multi-class classifier trained on a training dataset, the training dataset comprising regional methylation features from at least two primary malignancy types (“training data”), each primary malignancy type having a different location and / or histological subtype; iii) receiving an output from the machine learning model, the output indicating a primary malignancy type classification or an unknown type classification of the sample data; and iv) determining the primary cancer of the subject based on the output received in step (iii). The present invention also relates to a method of selecting a cancer treatment for a subject having a MUO and a method of treating a subject having a MUO. Systems and computer readable media are also provided.

Description

Technical Field

[0001] This invention relates to a computer-implemented method for identifying primary cancer of malignancy of unknown origin (MUO) in a subject. The invention also relates to a method for selecting cancer treatment for a subject with MUO and a method for treating a subject with MUO. Systems and computer-readable media are also provided. Background Technology

[0002] Malignant tumors of unknown origin (MUO) are a clinical condition representing a heterogeneous group of malignant tumors in which the most probable cancer cells have been identified in a subject, but the primary origin of these cancer cells is unknown. This condition also encompasses cancers of unknown primary (CUP), which represents a heterogeneous group of cancers in which secondary cancerous tumors have been identified, but the primary cancer remains unknown after standard clinical investigation. MUO and CUP account for approximately 1% to 3% of all cancer diagnoses.

[0003] Treatment options for MUO and CUP remain limited in the absence of a clear identification of the primary source of the cancer cells. Many individuals with CUP receive standard, "one-size-fits-all" chemotherapy, despite significant clinical, pathological, and molecular heterogeneity. Because treatments are not targeted at an unknown primary source, they typically have limited efficacy at best.

[0004] Current guidelines identify approximately 20% of patients with CUP as having a “favorable CUP.” This is based on clinicopathological features corresponding to cancer types that are more responsive to treatment. Even for those classified as having a favorable CUP, obtaining a diagnosis of the primary tumor can take a considerable amount of time, involving multiple investigations and pathological reviews from a very small amount of tissue obtained from an invasive tumor biopsy. The remaining 80% of patients with CUP are classified as having a “disadvantageous CUP,” which is considered to have a more negative prognosis. Therefore, CUP remains a common cause of cancer-related death. In fact, in England, only about 15% of patients with CUP survive for one year or longer, and only about 10% survive for three years or longer.

[0005] The role of biomarker-driven precision oncology and the emergence of immunotherapy are rapidly transforming standard-of-care (SOC) treatment and improving overall survival across a wide range of tumor types. However, these approaches remain largely inaccessible to individuals with chronic uterine artery disease (CUP), particularly those with unfavorable CUP. With the exception of a few tumor-agnostic treatments, most targeted therapies have demonstrated tumor-type-dependent potency, exemplified by the activity of targeted inhibitors in B-RAF-mutant melanoma compared to their inactivity in colorectal cancer. Furthermore, immunotherapy is increasingly indicated by the presence of tumor-type-validated biomarkers.

[0006] Molecular characterization methods for predicting the tissue of origin (also known as the tumor of origin) in patients with chronic ulcerative colitis (CUP) are considered a pathway to better treatment stratification. However, the scarcity of tumor tissue in patients with CUP means that the detection of the tumor of origin remains challenging. Furthermore, current tissue of origin prediction does not necessarily identify a specific primary cancer, given that cancers from the same tissue of origin can differ due to clinical, pathological, and molecular heterogeneity. Therefore, improvements in the diagnosis of MUO, particularly CUP, remain necessary. Accordingly, this should provide clinicians with more treatment options that improve therapeutic efficacy.

[0007] The present invention was designed in view of the above considerations. Summary of the Invention

[0008] In summary, the inventors have developed an improved method for identifying primary cancer in unexplained malignant tumors (MUO) of a subject. Since obtaining sufficient tumor tissue for molecular profiling analysis in MUO and CUP remains challenging, this invention utilizes a liquid biopsy approach to predict the primary cancer of a subject. This provides improved detection sensitivity due to the increased availability and quality of tumor tissue and associated DNA from the liquid biopsy sample. Computer-based analysis of regional methylation signatures in the liquid biopsy advantageously enables accurate stratification of the subject within a rapid turnaround time. Accordingly, this allows for the selection of treatment options targeting the primary cancer of the subject, thereby improving treatment efficacy. Furthermore, rapid identification of the primary cancer of the subject allows clinicians to administer treatment earlier than previously possible, further improving treatment efficacy. In some embodiments, the method may be referred to as “CUPiD” (Cancer of Unknown Primary Identification).

[0009] Therefore, in a first aspect, the present invention provides a computer-implemented method for determining the primary carcinoma of a malignant tumor of unknown origin (MUO) of a subject, the method comprising:

[0010] i) Provide sequence data of cell-free DNA (cfDNA) obtained from a liquid sample from the subject, the sequence data including regional methylation features (“sample data”).

[0011] ii) Feeding sample data as input to a machine learning model, the machine learning model comprising a multi-class classifier trained on a training dataset containing regional methylation features (“training data”) from at least two primary malignant tumor types, each primary malignant tumor type having a different location and / or histological subtype;

[0012] iii) Receive output from a machine learning model, the output indicating the primary malignant tumor type classification or unknown type classification of the sample data; and

[0013] iv) Determine the primary cancer of the object based on the output received in step (iii).

[0014] In the context of this invention, the term "regional methylation signature" will be understood to refer to the DNA methylation state of at least one region of the genome. A genomic region will be understood to refer to a portion of the genome, which may also be referred to as a genomic region. Generally, "DNA methylation state" will be understood to refer to the level of DNA methylation. DNA methylation can be either hypermethylation or hypomethylation.

[0015] In some implementations, liquid samples include blood, urine, or plasma samples.

[0016] In some implementations, the machine learning model includes decision trees, logistic regression models, artificial neural networks, support vector machines (SVMs), Naive Bayes, or k-nearest neighbors algorithms, optionally wherein the decision tree includes gradient boosting or random forest algorithms.

[0017] A machine learning model may contain at least 60 individual classifiers, and optionally at least 100 individual classifiers.

[0018] In some implementations, at least two primary malignant tumor types include at least three primary malignant tumor types, optionally at least four primary malignant tumor types, wherein each primary malignant tumor type has a different location and / or histological subtype.

[0019] In some implementation schemes, at least two primary malignant tumor types are selected from:

[0020] a) Primary malignant hepato-pancreatobiliary tumor;

[0021] b) Primary malignant female reproductive tract tumors;

[0022] c) Primary malignant upper or lower gastrointestinal tumors;

[0023] d) Primary malignant lung tumor;

[0024] e) Primary malignant breast tumor; and

[0025] f) Primary malignant urological tumors.

[0026] At least one of the two primary malignant tumor types may include primary malignant hepatopancreatic biliary tumors, optionally primary malignant cholangiocarcinoma tumors.

[0027] In some implementations, at least two primary malignant tumor types include at least 20 primary malignant tumor types, each with a different location and / or histological subtype.

[0028] In some implementation schemes, at least two primary malignant tumor types are selected from: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), invasive breast carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynecological carcinoma (Gynae), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), acute myeloid leukemia (LAML), low-grade glioma (LGG), and liver hepatocellular carcinoma. The following are cancer cell carcinomas: LIHC (lower GI), Lung adenocarcinoma (LUAD), Lung squamous cell carcinoma (LUSC), Mesothelioma (MESO), Pancreatic adenocarcinoma (PAAD), Pheochromocytoma and paraganglioma (PCPG), Prostate adenocarcinoma (PRAD), Sarcoma (SARC), Skin cutaneous melanoma (SKCM), Testicular germ cell tumor (TGCT), Thyroid carcinoma (THCA), Thymoma (THYM), Gastric and esophageal adenocarcinoma (upper GI), Head and neck and esophageal squamous cell carcinoma (upper Sq), and Uveal melanoma (UVM).

[0029] In some implementations, the training data also includes regional methylation features from at least one non-cancer control. It should be understood that regional methylation features from non-cancer controls include the DNA methylation status of at least one region of the genome obtained from samples from non-cancer controls.

[0030] In some implementations, the output includes a probability score for each category.

[0031] In some implementations, the methylation signature of each region includes the DNA methylation status of multiple genomic regions. Each genomic region may contain approximately 200 to approximately 2000 base pairs in length.

[0032] It should be understood that the training data includes regional methylation features from one primary malignant tumor type and regional methylation features from at least one other primary malignant tumor type. Typically, regional methylation features from each primary malignant tumor type include the same genomic regions. By "same region," this will be understood as referring to genomic portions with the same genomic coordinates and length. However, it should be understood that the DNA methylation status of each region may differ between each primary malignant tumor type. Therefore, the regional methylation features for each primary malignant tumor type are unique "features" for each primary malignant tumor type.

[0033] In some embodiments, the regional methylation signatures from each primary malignant tumor type include the DNA methylation status of at least 5, at least 10, at least 20, at least 30, at least 50, or at least 100 genomic regions. The genomic regions may be selected from the genomic regions listed in Table 1 (which may also be referred to as DMRs). In some embodiments, the genomic regions are selected from the genomic regions listed in Table 2. In some embodiments, the genomic regions include at least the first five genomic regions listed in Table 2. In some embodiments, the genomic regions include all the genomic regions listed in Table 2.

[0034] In some embodiments, the genomic regions are selected from the genomic regions listed in Table 5. In some embodiments, the genomic regions are selected from the list in Table 6. In some embodiments, the genomic regions include at least the first five genomic regions listed in Table 6. In some embodiments, the genomic regions are selected from the genomic regions listed in Table 7.

[0035] In some implementations, the genomic regions in each region methylation feature of the training data include genomic regions identified from pairwise comparisons of genomic region DNA methylation states between at least two primary malignant tumor types.

[0036] The genomic regions in the training data may include at least one, at least five, at least ten, at least twenty, at least thirty, at least fifty, at least one hundred, at least one hundred and fifty-five genomic regions with the greatest differences in methylation status identified from pairwise comparisons.

[0037] In some implementations, genomic regions are selected based on a function of similarity between each type of primary malignancy.

[0038] Training data may include regional methylation features derived from primary malignant tumor tissue samples obtained from multiple individuals, each with a known primary malignant cancer.

[0039] In some implementations, the regional methylation features for each primary malignant tumor type in the training data include the DNA methylation status of at least one genomic region derived from a primary malignant tumor tissue sample and the DNA methylation status of at least one genomic region derived from at least one non-cancer control cfDNA.

[0040] In some embodiments, the sample data also includes a tumor fraction (TF) of cell-free DNA (cfDNA) obtained from the liquid sample. In some embodiments, the tumor fraction (TF) is at least about 3%.

[0041] In some embodiments, MUO includes cancer of unknown primary origin (CUP). The subject may have an unfavorable CUP. In some embodiments, the subject is known to have secondary cancer, but the primary cancer is unknown. In some embodiments, the subject is known to have secondary adenocarcinoma, secondary squamous cell carcinoma, or secondary poorly differentiated carcinoma, but the primary cancer is unknown.

[0042] In some implementation schemes, the primary cancer in step iv) is selected from: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), invasive breast carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynecological carcinoma (Gynae), chromophobe renal cell carcinoma (KICH), clear cell renal cell carcinoma (KIRC), papillary renal cell carcinoma (KIRP), acute myeloid leukemia (LAML), low-grade glioma of the brain (LGG), liver cell carcinoma, etc. Primary cancers include: Lipocarcinoma of the lung (LIHC), colon and rectum (lower GI), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), mesothelioma (MESO), pancreatic adenocarcinoma (PAAD), pheochromocytoma and paraganglioma (PCPG), prostate adenocarcinoma (PRAD), sarcoma (SARC), cutaneous melanoma (SKCM), testicular germ cell tumor (TGCT), thyroid cancer (THCA), thymoma (THYM), gastric and esophageal adenocarcinoma (upper GI), head, neck and esophageal squamous cell carcinoma (upper Sq), uveal melanoma (UVM), and unclassified primary cancers.

[0043] According to a second aspect, the present invention provides a method for identifying primary cancer in a subject having a malignant tumor of unknown origin (MUO), the method comprising:

[0044] i) Analyze cell-free DNA (cfDNA) obtained from liquid samples from the subject to obtain sequence data containing regional methylation features (“sample data”); and

[0045] ii) The sample data are subjected to a first aspect method to determine the primary cancer of the subject.

[0046] According to a third aspect, the present invention provides a method for selecting cancer treatment for a subject with a malignant tumor of unknown origin (MUO), the method comprising:

[0047] i) Using the first approach to determine the most probable primary cancer in the subject; and

[0048] ii) Select anticancer therapy that targets the most likely primary cancer.

[0049] According to another aspect, the present invention provides a method for treating a subject with a malignant tumor of unknown origin (MUO), the method comprising:

[0050] i) Using the first approach to determine the most probable primary cancer in the subject; and

[0051] ii) Administer anticancer therapy targeting the most likely primary cancer.

[0052] According to another aspect, the present invention provides a system comprising:

[0053] i) processor; and

[0054] ii) A computer-readable medium containing instructions that, when executed by a processor, cause the processor to perform the steps of the method of the first aspect.

[0055] The present invention also provides one or more computer-readable media comprising instructions that, when executed by one or more processors, cause one or more processors to perform the steps of the method of the first aspect.

[0056] The present invention includes combinations of the described aspects and preferred features, unless such combinations are clearly unacceptable or explicitly avoided. Attached Figure Description

[0057] The embodiments and experiments illustrating the principles of the present invention will now be discussed with reference to the accompanying drawings, wherein:

[0058] Figure 1 A flowchart for the diagnostic classification of the CUP cohort following a retrospective review of clinical data.

[0059] Figure 2 : A schematic diagram of the development of the CUPID classifier.

[0060] Figure 3 : A chart showing the number of arrays used in each cancer category.

[0061] Figure 4 Example volcano plot, showing the β-value differences and false detection rate-adjusted p-values ​​of 59,918 differentially methylated regions found between 79 ACC and 409 BLCA transformation arrays. The 250 highlight windows with the largest β-value differences between each class were selected to construct the classifier.

[0062] Figure 5 A two-dimensional UMAP was applied to 9,017 transformed methylation arrays, using 22,179 differentially methylated regions (DMRs) selected by tumor category comparison. Category labels were superimposed on the centroid of the category members.

[0063] Figure 6A) Area Under the Receiver Operator Curve (AUROC or AUC) values ​​of 100 individual classifiers, evaluated on held-out mixes not used to train the classifiers (10,611–11,508 held-out mixes per classifier). Cancer category abbreviations are defined in Table 3. B) Receiver Operator Curve (ROC) for each category, evaluated using the ensemble classifier applied to held-out mixes for each category.

[0064] Figure 7 A) Number of T7-MBD-Seq samples in each category in the independent cfDNA testing cohort. B) Relative enrichment score (relH) of methylated enriched samples (MeCap) versus non-enriched samples among the 170 cfDNA samples in the testing cohort.

[0065] Figure 8 Performance of CUPiD in a test cohort of 170 cfDNA samples (composed of 143 samples from known primary tumor types and 27 non-cancer control groups). (This indicates that CUPiD prediction)

[0066] Figure 9 A) Tumor fractions estimated from 170 cfDNA samples using ichorCNA via shallow whole-genome sequencing of non-enriched samples. 3% detection limit cutoff is shown. B) Consistency between CUPID predictions and tumor fractions.

[0067] Figure 10 A mutation profile analysis of 641 gene targets in the cfDNA of 40 CUP patients was performed and compared with matched germlines. The oncoplot shows mutations in genes tagged as oncogenic by oncoKB, with actionable mutations highlighted by an inserted asterisk. The top plot shows the mean variant allele frequency (VAF) of all identified alterations for each patient. The bottom plot is annotated according to subsequent primary diagnosis and CUPID prediction. Cancer category abbreviations are defined in Table 3.

[0068] Figure 11 The alluvial plot shows how tumor type-enriched alteration (TTEA) (left) and CUPiD prediction (right) correspond to clinical classification (center).

[0069] Figure 12In cfDNA from 41 CUP patients, the tumor fraction estimated by ichorCNA and determined from non-enriched samples was correlated with the Pearson correlation of the mean variant allele frequency (VAF) of mutations identified by sequencing of 641 gene panels.

[0070] Figure 13 Sankey plot: Correlation between CUPiD prediction and subsequent diagnosis or clinical suspicion of primary tumor.

[0071] Figure 14 Estimated tumor scores (from ichorCNA) from 41 cfDNA samples from CUP patients, grouped by CUP predicted status and colored by predicted category. Dashed lines represent 3% tumor scores. Category abbreviations are defined in Table 3.

[0072] Figure 15 Tumor type predictions made by CUPID in the CUP cohort (n=32), excluding cases where no predictions were made.

[0073] Figure 16 Swimmers plot of fifteen CUP patients with clinically resolved primary tumor diagnoses. Timeline of the diagnostic investigation from CUP diagnosis to death or data lock. Final primary tumor diagnosis time is color-coded. Annotations by final diagnosis, CUPiD prediction, and concordance.

[0074] Figure 17 A) For different N values, the AUC values ​​of 100 individual sub-classifiers, each constructed using the N most important DMRs from the full CUPID classifier. Each sub-classifier was evaluated on a holdover mix of samples not used to train that sub-classifier. Black dots indicate the AUC of the ensemble classifier. B) These classifiers were applied to a test cohort comprising 170 cfDNA samples.

[0075] Figure 18 A) For different N values, the AUC values ​​of 100 individual sub-classifiers, each sub-classifier constructed using the N most significant DMRs between each comparison. Each sub-classifier was evaluated on a holdover mix of samples not used to train that sub-classifier. Black dots indicate the AUC of the ensemble classifier. B) These classifiers were applied to a test cohort comprising 170 cfDNA samples.

[0076] Figure 19A) After filtering out the 22,179 windows present in the CUPID classifier, AUC values ​​for 100 individual sub-classifiers were calculated for different N values, with each sub-classifier constructed using the N most significant DMRs between each comparison. Each sub-classifier was evaluated on a retainer mix of samples not used to train that sub-classifier. Black dots indicate the AUC of the ensemble classifier. B) These classifiers were applied to a test cohort comprising 170 cfDNA samples.

[0077] Figure 20 A) Comparison of AUC performance between the non-CUPiD classifier and the CUPID classifier, where the non-CUPiD classifier only uses windows not used in the CUPID classifier. B) Comparison of performance between the non-CUPiD classifier and the CUPID classifier on a cohort of 170 sample cfDNA tests.

[0078] Figure 21 Performance of CUPID v1.1 in test cohorts of known cancer samples (n=143, “known”), 167 cfDNA samples (“combinations”) (consisting of 143 known samples from the “known” cohort plus an additional 24 known samples from skin melanoma and breast cancer, labeled “combinations”), and 41 CUP subjects (“CUP”). The sensitivity and false negative tumor rate (FNTR) for the known TARGET cohort were 0.84 and 0.03, respectively. The sensitivity and FNTR for the combination cohort (n=167) were 0.86 and 0.02, respectively. The sensitivity and FNTR for the CUP cohort (n=41) were 0.78 and 0.02, respectively.

[0079] Figure 22A broader output classification is used to determine the most likely primary cancer. A broad cancer output group is applied to the CUPIDv1.1 model. Each bar represents a single object. The Y-axis represents the prediction score. A sample is correctly classified if the cumulative cancer type prediction score of 0.5 or higher falls within the correct broad category. The dashed line indicates a classification threshold of 0.5. The broad cancer output groups shown are: (A) Breast Cancer: The left image shows samples correctly classified in the breast cancer broad category. The right image shows samples incorrectly classified in the breast cancer broad category, where the correct tumor type is lung cancer; (B) Female Reproductive Cancer: The right image shows samples correctly classified in this broad category; (C) Kidney Cancer: The right image shows samples correctly classified in the kidney cancer broad category; (D) Lower Abdominal Cancer: The right image shows samples correctly classified in this category; (E) Lung Cancer: The right image shows samples correctly classified in this category; (F) Samples that remain unclassified using the broad classification method. The sum of all predicted scores within the broad category remains below 0.5, (G) rare cancer; the figure shows samples correctly classified in this category and (H) upper abdominal cancer; the left figure shows samples correctly classified in the broad category, and the right figure shows samples incorrectly classified in the broad category, where the correct tumor type is lung cancer. The right side of the figures in (A) to (H) shows a legend showing a more specific classification. (I) shows a bar chart of predicted scores for four samples whose predicted status changed when the broad classification method was applied to a known cancer cohort. The top figure shows the broad category for lung, and the bottom figure shows the broad category for upper abdomen. Cancer category abbreviations are defined in Table 3.

[0080] Figure 23 Broad output classification improves the sensitivity of CUPID v1.1.

[0081] Figure 24 For different N values, the AUC value of the CUPID v1.1 ensemble classifier is constructed using the N most significant DMRs between each pairwise comparison, as detailed on the X-axis. Evaluation is performed on the retained mixed samples.

[0082] Figure 25 For different N values, the AUC values ​​of the CUPID v1.1 ensemble classifiers are constructed using the N most significant DMRs between each pairwise comparison, as detailed on the X-axis. Evaluations were performed on retainer mix samples differentiated by tumor content (TC) below 3% (left panel), between 3% and 5% (middle panel), and above 5% (right panel).

[0083] Figure 26 : Figure 24 The proportion of correct, incorrect, and unpredictable results for each number of DMRs in each pairwise comparison of the computer simulation mixture set.

[0084] Figure 27When tested on 143 cfDNA sample cohorts from TARGET and CUP (n=41), the proportion of samples in each number of pairs of most significant DMRs predicted correctly, incorrectly, or not predicted in CUPiDv.1.1.

[0085] Figure 28 A graph ranking the importance of DMR variables in CUPID v1 versus v1.1 using all DMRs. Detailed Implementation

[0086] According to a first aspect, the present invention provides a computer-implemented method for determining the primary carcinoma of a malignant tumor of unknown origin (MUO) of a subject, the method comprising:

[0087] i) Provided sequence data of cell-free DNA (cfDNA) obtained from a liquid sample from the subject, the sequence data including regional methylation features (“sample data”).

[0088] ii) Feeding sample data as input to a machine learning model that includes a multi-class classifier trained on a training dataset, the training dataset comprising regional methylation features (“training data”) from at least two primary malignant tumor types, each primary malignant tumor type having a different location and / or histological subtype;

[0089] iii) Receive output from a machine learning model, the output indicating the primary malignant tumor type classification or unknown type classification of the sample data; and

[0090] iv) Determine the primary cancer of the object based on the output received in step (iii).

[0091] As used in this article, the term "primary carcinoma" refers to the original or first-degree malignant tumor in the subject.

[0092] As used herein, “object” refers to an animal or a human. An object can be a mammal (e.g., a human, a non-human primate, a cat, a dog, a horse, a donkey, a sheep, a pig, a goat, a cow, a mouse, a rat, a rabbit, or a guinea pig). The object is preferably a human. The object can be an adult human (at least 18 years of age). In some embodiments, the object is a child human (under 18 years of age).

[0093] For “malignant tumor of unknown origin (MUO),” this will be understood to refer to a condition in which the most probable cancer cells and / or symptoms have been identified in the subject, but the primary origin of the cancer cells is unknown. In some implementations, MUO includes cancer of unknown primary origin (CUP). Cancer of unknown primary origin (CUP) will be understood to refer to a cancerous tumor that has been identified in the subject, but the primary cancer remains unknown after standard clinical investigation.

[0094] The machine learning model of the present invention includes a multi-class classifier trained on a training dataset comprising regional methylation features from at least two primary malignant tumor types, each primary malignant tumor type having a different location and / or histological subtype. In some embodiments, each primary malignant tumor type has a different location and histological subtype. The different location and / or histological subtype of each primary malignant tumor type ensures that each primary malignant tumor type on which the classifier is trained is clearly defined and distinctly different from any other primary malignant tumor type on which the classifier is trained. Advantageously, this ensures that the classifier, when receiving sample data as input, is able to provide an output indicating an accurate and sensitive classification of the primary tumor type.

[0095] As used herein, the term "different location" will be understood to mean that each primary malignant tumor type has an anatomical location in the subject that is different from that of each other primary malignant tumor type when in situ. Anatomical location may include a specific organ or group of organs of the subject. For example, location may include the gastrointestinal tract, head, neck, liver, cervix, bile duct, bladder, breast, blood, kidney, lung, pancreas, prostate, skin, muscle, tendon, fat, lymphatic system, blood vessel, nerve, bile duct, penis, bone, testis, thyroid gland, thymus, mesothelial tissue, ovary, uterus, adrenal gland, heart, spleen, or gallbladder.

[0096] The head may include one or more of the following: esophagus, brain, tongue, throat, salivary glands, eyes, and cheeks.

[0097] The gastrointestinal tract may include one or more of the stomach, esophagus, colon, rectum, small intestine, and large intestine.

[0098] In some implementations, different locations include specific locations within an organ or object region—for example, the upper or lower gastrointestinal tract.

[0099] In the context of this invention, the term "histological subtype" will be understood as a classification of the type of primary malignant tumor based on at least the morphological appearance of tumor cells. It should be understood that the morphological appearance of tumor cells often differs between different histological subtypes (e.g., carcinoma and neuroendocrine). Therefore, histological subtypes can include histological classifications of the type of primary malignant tumor. In some embodiments, histological subtypes are classifications of the type of primary malignant tumor based on the morphological appearance of tumor cells and the rate of tumor cell growth and division.

[0100] The term “cell-free DNA” (cfDNA) as used herein may also be referred to as circulating cell-free DNA. Generally, “cell-free DNA” refers to DNA that is not contained within intact cells (although a sample may intentionally contain or contain one or more cells as trace impurities). Therefore, cfDNA typically includes DNA that has been released from cells into a biological fluid of the subject, such as circulating biological fluids like blood, plasma, urine, or cerebrospinal fluid.

[0101] Regional methylation characteristics can include the DNA methylation status of multiple genomic regions. It should be understood that the term "genomic region" is used interchangeably with the term "genomic window." Methylation status can include hypermethylation or hypomethylation. Typically, methylation status includes a measure of the degree of high or low methylation observed in a genomic region. Methylation status can be included in methylation sequencing reads.

[0102] In the context of this invention, a genomic region will be understood as a portion of the genome. The length of a genomic region may be at least about 50 base pairs, at least about 100 base pairs, at least about 150 base pairs, at least about 200 base pairs, or at least about 250 base pairs. In some implementations, the length of the genomic region does not exceed approximately 2000 base pairs, approximately 1900 base pairs, approximately 1800 base pairs, approximately 1700 base pairs, approximately 1600 base pairs, approximately 1500 base pairs, approximately 1400 base pairs, approximately 1300 base pairs, approximately 1200 base pairs, approximately 1100 base pairs, approximately 1000 base pairs, approximately 900 base pairs, approximately 800 base pairs, approximately 700 base pairs, approximately 600 base pairs, approximately 550 base pairs, approximately 500 base pairs, approximately 450 base pairs, approximately 400 base pairs, or approximately 350 base pairs. In some implementations, the length of each genomic region is approximately 200 to approximately 2000 base pairs. In some embodiments, each genomic region is about 200 to about 1000 base pairs in length. In some embodiments, each genomic region is about 200 to about 400 base pairs in length. In some embodiments, each genomic region is about 300 base pairs in length. Genomic regions will be discussed in further detail below.

[0103] In some implementations, the genomic region contains one or more CpG islands or CpG sites.

[0104] Types of primary malignant tumors

[0105] In some implementations, at least two primary malignant tumor types include at least three primary malignant tumor types, optionally at least four primary malignant tumor types, wherein each primary malignant tumor type has a different location and / or histological subtype.

[0106] In some implementations, at least two primary malignant tumor types include at least five, at least six, at least seven, at least eight, at least nine, or at least ten primary malignant tumor types. It should be understood that each primary malignant tumor type has a different location and / or histological subtype relative to the other primary malignant tumor types. In some implementations, at least two primary malignant tumor types include at least eleven, at least twelve, at least thirteen, at least fourteen, or at least fifteen primary malignant tumor types.

[0107] In some implementations, at least two primary malignant tumor types include at least 20 primary malignant tumor types, each with a different location and / or histological subtype.

[0108] In some implementation schemes, at least two primary malignant tumor types include at least 21 primary malignant tumor types, at least 22 primary malignant tumor types, at least 23 primary malignant tumor types, at least 24 primary malignant tumor types, or at least 25 primary malignant tumor types.

[0109] In some implementation schemes, at least two primary malignant tumor types include at least 29 primary malignant tumor types.

[0110] In some implementation schemes, at least two primary malignant tumor types include no more than 40 primary malignant tumor types, no more than 39 primary malignant tumor types, no more than 38 primary malignant tumor types, no more than 37 primary malignant tumor types, no more than 36 primary malignant tumor types, or no more than 35 primary malignant tumor types.

[0111] In some implementation schemes, at least two primary malignant tumor types include no more than 30 primary malignant tumor types.

[0112] In some embodiments, at least two primary malignant tumor types include 10 to 40 primary malignant tumor types. At least two primary malignant tumor types may include 20 to 30 primary malignant tumor types. In some embodiments, at least two primary malignant tumor types include 29 primary malignant tumor types.

[0113] In some implementations, at least two primary malignant tumor types include at least one rare primary cancer tumor type. The term "rare cancer" or "rare primary cancer" is generally defined as a cancer with an incidence rate of <6 / 100,000 in the general population. Rare cancers may include pediatric rare cancers or cancers in adolescents and young adults. Further information on rare cancers can be found in Keat, N., K. Law, M. Seymour, et al., International rare cancers initiative. Lancet Oncol, 2013. 14(2): p. 109-10, which is incorporated herein by reference in its entirety.

[0114] In some implementation schemes, at least two primary malignant tumor types are selected from:

[0115] a) Primary malignant hepatobiliary tumors;

[0116] b) Primary malignant female reproductive tract tumors;

[0117] c) Primary malignant upper or lower gastrointestinal tumors;

[0118] d) Primary malignant lung tumor;

[0119] e) Primary malignant breast tumor; and

[0120] f) Primary malignant urological tumors.

[0121] At least one of the two primary malignant tumor types may include primary malignant hepatopancreatic biliary tumors, optionally primary malignant cholangiocarcinoma tumors.

[0122] In some implementation schemes, at least two primary malignant tumor types are selected from: adrenal cancer, anal cancer, bile duct cancer, bladder cancer, bone cancer, brain / CNS tumors, breast cancer, Castleman disease, cervical cancer, colorectal cancer, endometrial cancer, esophageal cancer, Ewing family tumors, eye cancer, gallbladder cancer, gastrointestinal carcinoid tumors, gastrointestinal stromal tumors (GIST), gestational trophoblastic disease, Hodgkin's disease, Kaposi's sarcoma, renal cancer, laryngeal and hypopharyngeal cancer, and leukemia (acute lymphoblastic, acute myeloid, chronic lymphoblastic, chronic myeloid, chronic myeloid, chronic granulocytic). Myelomonocytic, liver cancer, lung cancer (non-small cell, small cell, lung carcinoid), lymphoma, cutaneous lymphoma, malignant mesothelioma, multiple myeloma, myelodysplastic syndrome, nasal and paranasal sinus carcinoma, nasopharyngeal carcinoma, neuroblastoma, non-Hodgkin lymphoma, oral and oropharyngeal carcinoma, osteosarcoma, ovarian cancer, penile cancer, pituitary adenoma, prostate cancer, retinoblastoma, rhabdomyosarcoma, salivary gland cancer, sarcoma—adult soft tissue cancer, skin cancer (basal and squamous cell, melanoma, Merkel cell), small intestine cancer, gastric cancer, testicular cancer, thymic cancer, thyroid cancer, uterine sarcoma, vaginal cancer, vulvar cancer, Waldenström macroglobulinemia, and Wilms tumor.

[0123] At least two primary malignant tumor types may include breast cancer and at least one other primary malignant tumor type. Breast cancer may include hormone-positive breast cancer or human epidermal growth factor-receptor 2 (HER2)-positive breast cancer. Hormone-positive breast cancer may include estrogen receptor-positive (ER+) breast cancer, progesterone receptor-positive (PR+) breast cancer, or ER+PR+ breast cancer. In some implementations, breast cancer includes triple-negative breast cancer.

[0124] In some implementation schemes, at least two primary malignant tumor types are selected from: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), invasive breast carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynecological carcinoma (Gynae), chromophobe renal cell carcinoma (KICH), clear cell renal cell carcinoma (KIRC), papillary renal cell carcinoma (KIRP), acute myeloid leukemia (LAML), and low-grade glioma (LG). G), hepatocellular carcinoma (LIHC), colorectal and rectal adenocarcinoma (lower GI), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), mesothelioma (MESO), pancreatic adenocarcinoma (PAAD), pheochromocytoma and paraganglioma (PCPG), prostate adenocarcinoma (PRAD), sarcoma (SARC), cutaneous melanoma (SKCM), testicular germ cell tumor (TGCT), thyroid cancer (THCA), thymoma (THYM), gastric and esophageal adenocarcinoma (upper GI), head and neck and esophageal squamous cell carcinoma (upper Sq), and uveal melanoma (UVM).

[0125] For example, at least two primary malignant tumor types may include at least two of the following primary malignant tumor types: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), invasive breast carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynecological carcinoma (Gynae), renal chromophobe cell carcinoma (KICH), renal clear cell carcinoma (KIRC), renal papillary cell carcinoma (KIRP), acute myeloid leukemia (LAML), and low-grade cerebral glioma. Gestational tumor (LGG), hepatocellular carcinoma (LIHC), colorectal and rectal adenocarcinoma (lower GI), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), mesothelioma (MESO), pancreatic adenocarcinoma (PAAD), pheochromocytoma and paraganglioma (PCPG), prostate adenocarcinoma (PRAD), sarcoma (SARC), cutaneous melanoma (SKCM), testicular germ cell tumor (TGCT), thyroid cancer (THCA), thymoma (THYM), gastric and esophageal adenocarcinoma (upper GI), head and neck and esophageal squamous cell carcinoma (upper Sq), and uveal melanoma (UVM).

[0126] In some implementations, at least two primary malignant tumor types include at least three of the following primary malignant tumor types: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), invasive breast carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynecological carcinoma (Gynae), chromophobe renal cell carcinoma (KICH), clear cell renal cell carcinoma (KIRC), papillary renal cell carcinoma (KIRP), acute myeloid leukemia (LAML), and cerebral hypoplasia. Grade 1 glioma (LGG), hepatocellular carcinoma (LIHC), colorectal adenocarcinoma (lower GI), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), mesothelioma (MESO), pancreatic adenocarcinoma (PAAD), pheochromocytoma and paraganglioma (PCPG), prostate adenocarcinoma (PRAD), sarcoma (SARC), cutaneous melanoma (SKCM), testicular germ cell tumor (TGCT), thyroid cancer (THCA), thymoma (THYM), gastric and esophageal adenocarcinoma (upper GI), head and neck and esophageal squamous cell carcinoma (upper Sq), and uveal melanoma (UVM).

[0127] In some implementations, at least two primary malignant tumor types include at least four of the following primary malignant tumor types: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), invasive breast carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynecological carcinoma (Gynae), chromophobe renal cell carcinoma (KICH), clear cell renal cell carcinoma (KIRC), papillary renal cell carcinoma (KIRP), acute myeloid leukemia (LAML), and cerebral hypoplasia. Grade 1 glioma (LGG), hepatocellular carcinoma (LIHC), colorectal adenocarcinoma (lower GI), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), mesothelioma (MESO), pancreatic adenocarcinoma (PAAD), pheochromocytoma and paraganglioma (PCPG), prostate adenocarcinoma (PRAD), sarcoma (SARC), cutaneous melanoma (SKCM), testicular germ cell tumor (TGCT), thyroid cancer (THCA), thymoma (THYM), gastric and esophageal adenocarcinoma (upper GI), head and neck and esophageal squamous cell carcinoma (upper Sq), and uveal melanoma (UVM).

[0128] In some implementation schemes, at least two primary malignant tumor types include at least five of the following primary malignant tumor types: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), invasive breast carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynecological carcinoma (Gynae), chromophobe renal cell carcinoma (KICH), clear cell renal cell carcinoma (KIRC), papillary renal cell carcinoma (KIRP), acute myeloid leukemia (LAML), and cerebral hypoplasia. Grade 1 glioma (LGG), hepatocellular carcinoma (LIHC), colorectal adenocarcinoma (lower GI), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), mesothelioma (MESO), pancreatic adenocarcinoma (PAAD), pheochromocytoma and paraganglioma (PCPG), prostate adenocarcinoma (PRAD), sarcoma (SARC), cutaneous melanoma (SKCM), testicular germ cell tumor (TGCT), thyroid cancer (THCA), thymoma (THYM), gastric and esophageal adenocarcinoma (upper GI), head and neck and esophageal squamous cell carcinoma (upper Sq), and uveal melanoma (UVM).

[0129] In the context of this invention, the abbreviation "Upper Sq" will be understood as referring to a grouping of squamous cell carcinomas of the head, neck, and esophagus. Due to their similar anatomical and histological subtypes, the inventors grouped these different tumor types together in the "Upper Sq" group. Therefore, it should be understood that, in the context of this invention, the term "Upper Sq" may cover one or a combination of head and neck and esophageal squamous cell carcinomas.

[0130] In some embodiments, at least one of at least two primary malignant tumor types includes an upper Sq primary malignant tumor type. The upper Sq primary malignant tumor type may include primary malignant head squamous cell carcinoma, primary malignant neck squamous cell carcinoma, or primary malignant esophageal squamous cell carcinoma. In some embodiments, the upper Sq primary malignant tumor type may include primary malignant head squamous cell carcinoma and primary malignant neck squamous cell carcinoma.

[0131] As used herein, the term "non-squamous gynecological cancer" may refer to one or more of endocervical adenocarcinoma, ovarian cystadenocarcinoma, endometrial cancer, and uterine carcinosarcoma. In some embodiments, at least one of at least two primary malignant tumor types includes non-squamous gynecological cancer. Non-squamous gynecological cancers may include endocervical adenocarcinoma, ovarian cystadenocarcinoma, endometrial cancer, and uterine carcinosarcoma. In some embodiments, non-squamous gynecological cancers include endocervical adenocarcinoma. In some embodiments, non-squamous gynecological cancers include ovarian cystadenocarcinoma. In some embodiments, non-squamous gynecological cancers include endometrial cancer. In some embodiments, non-squamous gynecological cancers include uterine carcinosarcoma.

[0132] At least two primary malignant tumor types may be included, such as adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), invasive breast carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynecological carcinoma (Gynae), renal chromophobe cell carcinoma (KICH), renal clear cell carcinoma (KIRC), renal papillary cell carcinoma (KIRP), acute myeloid leukemia (LAML), low-grade glioma of the brain (LGG), and liver cell carcinoma. Lung adenocarcinoma (LIHC), colon and rectal adenocarcinoma (lower GI), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), mesothelioma (MESO), pancreatic adenocarcinoma (PAAD), pheochromocytoma and paraganglioma (PCPG), prostate adenocarcinoma (PRAD), sarcoma (SARC), cutaneous melanoma (SKCM), testicular germ cell tumor (TGCT), thyroid cancer (THCA), thymoma (THYM), gastric and esophageal adenocarcinoma (upper GI), head and neck and esophageal squamous cell carcinoma (upper Sq), and uveal melanoma (UVM).

[0133] Therefore, in some implementation schemes, at least two primary malignant tumor types include the following 29 primary malignant tumor types: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), invasive breast carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynecological carcinoma (Gynae), chromophobe renal cell carcinoma (KICH), clear cell renal cell carcinoma (KIRC), papillary renal cell carcinoma (KIRP), acute myeloid leukemia (LAML), and cerebral hypoplasia. Grade 1 glioma (LGG), hepatocellular carcinoma (LIHC), colorectal adenocarcinoma (lower GI), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), mesothelioma (MESO), pancreatic adenocarcinoma (PAAD), pheochromocytoma and paraganglioma (PCPG), prostate adenocarcinoma (PRAD), sarcoma (SARC), cutaneous melanoma (SKCM), testicular germ cell tumor (TGCT), thyroid cancer (THCA), thymoma (THYM), gastric and esophageal adenocarcinoma (upper GI), head and neck and esophageal squamous cell carcinoma (upper Sq), and uveal melanoma (UVM).

[0134] In some implementations, the training data also includes regional methylation features from at least one non-cancer control. In the context of this invention, the term "non-cancer control" refers to an individual known not to have cancer. A "non-cancer control" individual can therefore be referred to as healthy.

[0135] Training data may include regional methylation features from at least two, at least five, at least 10, at least 20, at least 30, at least 40, at least 50, at least 60, at least 70, or at least 80 non-cancer controls. In some embodiments, the training data includes regional methylation features from at least 70 non-cancer controls, optionally from at least 75. In some embodiments, the training data includes regional methylation features from about 79 non-cancer controls. Advantageously, providing regional methylation features from at least one non-cancer control avoids false positives. In particular, including regional methylation features from at least one non-cancer control enables the output of an "unknown type classification" for sample data in which the regional methylation features are not sufficiently aligned with any regional methylation features of the primary malignant tumor type in the training data. Therefore, an unknown type classification output may occur when the subject does not have cancer or has cancer that is significantly different from the cancer of the primary malignant tumor type in the training data. Alternatively, an unknown type classification may occur when the amount of tumor DNA present in a liquid biopsy is too low to be detected and / or analyzed.

[0136] In some embodiments, the regional methylation signature from at least one non-cancer control includes a regional methylation signature obtained from cfDNA from at least one non-cancer control. In some embodiments, the regional methylation signature from at least one non-cancer control includes a regional methylation signature obtained from tissue DNA from at least one non-cancer control or from an organ or tissue known not to contain cancer, which may be further referred to as non-cancer tissue, such as non-cancer liver tissue. In some embodiments, the regional methylation signature from at least one non-cancer control includes a regional methylation signature obtained from cfDNA from at least one non-cancer control and a regional methylation signature obtained from liver tissue DNA from at least one non-cancer control. The inventors have advantageously discovered that including a regional methylation signature obtained from liver tissue DNA from at least one non-cancer control improves the accuracy of the method, particularly by reducing the chance that the model output incorrectly indicates that a subject without cancer has a primary liver cancer tumor.

[0137] The subject may have an unfavorable CUP. An "unfavorable CUP" will be understood as a CUP that is not a favorable CUP and is therefore considered a cancer that may have a low response to treatment. In some embodiments, the subject is known to have secondary cancer, but the primary cancer is unknown. In some embodiments, the subject is known to have secondary adenocarcinoma, secondary squamous cell carcinoma, or secondary poorly differentiated carcinoma, but the primary cancer is unknown.

[0138] In some implementation schemes, the primary cancer in step iv) is selected from: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), invasive breast carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynecological carcinoma (Gynae), chromophobe renal cell carcinoma (KICH), clear cell renal cell carcinoma (KIRC), papillary renal cell carcinoma (KIRP), acute myeloid leukemia (LAML), low-grade glioma of the brain (LGG), liver cell carcinoma, etc. Primary cancers include: Lipocarcinoma of the lung (LIHC), colon and rectum (lower GI), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), mesothelioma (MESO), pancreatic adenocarcinoma (PAAD), pheochromocytoma and paraganglioma (PCPG), prostate adenocarcinoma (PRAD), sarcoma (SARC), cutaneous melanoma (SKCM), testicular germ cell tumor (TGCT), thyroid cancer (THCA), thymoma (THYM), gastric and esophageal adenocarcinoma (upper GI), head, neck and esophageal squamous cell carcinoma (upper Sq), uveal melanoma (UVM), and unclassified primary cancers.

[0139] In some implementations, the primary cancer in step iv) is selected from breast cancer, female reproductive cancer, blood cancer, kidney cancer, lower abdominal cancer, lung cancer, upper abdominal cancer, brain cancer, male reproductive cancer, NET (neuroendocrine tumor), skin cancer, thyroid cancer, rare cancers, and unclassified primary cancers.

[0140] Training data

[0141] The training data included regional methylation features from primary malignant tumor types and regional methylation features from at least one other primary malignant tumor type. Each primary malignant tumor type had different locational and / or histological subtypes.

[0142] As described herein, each regional methylation signature includes the DNA methylation status of at least one genomic region. Typically, regional methylation signatures for different primary malignant tumor types include the DNA methylation status of the same genomic region. However, it should be understood that the DNA methylation status of regions may differ between primary malignant tumor types. Therefore, the regional methylation signature for each primary malignant tumor type will be different.

[0143] Each region methylation feature in the training data may include the DNA methylation states of at least 5, 10, 20, 30, 50, or 100 genomic regions. In some embodiments, each region methylation feature in the training data includes the DNA methylation states of at least 50, 100, 150, 200, or 250 genomic regions. In some embodiments, each region methylation feature in the training data includes the DNA methylation states of at least 1000, 5000, 10000, 15000, 20000, or 25000 genomic regions. In some embodiments, each region methylation feature in the training data includes the DNA methylation states of at least 5000 genomic regions. In some embodiments, each region methylation feature in the training data includes at least 1 × 10⁻⁶ genomic regions. 6 At least 2×10 6 At least 3×10 6 At least 4×10 6 At least 5×10 6 At least 6×10 6 At least 7×10 6 At least 8×10 6 Or at least 9×10 6 The DNA methylation status of each genomic region.

[0144] The inventors have observed that in embodiments comprising multiple genomic regions, a significant proportion of these genomic regions can be close to each other. This may indicate that certain regions of the genome are particularly indicative of cancer status. Therefore, when a regional methylation signature comprises the DNA methylation status of multiple genomic regions, at least 10%, at least 20%, at least 30%, at least 40%, or at least 50% of the genomic regions in each regional methylation signature may be no more than 1000 base pairs away from at least one other genomic region in the regional methylation signature. In some embodiments, at least 10%, at least 20%, at least 30%, at least 40%, or at least 50% of the genomic regions in each regional methylation signature are no more than 500 base pairs away from at least one other genomic region in the regional methylation signature. In some embodiments, at least 60%, optionally at least 70%, of the genomic regions in each regional methylation signature are no more than 1000 base pairs away from at least one other genomic region in the regional methylation signature. In some embodiments, at least 60%, optionally at least 70%, of the genomic regions in each methylation feature are separated from at least one other genomic region in the methylation feature by no more than 500 base pairs. In some embodiments, at least 80% of the genomic regions in each methylation feature are separated from at least one other genomic region in the methylation feature by no more than 1000 base pairs, optionally no more than 500 base pairs.

[0145] In some embodiments, at least 10%, at least 20%, at least 30%, at least 40%, or at least 50% of the genomic region in each regional methylation feature is adjacent to or overlaps with at least one other genomic region in the regional methylation feature. For example, at least 60%, optionally at least 70%, of the genomic region in each regional methylation feature may be adjacent to or overlap with at least one other genomic region in the regional methylation feature. In some embodiments, at least 80% of the genomic region in each regional methylation feature is adjacent to or overlaps with at least one other genomic region in the regional methylation feature.

[0146] In some implementations, each region's methylation signature includes the DNA methylation status of all genomic regions in the genome. This may also be referred to as the whole-genome DNA methylation status.

[0147] In some embodiments, the genomic regions in each regional methylation feature of the training data include genomic regions identified from pairwise comparisons of genomic region DNA methylation states between at least two primary malignant tumor types. In some embodiments, the genomic regions in each regional methylation feature of the training data include genomic regions identified from pairwise comparisons of genomic region DNA methylation states between at least two primary malignant tumor types and at least one non-cancer control. In some embodiments, the genomic regions in each regional methylation feature of the training data include genomic regions identified from pairwise comparisons of genomic region DNA methylation states between at least two primary malignant tumor types and non-cancer tissue. In the context of this invention, pairwise comparisons of DNA methylation states will be understood as comparisons of genomic region DNA methylation states between two different types, such as between two different primary malignant tumor types, between a primary malignant tumor type and a non-cancer control, or between a primary malignant tumor type and non-cancer tissue. Genomic regions identified from pairwise comparisons may include genomic regions identified as having the greatest difference in DNA methylation states between two types, such as between two primary malignant tumor types.

[0148] The genomic regions in the training data may include those that exhibit the greatest difference in methylation state when compared using pairwise comparisons (e.g., between two different primary malignant tumor types or between a primary malignant tumor type and a non-cancer control). Therefore, the genomic regions in the training data may include at least one, at least five, at least ten, at least twenty, at least thirty, at least fifty, at least one hundred, at least one hundred, or at least two hundred genomic regions identified from pairwise comparisons that exhibit the greatest difference in methylation state. In some embodiments, the training data includes approximately 250 genomic regions identified from pairwise comparisons that exhibit the greatest difference in methylation state. The difference in methylation state can be measured by statistical analysis. In such embodiments, the largest difference may be the statistically most significant difference in methylation state. Statistical analysis may include setting a threshold for an adjusted p-value and ranking the resulting regions based on differences in β-values.

[0149] The genomic regions in the training data may include those that show the greatest difference in hypomethylation when compared using pairwise comparisons (e.g., between two different primary malignant tumor types or between a primary malignant tumor type and a non-cancer control). Alternatively, the genomic regions in the training data may include those that show the greatest difference in hypermethylation when compared using pairwise comparisons (e.g., between two different primary malignant tumor types or between a primary malignant tumor type and a non-cancer control).

[0150] In some implementations, the genomic regions in the training data include those that exhibit the greatest difference in hypomethylation and the greatest difference in hypermethylation when compared using pairwise comparisons (e.g., between two different primary malignant tumor types or between a primary malignant tumor type and a non-cancer control). Such pairwise comparisons may also be referred to as “up and down” pairwise comparisons.

[0151] For example, in a pairwise comparison between primary malignant tumor type A and primary malignant tumor type B, primary malignant tumor type A may be hypermethylated in two genomic regions and hypomethylated in one genomic region across at least three genomic regions being compared. Therefore, primary malignant tumor type B will be hypomethylated in two genomic regions and hypermethylated in one genomic region.

[0152] The genomic regions in each region methylation feature of the training data may include one, at least 5, at least 10, at least 20, at least 30, at least 50, at least 100, at least 150, at least 200, or at least 250 genomic regions with the greatest difference in hypermethylation. The genomic regions in each region methylation feature of the training data may include one, at least 5, at least 10, at least 20, at least 30, at least 50, at least 100, at least 150, at least 200, or at least 250 genomic regions with the greatest difference in hypomethylation.

[0153] In some implementations, the training data includes approximately 250 genomic regions identified from pairwise comparisons that exhibit the greatest differences in hypermethylation.

[0154] In some implementations, the training data includes approximately 250 genomic regions identified from pairwise comparisons that exhibit the greatest differences in hypomethylation.

[0155] In some implementations, the training data includes approximately 250 genomic regions identified from pairwise comparisons that have the greatest differences in hypermethylation and approximately 250 genomic regions identified from pairwise comparisons that have the greatest differences in hypomethylation.

[0156] The genomic regions in each region methylation feature of the training data may include at least 500, at least 1000, at least 2000, at least 3000, at least 4000, at least 5000, at least 6000, at least 7000, at least 8000, at least 9000, at least 10000, or at least 15000 genomic regions identified as having the greatest differential methylation in one or more pairwise comparisons.

[0157] One or more pairwise comparisons may include two or more pairwise comparisons, preferably three or more pairwise comparisons if at least two primary malignant tumor types include at least three primary malignant tumor types or at least two primary malignant tumor types and a non-cancer control.

[0158] In some implementations, the number of pairwise comparisons is calculated using the formula n(n-1) / 2, where n includes the number of primary malignant tumor types. In implementations where the training data includes regional methylation features from at least one non-cancer control, n may include the number of primary malignant tumor types + 1, where + 1 represents a non-cancer control.

[0159] In embodiments that include at least four types of primary malignant tumors, one or more pairwise comparisons may include at least six pairwise comparisons. In embodiments that include at least five types of primary malignant tumors, one or more pairwise comparisons may include at least ten pairwise comparisons.

[0160] For example, when at least two primary malignant tumor types include at least three primary malignant tumor types (which may also be referred to as first, second, and third primary malignant tumor types), pairwise comparisons may include:

[0161] a) Pairwise comparison of genomic region methylation status between the first and second primary malignant tumor types;

[0162] b) Pairwise comparison of genomic region methylation status between the first and third primary malignant tumor types; and

[0163] c) Pairwise comparison of genomic region methylation status between the second and third primary malignant tumor types.

[0164] When at least two primary malignant tumor types include at least four primary malignant tumor types, pairwise comparisons may include:

[0165] a) Pairwise comparison of genomic region methylation status between the first and fourth primary malignant tumor types;

[0166] b) Pairwise comparison of genomic region methylation status between the first and third primary malignant tumor types;

[0167] c) Pairwise comparison of genomic region methylation status between the first and second primary malignant tumor types;

[0168] d) Pairwise comparison of methylation status of genomic regions between the second and fourth primary malignant tumor types;

[0169] e) Pairwise comparison of genomic region methylation status between the third and fourth primary malignant tumor types; and

[0170] f) Pairwise comparison of genomic region methylation status between the second and third primary malignant tumor types.

[0171] The genomic regions in each region methylation feature of the training data may include at least the top 5, top 10, top 15, top 20, top 25, top 30, top 50, top 100, top 150, top 200, or at least top 250 genomic regions identified from at least one pairwise comparison. The term "top 5" should be understood as referring to the five genomic regions identified as having the maximum differential methylation in at least one pairwise comparison. The maximum differential methylation may include the maximum hypermethylation or the maximum hypomethylation. In some embodiments, the maximum differential methylation includes both maximum hypermethylation and maximum hypomethylation.

[0172] In some embodiments, the multiple genomic regions include no more than about 500,000 genomic regions, no more than about 450,000 genomic regions, no more than about 400,000 genomic regions, no more than about 350,000 genomic regions, no more than about 300,000 genomic regions, or no more than about 250,000 genomic regions. In some embodiments, the multiple genomic regions include no more than about 50,000 genomic regions, no more than about 40,000 genomic regions, no more than about 30,000 genomic regions, or no more than about 25,000 genomic regions.

[0173] In embodiments where genomic regions are identified from multiple pairwise comparisons, the genomic regions may include at least the top 5, top 10, top 15, top 20, top 25, top 30, top 50, top 100, top 150, top 200, or at least the top 250 genomic regions identified from each pairwise comparison as having the greatest differential hypermethylation. In some embodiments, the genomic regions include at least the top 5, top 10, top 15, top 20, top 25, top 30, top 50, top 100, top 150, top 200, or at least the top 250 genomic regions identified from each pairwise comparison as having the greatest differential hypomethylation. In some implementations, the genomic regions include at least the top 5, top 10, top 15, top 20, top 25, top 30, top 50, top 100, top 150, top 200, or at least the top 250 genomic regions identified as having the greatest differential hypomethylation, and at least the top 5, top 10, top 15, top 20, top 25, top 30, top 50, top 100, top 150, top 200, or at least the top 250 genomic regions identified as having the greatest differential hypermethylation.

[0174] For example, the genomic regions may include approximately 250 genomic regions identified in each pairwise comparison. The genomic regions may include approximately 250 genomic regions identified as having the highest degree of hypermethylation and approximately 250 genomic regions identified as having the highest degree of hypomethylation. Therefore, in an embodiment involving, for example, six pairwise comparisons, the genomic regions may include approximately 500 genomic regions identified from each pairwise comparison (250 with the highest degree of hypermethylation and 250 with the highest degree of hypomethylation, thus approximately 3000 genomic regions in total).

[0175] In some implementations, the non-cancer control in the paired comparison is a non-cancer liver control, which may also be referred to as a normal liver category. The non-cancer control in the paired comparison may not include any non-liver controls.

[0176] Before pairwise comparisons, genomic regions from non-cancer controls with a median methylation β value of at least 0.05 can be excluded from the training data. For example, genomic regions from non-cancer controls with a median methylation β value of at least about 0.25 can be excluded from the training data before pairwise comparisons. In some embodiments, genomic regions from non-cancer controls with a median methylation β value of at least about 0.5 are excluded from the training data before pairwise comparisons. Alternatively, genomic regions from non-cancer controls with the aforementioned methylation β values ​​can be excluded from the training data after pairwise comparisons.

[0177] The inventors have identified several genomic regions that exhibit the greatest differences in methylation status in pairwise comparisons of different primary malignant tumor types. Advantageously, the inventors have used the DNA methylation status of these specific genomic regions as training data for training this machine learning model. This has resulted in a machine model capable of accurately distinguishing and assigning primary malignant tumor types. In particular, the inventors have identified the top 250 genomic regions that exhibit the greatest differences in methylation status when analyzed in pairwise comparisons between primary malignant tumor types. Increased methylation in a genomic region of one primary malignant tumor type compared to another can indicate an association with that primary malignant tumor type. Alternatively, increased hypomethylation in a genomic region of one primary malignant tumor type compared to another can indicate an association with that primary malignant tumor type. Therefore, regions identified as having the greatest differences in methylation status in pairwise comparisons of different primary malignant tumor types are regions for which the inventors have found particularly useful training data. The inventors have advantageously used such training data to generate accurate and effective machine learning models that can subsequently indicate the primary malignant tumor type of an object.

[0178] Table 1 includes the top 250 genomic regions identified by the inventors from each of 30 pairwise comparisons. These genomic regions were used as training data for “CUPiD v1” in Examples 1 through 6. CUPID v1 is a machine learning model trained on a training dataset that includes methylation features of regions from 29 primary malignant tumor types, general non-cancer controls, and normal liver categories. The first three columns of Table 1 indicate the genomic coordinates of each genomic region. The fourth column lists the primary malignant tumor type categories in which a particular genomic region is highly methylated (i.e., higher than another category) in at least one pairwise comparison. The fifth column shows the pairwise comparisons in which the genomic region is located among the top 250 most highly methylated genomic regions. For example, TGCT>Gynae means that a particular genomic region is located among the top 250 genomic regions with the greatest methylation difference between TGCT and Gynae (higher methylation in TGCT).

[0179] Table 3 of this application provides a complete definition for each abbreviation used in columns 4 and 5 of Table 1. For example, TGCT is used to refer to testicular germ cell tumors, while Gynae is used to refer to non-squamous gynecological cancers (cervical endometrial adenocarcinoma, ovarian cystadenocarcinoma, endometrial cancer, uterine carcinosarcoma).

[0180] Therefore, in some implementations, the genomic regions are selected from the genomic regions listed in Table 1.

[0181] In an embodiment where the primary malignant tumor type is adrenocortical carcinoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "ACC >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is urothelial carcinoma of the bladder, the genomic region may be selected from any genomic region listed in Table 1 that presents "BLCA >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is invasive breast cancer, the genomic region may be selected from any genomic region listed in Table 1 that presents "BRCA >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is squamous cell carcinoma of the cervix, the genomic region may be selected from any genomic region listed in Table 1 that presents "CervSq >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is cholangiocarcinoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "CHOL >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is diffuse large B-cell lymphoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "DLBC >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is glioblastoma multiforme, the genomic region may be selected from any genomic region listed in Table 1, where “GBM >” is presented in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is non-squamous gynecological carcinoma, the genomic region may be selected from any genomic region listed in Table 1, where “Gynae >” is presented in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is chromophobe renal carcinoma, the genomic region may be selected from any genomic region listed in Table 1, where “KICH >” is presented in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is clear cell renal carcinoma, the genomic region may be selected from any genomic region listed in Table 1, where “KIRC >” is presented in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is papillary renal carcinoma, the genomic region may be selected from any genomic region listed in Table 1, where “KIRP >” is presented in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is acute myeloid leukemia, the genomic region may be selected from any genomic region listed in Table 1, where “LAML >” is presented in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is low-grade glioma of the brain, the genomic region may be selected from any genomic region listed in Table 1 that presents "LGG >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is hepatocellular carcinoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "LIHC >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is colorectal adenocarcinoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "lower GI >" in the fourth column of Table 1.In an embodiment where the primary malignant tumor type is lung adenocarcinoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "LUAD >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is lung squamous cell carcinoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "LUSC >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is mesothelioma, the genomic region may be selected from any genomic region listed in Table 1 that presents "MESO >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is pancreatic adenocarcinoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "PAAD >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is pheochromocytoma and paraganglioma, the genomic region may be selected from any genomic region listed in Table 1 that presents "PCPG >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is prostate adenocarcinoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "PRAD >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is sarcoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "SARC >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is cutaneous melanoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "SKCM >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is testicular germ cell tumor, the genomic region may be selected from any genomic region listed in Table 1 that presents "TGCT >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is thyroid cancer, the genomic region may be selected from any genomic region listed in Table 1 that presents "THCA >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is thymoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "THYM >" in the fourth column of Table 1. In an embodiment where the primary malignant tumor type is gastric and esophageal adenocarcinoma, the genomic region may be selected from any genomic region listed in Table 1 that presents "Upper GI >" in the fourth column of Table 1. In an implementation where the primary malignant tumor type is squamous cell carcinoma of the head and neck or esophagus, the genomic region may be selected from any genomic region listed in Table 1 that presents “Upper Sq >” in the fourth column of Table 1. In an implementation where the primary malignant tumor type is uveal melanoma, the genomic region may be selected from any genomic region listed in Table 1 that presents “UVM >” in the fourth column of Table 1.

[0182] In some embodiments, the genomic regions are selected from those listed in Table 2. The genomic regions listed in Table 2 have been identified by the inventors as the most informative genomic regions for distinguishing primary malignant tumor types, as determined by variable importance scores through pairwise comparisons between different primary malignant tumor categories. The first column identifies the chromosome on which the genomic region is found. The start and end points of the genomic regions are subsequently listed in columns 2 and 3 of Table 2, respectively. These genomic regions have been ranked in Table 2 by variable importance scores – those at the beginning of the table are considered more important to this model than those at the bottom (see column 4 for details of average importance). In some embodiments, the genomic regions include at least five, at least 10, at least 20, at least 30, at least 40, or at least 50 genomic regions listed in Table 2. In some embodiments, the genomic regions include at least the top five, top 10, top 20, top 30, top 40, top 50, top 60, top 70, top 80, or top 90 genomic regions in Table 2. The term "top 10" should be understood to refer to, for example, the top 10 genomic regions listed in Table 2. In some embodiments, the genomic regions include the top 100 genomic regions listed in Table 2. In some embodiments, the genomic regions include the top 150 genomic regions listed in Table 2. In some embodiments, the genomic regions include the top 200 genomic regions listed in Table 2. In some embodiments, the genomic regions include all 250 genomic regions listed in Table 2.

[0183] In some embodiments, the genomic regions are selected from those listed in Table 9. The genomic regions listed in Table 9 have been identified by the inventors as the top 500 most informative genomic regions distinguishing primary malignant tumor types, as determined by variable importance scores in pairwise comparisons between different primary malignant tumor categories in CUPID v1. The first column identifies the chromosome on which the genomic region is found. The start and end points of the genomic regions are subsequently listed in columns 2 and 3 of Table 9, respectively. These genomic regions have been ranked in Table 9 by variable importance scores – those at the beginning of the table are considered more important to this model than those at the bottom (see column 4 for details of average importance). Therefore, Table 9 includes all 250 genomic regions listed in Table 2 plus an additional 250 genomic regions. It should be understood that these additional 250 genomic regions are also listed in Table 1. In some embodiments, the genomic regions include at least five, at least 10, at least 20, at least 30, at least 40, or at least 50 genomic regions listed in Table 9. In some implementations, the genomic regions include at least the first 300, 350, 400, or 450 genomic regions listed in Table 9. In some implementations, the genomic regions include all 500 genomic regions listed in Table 9.

[0184] In some implementations, the genomic regions are selected from those listed in Table 5. Table 5 includes the first 250 genomic regions identified by the inventors from pairwise comparisons of a later version of the model (“CUPiD v1.1”). These pairwise comparisons were between the same 29 primary malignant tumor categories and normal liver categories as in CUPID v1.

[0185] Unlike Table 1, the pairwise comparisons in Table 5 do not include any comparisons with non-cancer control samples other than the normal liver category. The genomic regions in Table 5 were used for training data in Examples 8 through 10. 11,809 identical genomic regions were found in both Tables 1 and 5. Therefore, in some embodiments, the genomic regions are selected from those present in both Tables 1 and 5.

[0186] In an embodiment where the primary malignant tumor type is adrenocortical carcinoma, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "ACC >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is urothelial carcinoma of the bladder, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "BLCA >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is invasive breast cancer, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "BRCA >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is squamous cell carcinoma of the cervix, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "CervSq >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is cholangiocarcinoma, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "CHOL >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is diffuse large B-cell lymphoma, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "DLBC >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is glioblastoma multiforme, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "GBM >" in column four. In an embodiment where the primary malignant tumor type is non-squamous gynecological carcinoma, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "Gynae >" in column four. In an embodiment where the primary malignant tumor type is chromophobe renal carcinoma, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "KICH >" in column four. In an embodiment where the primary malignant tumor type is clear cell renal carcinoma, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "KIRC >" in column four. In an embodiment where the primary malignant tumor type is papillary renal carcinoma, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "KIRP >" in column four. In an embodiment where the primary malignant tumor type is acute myeloid leukemia, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "LAML >" in column four. In an embodiment where the primary malignant tumor type is low-grade glioma of the brain, the genomic region may be selected from any genomic region listed in Table 5 that presents "LGG >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is hepatocellular carcinoma, the genomic region may be selected from any genomic region listed in Table 5 that presents "LIHC >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is colorectal adenocarcinoma, the genomic region may be selected from any genomic region listed in Table 5 that presents "lower GI >" in the fourth column of Table 5.In an embodiment where the primary malignant tumor type is lung adenocarcinoma, the genomic region may be selected from any genomic region listed in Table 5 that presents "LUAD >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is lung squamous cell carcinoma, the genomic region may be selected from any genomic region listed in Table 5 that presents "LUSC >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is mesothelioma, the genomic region may be selected from any genomic region listed in Table 5 that presents "MESO >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is pancreatic adenocarcinoma, the genomic region may be selected from any genomic region listed in Table 5 that presents "PAAD >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is pheochromocytoma and paraganglioma, the genomic region may be selected from any genomic region listed in Table 5 that presents "PCPG >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is prostate adenocarcinoma, the genomic region may be selected from any genomic region listed in Table 5 that presents "PRAD >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is sarcoma, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "SARC >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is cutaneous melanoma, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "SKCM >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is testicular germ cell tumor, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "TGCT >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is thyroid cancer, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "THCA >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is thymoma, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "THYM >" in the fourth column of Table 5. In an embodiment where the primary malignant tumor type is gastric and esophageal adenocarcinoma, the genomic region may be selected from any genomic region listed in Table 5, specifically those displaying "Upper GI >" in the fourth column of Table 5. In an implementation scheme where the primary malignant tumor type is squamous cell carcinoma of the head and neck or esophagus, the genomic region may be selected from any genomic region listed in Table 5 that presents "Upper Sq >" in the fourth column of Table 5. In an implementation scheme where the primary malignant tumor type is uveal melanoma, the genomic region may be selected from any genomic region listed in Table 5 that presents "UVM >" in the fourth column of Table 5.

[0187] In some embodiments, the genomic regions are selected from those listed in Table 6. The genomic regions listed in Table 6 have been identified by the inventors as the top 250 genomic regions for distinguishing primary malignant tumor types, as determined by variable importance scoring in CUPIDv1.1. The first column identifies the chromosome on which the genomic region is found. The start and end points of the genomic regions are subsequently listed in columns 2 and 3 of Table 6, respectively. These genomic regions have been ranked in Table 6 by variable importance scoring – those at the beginning of the table are considered more important than those at the bottom (see column 4 for details of average importance). In some embodiments, the genomic region includes at least five, at least 10, at least 20, at least 30, at least 40, or at least 50 genomic regions listed in Table 6. In some embodiments, the genomic region includes at least the top five, top 10, top 20, top 30, top 40, top 50, top 60, top 70, top 80, or top 90 genomic regions in Table 6. The term "top 10" should be understood to refer to, for example, the top 10 genomic regions listed in Table 6. In some embodiments, this genomic region includes the top 100 genomic regions listed in Table 6. In some embodiments, this genomic region includes the top 150 genomic regions listed in Table 6. In some embodiments, this genomic region includes the top 200 genomic regions listed in Table 6. In some embodiments, this genomic region includes all 250 genomic regions listed in Table 6.

[0188] Many genomic regions in Table 2 are also shared with Table 6. A total of 174 genomic regions listed in Table 2 are also present in Table 6. These are listed in Table 7. It should be understood that these 174 genomic regions are also present in Tables 1 and 5. Therefore, in some embodiments, genomic regions include multiple regions shared by both Table 2 and Table 6. In other words, genomic regions may include multiple regions selected from Table 7. For example, genomic regions may include at least five, at least 10, at least 20, at least 30, at least 40, or at least 50 of the genomic regions shared by Tables 2 and Table 6 (as shown in Table 7). Four of the first five genomic regions in Table 2 are also among the first seven genomic regions in Table 6. Furthermore, six of the first ten genomic regions in Table 2 are also among the first ten in Table 6. Therefore, in some embodiments, genomic regions include at least the first two, three, four, and five genomic regions listed in Table 2, which are also present in Table 6. In some implementations, the genomic regions include at least five of the first 10 genomic regions in Table 2, optionally at least six, at least seven, at least eight, or at least nine. Eight of the first 10 genomic regions in Table 6 (genomic regions ranked 2 through 9) are also present in Table 2. Therefore, in some implementations, the genomic regions include the genomic regions ranked 2 through 9 in Table 6.

[0189] In some implementations, the genomic regions include all the genomic regions listed in Table 7.

[0190] In some embodiments, the genomic regions are selected from those listed in Table 8. The genomic regions listed in Table 8 have been identified by the inventors as the top 500 comprehensive genomic regions distinguishing primary malignant tumor types, as determined by variable importance scoring in CUPIDv1.1. The first column identifies the chromosome on which the genomic region is found. The start and end points of the genomic regions are subsequently listed in columns 2 and 3 of Table 8, respectively. These genomic regions have been ranked in Table 8 by variable importance scoring – those at the beginning of the table are considered more important than those at the bottom (see column 4 for details of average importance). In some embodiments, the genomic regions include at least five, at least 10, at least 20, at least 30, at least 40, or at least 50 genomic regions listed in Table 8. In some embodiments, the genomic regions include the top 300 genomic regions listed in Table 8. In some embodiments, the genomic regions include the top 350 genomic regions listed in Table 8. In some embodiments, the genomic regions include the top 400 genomic regions listed in Table 8. In some embodiments, the genomic regions include the top 450 genomic regions listed in Table 8. In some implementations, the genomic regions include all the genomic regions listed in Table 8.

[0191] The 224 genomic regions listed in Table 8 are also listed in Table 2 (the first 250 genomic regions of CUPID v1). Therefore, in some implementations, the genomic regions include those common to both Table 2 and Table 8. For example, the genomic regions may include at least 150, at least 200, or all 224 of the genomic regions common to both Table 8 and Table 2.

[0192] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75%, or at least 90% of the genomic regions used in any aspect of the present invention are selected from those shown in any of Tables 1, 2, 5, 6, 7, 8, and 9.

[0193] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75%, or at least 90% of the genomic regions used in any aspect of the invention are selected from those shown in Table 1 or Table 2.

[0194] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75%, or at least 90% of the genomic regions used in any aspect of the present invention are selected from those shown in Tables 5, 6, 7, 8, or 9.

[0195] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75%, or at least 90% of the genomic regions used in any aspect of the present invention are selected from those shown in Tables 6, 7, 8, or 9.

[0196] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75%, or at least 90% of the genomic regions used in any aspect of the invention are selected from those shown in Table 6.

[0197] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75%, or at least 90% of the genomic regions used in any aspect of the invention are selected from those shown in Table 7.

[0198] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75%, or at least 90% of the genomic regions used in any aspect of the invention are selected from those shown in Table 1.

[0199] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75%, or at least 90% of the genomic regions used in any aspect of the present invention are selected from those shown in Table 5.

[0200] In some embodiments, the genomic regions do not include any of the genomic regions in Table 7. In some embodiments, the genomic regions do not include any of the genomic regions in Table 2. In some embodiments, the genomic regions do not include any of the genomic regions in Table 6. In some embodiments, the genomic regions do not include any of the genomic regions in Tables 2 and 6.

[0201] In some implementations, the genomic regions do not include any of the genomic regions in Table 8.

[0202] In some implementations, the genomic regions do not include any of the genomic regions in Table 9.

[0203] In some embodiments, the genomic regions are selected from those listed in Tables 1 and / or 5 but not in Tables 2, 6, 7, 8, and 9. In some embodiments, the genomic regions are selected from those listed in Tables 1 and / or 5 but not in Table 7. In some embodiments, the genomic regions are selected from those listed in Tables 1 and / or 5 but not in Table 2. In some embodiments, the genomic regions are selected from those listed in Tables 1 and / or 5 but not in Table 6. In some embodiments, the genomic regions are selected from those listed in Tables 1 and / or 5 but not in Table 8. In some embodiments, the genomic regions are selected from those listed in Tables 1 and / or 5 but not in Tables 2 and 6. In some embodiments, the genomic regions are selected from those listed in Tables 1 and / or 5 but not in Table 9.

[0204] In some embodiments, less than 50% of the genomic regions are selected from those shown in Table 2. In some embodiments, less than 50% of the genomic regions are selected from those shown in Table 6. In some embodiments, less than 50% of the genomic regions are selected from those shown in Table 7. In some embodiments, less than 50% of the genomic regions are selected from those shown in both Tables 2 and 6. In some embodiments, less than 50% of the genomic regions are selected from those shown in Table 8. In some embodiments, less than 50% of the genomic regions are selected from those shown in Table 9.

[0205] In some embodiments, less than 40%, optionally less than 30%, of the genomic regions is selected from those shown in Table 2. In some embodiments, less than 40%, optionally less than 30%, of the genomic regions is selected from those shown in Table 6. In some embodiments, less than 40%, optionally less than 30%, of the genomic regions is selected from those shown in Table 7. In some embodiments, less than 40%, optionally less than 30%, of the genomic regions is selected from those shown in both Tables 2 and 6. In some embodiments, less than 40%, optionally less than 30%, of the genomic regions is selected from those shown in Table 8. In some embodiments, less than 40%, optionally less than 30%, of the genomic regions is selected from those shown in Table 9.

[0206] In some embodiments, less than 20% of the genomic regions are selected from those shown in Table 2. In some embodiments, less than 20% of the genomic regions are selected from those shown in Table 6. In some embodiments, less than 20% of the genomic regions are selected from those shown in Table 7. In some embodiments, less than 20% of the genomic regions are selected from those shown in both Tables 2 and 6. In some embodiments, less than 20% of the genomic regions are selected from those shown in Table 8. In some embodiments, less than 20% of the genomic regions are selected from those shown in Table 9.

[0207] In some implementations, genomic regions are selected based on a function of similarity between each primary malignancy type. For example, if the primary malignancy types originate from similar locations or histological subtypes, a larger number of genomic regions can be selected to improve the accuracy of distinguishing between two similar primary malignancy types.

[0208] In some embodiments, the training data includes regional methylation signatures derived from primary malignant tumor tissue samples obtained from multiple individuals known to have primary malignant cancer. In other words, the DNA methylation status of genomic regions in the training data may include the DNA methylation status of genomic regions obtained from primary malignant tumor tissue samples. Primary malignant tumor tissue samples may be obtained from multiple individuals diagnosed with a specific primary malignant cancer. In other embodiments, the training data includes regional methylation signatures of cfDNA derived from multiple individuals diagnosed with a specific primary malignant cancer. Alternatively, the training data may include regional methylation signatures of cfDNA derived from multiple individuals previously labeled as having CUP but subsequently having their primary cancer identified. It should be understood that in such embodiments, the training data includes regional methylation signatures of cfDNA derived from liquid samples from multiple individuals.

[0209] In some implementations, the regional methylation features of each primary malignant tumor type in the training data include the DNA methylation status of genomic regions that are different from the regional methylation features of each other primary malignant tumor type.

[0210] In the context of this invention, it should be understood that the training data includes regional methylation features for each primary malignant tumor type. The regional methylation features for each primary malignant tumor type may include the DNA methylation state of at least one genomic region derived from a primary malignant tumor tissue sample and the DNA methylation state of at least one genomic region derived from at least one non-cancer control cfDNA. This may also be referred to as a “computer-simulated mix.” Advantageously, the computer-simulated mix simulates the proportion of circulating tumor (ct) DNA typically expected in cfDNA, but using data from more readily available tumor tissue samples. Furthermore, the computer-simulated mix enables the generation of a large dataset from initial small volumes of DNA data from primary malignant tumor tissue samples on which the model can be trained. Moreover, it is difficult to determine how much DNA from primary malignant tumor tissue samples might shed into the biological fluid of the object to form ctDNA. Therefore, by “diluting” the DNA methylation state of the genomic region derived from the primary malignant tumor tissue sample with the DNA methylation state derived from at least one non-cancer control cfDNA, this provides higher accuracy for the training data and thus for the model.

[0211] The methylation status of genomic regions can be obtained from MBD-Seq reads. In some implementations, MBD-Seq reads are equivalent MBD-Seq reads obtained by converting β values ​​to equivalent MBD-Seq reads. As those skilled in the art will understand, β values ​​are a known quantitative form of methylation sequencing. β values ​​are methylation estimates measured by probes. β values ​​can be obtained from methylation arrays, bisulfite sequencing, or enzymatic methylation sequencing.

[0212] Converting β values ​​from a methylation array to equivalent MBD-Seq reads can be done as described by Chemi et al., 2022, the entire contents of which are incorporated herein by reference. For example, converting β values ​​to equivalent MBD-Seq reads may include calculating the β values ​​for an MBD-Seq window using a suitable procedure (e.g., qsea). The β values ​​used for fitting can be obtained by calculating a typical average enrichment profile of the T7-MBD-Seq method. In particular, the β values ​​used for fitting can be obtained by calculating an average enrichment profile from multiple non-cancer control cfDNA samples. The average enrichment profile may include fitted β values ​​for each window CG density, rounded to one decimal place, for different numbers of counts in each window. This profile can then be used for comparison to determine how many counts are needed for each β value at each CG density and can be used to convert the β values ​​to estimated T7-MBD-seq counts.

[0213] Machine learning models

[0214] In some implementations, the machine learning model includes one or more of decision trees, logistic regression models, artificial neural networks, support vector machines (SVM), Naive Bayes, linear discriminant analysis, or k-nearest neighbors algorithms, optionally wherein the decision tree includes gradient boosting or random forest algorithms.

[0215] In some implementations, the machine learning model includes one or more of decision trees or logistic regression models.

[0216] Logistic regression models can include penalized logistic regression models. Penalized logistic regression models can include L1 (lasso logistic regression) penalty terms, L2 (ridge logistic regression) penalty terms, or a combination of L1 and L2 penalties (elastic network logistic regression).

[0217] In some implementations, the machine learning model includes a decision tree containing a gradient boosting algorithm.

[0218] In some implementations, the machine learning model includes at least 51, at least 52, at least 53, at least 54, at least 55, at least 56, at least 57, at least 58, or at least 59 individual classifiers. The machine learning model may include at least 60 individual classifiers, and optionally at least 100 individual classifiers. In some implementations, the machine learning model includes 100 individual classifiers.

[0219] In some implementations, each individual classifier contains a limiting gradient boosting tree.

[0220] In some implementations, the output includes a probability score for each category. In other implementations, the output includes a ranking probability score for each category.

[0221] When the probability score for the primary malignant tumor type classification is greater than 0.5, the output may include the primary malignant tumor type classification of the sample data. When the probability score for the unknown type classification is greater than 0.5, or when the probability score for each classification is less than 0.5, the output may include the unknown type classification of the sample data.

[0222] In some implementations, when the probability score for the primary malignant tumor type classification is greater than 0.3, the primary malignant tumor type classification containing the sample data is output. In some implementations, when the probability score for the primary malignant tumor type classification is greater than 0.4, the primary malignant tumor type classification containing the sample data is output.

[0223] In some implementations, the output includes a ranking probability score for each category, and the subject's primary cancer is determined to be the primary malignant tumor type category with the highest ranking probability score. In implementations where the output includes a ranking probability score for each category, if the difference in probability scores between the highest and second-ranked primary malignant tumor type categories is greater than at least about 0.2, the subject's primary cancer can be determined to be the highest-ranked primary malignant tumor type category. In implementations where the output includes a ranking probability score for each category, if the difference in probability scores between the highest and second-ranked primary malignant tumor type categories is less than about 0.05, optionally less than about 0.04, and further optionally less than about 0.03, the subject's primary cancer can be determined to be an unknown or unclassified cancer. In other implementations, if the difference in probability scores between the highest and second-ranked primary malignant tumor type categories is less than about 0.05, optionally less than about 0.04, and further optionally less than about 0.03, the subject's primary cancer can be determined to be either the first or second-ranked primary malignant tumor type category. In some implementations, if the probability score difference between the highest-ranked and second-ranked primary malignant tumor type classifications is less than about 0.3, optionally less than about 0.2, and further optionally less than about 0.1, the subject's primary cancer can be identified as belonging to the first or second-ranked primary malignant tumor type classification.

[0224] For example, if:

[0225] a) The probability score difference between the highest-ranked and second-ranked primary malignancy type classifications is less than about 0.05, optionally less than about 0.04, and further optionally less than about 0.03; and

[0226] b) The primary malignant tumor types classified as the highest and second highest are primary malignant tumor types with the same or similar locations.

[0227] The primary cancer of the subject can then be classified as the first or second ranked primary malignant tumor type.

[0228] Examples of two primary malignant tumor types with the same or similar anatomical locations are cholangiocarcinoma and pancreatic adenocarcinoma. Cholangiocarcinoma can also be referred to as cancer of the bile duct. Therefore, cholangiocarcinoma and pancreatic adenocarcinoma share the same location (the pancreatic region).

[0229] In some implementations, the output primary malignant tumor type classification corresponds to the primary malignant tumor types on which the model was trained. For example, in an implementation where the output includes a probability score for each classification, the classification may correspond to the primary malignant tumor types on which the model was trained. In an implementation where the output includes a primary malignant tumor type classification, the classification may correspond to one of the primary malignant tumor types on which the model was trained.

[0230] Alternatively, the output primary malignancy type classification may include a broader classification than the primary malignancy types used to train the model. A broader classification may include classifications specifying a wider anatomical region. A broader classification may not distinguish histological subtypes. Exemplary broader classifications may include, but are not limited to, brain cancer, breast cancer, female reproductive cancer, blood cancer, kidney cancer, lower abdominal cancer, lung cancer, male reproductive cancer, neuroendocrine tumors, rare cancers, skin cancer, thyroid cancer, upper abdominal cancer, and unclassified primary cancers. A broader classification may not include non-cancer categories. While this invention can be used to accurately determine the anatomical location and histological subtype of primary cancer, the inventors have also found it applicable to the broad classification of cancers. Such broad classifications facilitate initial stratification of subjects to determine initial treatment strategies or may aid in further testing to specify a particular primary malignancy type.

[0231] In some implementations, the probability scores for each category are summed to form a cumulative score for one or more broader cancer categories. The subject's primary cancer may be identified as belonging to the broader cancer category with the highest-ranking cumulative score. In some implementations, the subject's primary cancer is identified as belonging to a broader cancer category with a cumulative score of at least about 0.5. In some implementations, the subject's primary cancer is identified as belonging to a broader cancer category with a cumulative score greater than 0.5.

[0232] Determining a broader cancer classification may include the final step in the first layer of a two-layer computer-implemented model, where steps (i) through (iv) represent the first layer. Therefore, in an implementation where the primary cancer of an object is determined to be a broader cancer classification, the method may further include step (v) inputting the output from step (iii) into a second machine learning model containing a multi-class classifier trained on training data as described herein. The training data may be the same as or different from the training data used to train the machine learning model in step (ii). The method may further include step (vi), which includes receiving the output from the second machine learning model, which indicates the primary malignancy type classification of the sample data, and determining the primary cancer of the object based on the output received in step (vi), wherein the primary cancer is a cancer subtype with different anatomical and histological subtypes.

[0233] Sample data

[0234] It should be understood that the regional methylation signature of the sample data includes the DNA methylation status of at least one region of the cfDNA genome. As used herein, “DNA methylation status” will be understood to refer to the level of DNA methylation. DNA methylation can be either hypermethylated or hypomethylated.

[0235] In some implementations, the regional methylation characteristics of sample data include the DNA methylation status of multiple genomic regions.

[0236] The regional methylation characteristics of sample data can be defined as those of the training data. In some implementations, the regional methylation characteristics of sample data include the DNA methylation status of genomic regions that are identical to those of the genomic regions in the training data. As used herein, “identical genomic regions” will be understood to refer to two or more genomic regions that have the same length and genomic coordinates but originate from different samples, such as those from a liquid biopsy sample and a primary malignant tumor sample, or those from two or more different primary malignant tumor samples. Therefore, while “identical” genomic regions may have the same length and genomic coordinates, they may each have different DNA methylation statuses.

[0237] Alternatively, the regional methylation features of the sample data may include the DNA methylation status of a subset of genomic regions used in the training data. While predictive performance may be reduced when the sample data is “incomplete” in terms of genomic regions, the impact is expected to be minor when the sample data contains most of the genomic regions used in the training data. For example, the regional methylation features of the sample data may include the DNA methylation status of at least 75%, at least 80%, at least 85%, at least 90%, or at least 95% of the genomic regions used in the training data.

[0238] Liquid samples can include any biological fluid. For example, liquid samples can include blood samples, plasma samples, serum, lymph, synovial fluid, ascites, interstitial or extracellular fluid, cerebrospinal fluid, saliva, mucus, semen, sweat, urine, or any other bodily fluid. In some embodiments, the liquid sample includes blood, urine, or plasma samples. In some embodiments, the liquid sample includes blood or plasma samples. In some embodiments, the liquid sample includes blood samples. In some embodiments, the liquid sample includes plasma samples.

[0239] Samples may be freshly obtained from the subject or may be samples that have been processed and / or stored prior to detection (e.g., frozen, fixed, or subjected to one or more purification, enrichment, or extraction steps). Samples are preferably from mammals (e.g., mammalian cell samples or samples from mammalian subjects, including, in particular, model animals such as mice, rats, etc.), and preferably from humans (e.g., human cell samples or samples from human subjects). Furthermore, samples may be transported and / or stored, and collection may be performed at a location remote from the genomic sequence data acquisition (e.g., sequencing) location, and / or computer-implemented method steps may be performed at a location remote from the sample collection location and / or remote from the genomic data acquisition (e.g., sequencing) location (e.g., computer-implemented method steps may be performed via a networked computer, such as via a "cloud" provider).

[0240] In some embodiments, the sample data also includes DNA sequence data obtained from a tissue sample from the subject. In such embodiments, it should be understood that the sample includes both liquid and tissue samples. Tissue samples can be obtained from any biological tissue. For example, tissue samples may include liver tissue, lymph node tissue, bone tissue, lung tissue, skin tissue, gastrointestinal tissue, or tissue from any other suitable organ or site. A person skilled in the art will know the suitable organ / site from which a tissue sample can be obtained. In some embodiments, the tissue sample includes a tumor sample. A “tumor sample” refers to a sample containing tumor cells or genetic material derived from tumor cells. A tumor sample can be a cell or tissue sample obtained directly from a tumor (e.g., a biopsy).

[0241] Tissue samples may include fresh tissue samples, frozen tissue samples (e.g., frozen at -20°C or -80°C), or formalin-fixed paraffin-embedded tissue samples.

[0242] In some implementations, the sample data also includes sequence data of DNA obtained from circulating tumor cells in a liquid sample.

[0243] In some embodiments, the liquid sample comprises 1 pg (i.e., picogram) to 1 µg, such as 1 pg to 200 ng or 1 ng to 200 ng of cfDNA, and / or the cfDNA sample comprises less than 10,000, less than 1,000, less than 300 or less than 100 haploid human genome equivalents.

[0244] In some embodiments, the sample data also includes the tumor fraction (TF) of cell-free DNA (cfDNA) obtained from the liquid sample. The TF of the sample data can be at least about 1%, at least about 1.5%, at least about 2%, at least about 2.5%, at least about 3%, at least about 3.5%, or at least about 4%. In some embodiments, the TF of the sample data is no more than about 80%, no more than about 75%, or no more than about 70%. In some embodiments, the TF of the sample data is from about 1% to about 75%.

[0245] As used in this article, the term “tumor score (TF)” will be understood as the proportion of cancer-derived DNA in cell-free DNA.

[0246] The inventors have advantageously discovered that while it is possible to identify the primary cancer of an object from sample data containing very low TF (e.g., greater than 0 but less than about 2% TF) based on the output of a machine learning model, the proportion of primary cancers identified as belonging to an unknown type classification in this group is higher than that from sample data containing higher TF.

[0247] In some embodiments, the sample data contains at least about 3% TF. The inventors have advantageously discovered that at least about 3% TF reduces the risk of the primary cancer being classified as unclassified and therefore unknown.

[0248] In some embodiments, the sample data also includes mutation profiles. In some embodiments, the training data also includes mutation profiles from at least two primary malignant tumor types. In some embodiments, the training data also includes mutation profiles from at least one non-cancer control.

[0249] A mutation spectrum can include the mutation states of multiple genes.

[0250] Many genes are known to mutate in certain cancers. Therefore, technicians will be able to easily select genes. For example, genes may include, but are not limited to, ALK, APC, AR, ARID1A, ATM, ATRX, GRIN2A, KDR, KMT2C, KMT2D, KRAS, LRP1B, MED12, NF1, NOTCH1, PIK3CA, RB1, SMAD4, SPEN, TP53, BRAF, BRCA1 / 2, EGFR, and IDH1.

[0251] In some implementations, the genes include PIK3CA, BRAF, and IDH1.

[0252] In other embodiments, the gene is selected from ALK, APC, AR, ARID1A, ATM, ATRX, GRIN2A, KDR, KMT2C, KMT2D, KRAS, LRP1B, MED12, NF1, NOTCH1, PIK3CA, RB1, SMAD4, SPEN, and TP53. For example, the gene may include ALK, APC, AR, ARID1A, ATM, ATRX, GRIN2A, KDR, KMT2C, KMT2D, KRAS, LRP1B, MED12, NF1, NOTCH1, PIK3CA, RB1, SMAD4, SPEN, and TP53.

[0253] The mutation state can include a mutated or non-mutated state. A mutation state can include at least one non-synonymous mutation. Mutation states can include frameshift deletions, frameshift insertions, splice site mutations, translation start site mutations, nonsense mutations, non-termination mutations, in-frame deletions, and / or missense mutations.

[0254] In some implementations, the sample data also includes somatic copy number alteration (SCNA) profiles. As those skilled in the art will understand, somatic copy number alterations are known to occur in certain cancers. "Somatic copy number alteration" will be understood as a dose change in certain genes, chromosome arms, or the entire chromosome relative to a non-cancer control. SCNA profiles for various tumor samples are publicly available and known, for example, from the Cancer Genome Atlas and the International Cancer Genome Consortium. In some implementations, the training data also includes SCNA profiles from at least two primary malignant tumor types.

[0255] In some embodiments, the sample data also include fragmentomics profiles. As those skilled in the art will appreciate, fragmentomics is a term used in studies describing the size of cfDNA fragments in subjects with cancer. In particular, recent studies have shown that the fragment sizes of cfDNA in cancer patients are much more diverse than those in non-cancer controls. Therefore, in some embodiments, the training data also includes fragmentomics profiles from at least two primary malignant tumor types. In some embodiments, the training data also includes fragmentomics profiles from at least one non-cancer control.

[0256] Methods for identifying the primary cancer in a patient, methods for selecting cancer treatment, and treatment methods.

[0257] According to a second aspect, the present invention provides a method for identifying primary cancer in a subject having a malignant tumor of unknown origin (MUO), the method comprising:

[0258] i) Analyze cell-free DNA (cfDNA) obtained from liquid samples from the subject to obtain sequence data containing regional methylation features (“sample data”); and

[0259] ii) The sample data are subjected to a first aspect method to determine the primary cancer of the subject.

[0260] In some implementations, analyzing cfDNA obtained from a liquid sample includes methylation sequencing of multiple genomic regions of the cfDNA. These multiple genomic regions may be as defined above with respect to the first aspect.

[0261] In some embodiments, methylation sequencing includes non-destructive methylation sequencing, such as non-destructive enzyme-based conversion methods to distinguish unmethylated cytosine from 5-methylcytosine and / or 5-hydroxymethylcytosine. This may also be referred to as enzymatic methyl sequencing (EM-seq). In some embodiments, methylation sequencing includes a bisulfite conversion step (sometimes referred to as "bisulfite sequencing"). In some embodiments, methylation sequencing includes direct measurement of base modifications, for example on a single-molecule sequencing platform, such as a nanopore sequencing platform or a PacBio sequencing platform.

[0262] In some implementations, methylation sequencing includes methylated DNA immunoprecipitation sequencing (MeDIP-seq). As those skilled in the art will understand, MeDIP-seq utilizes a 5-methylcytosine monoclonal antibody to capture methylated fragments from single-stranded DNA.

[0263] In some implementations, methylation sequencing includes methyl-binding domain sequencing (MBD-seq). As used herein, MBD-seq will be understood to refer to the capture of double-stranded methylated DNA fragments using methyl-binding domains from MECP2.

[0264] In some preferred embodiments, methylation sequencing includes T7-MBD-seq, as described by Chemi et al. 2022, the entire contents of which are incorporated herein by reference. In short, T7-MBD-seq involves barcoding and merging fragmented DNA, followed by incubation with MBD2. Next-generation sequencing (NGS) libraries can then be generated for the methylation-enriched portions.

[0265] Therefore, in some embodiments, analyzing cfDNA includes performing methylation sequencing and next-generation sequencing. Preferably, analyzing cfDNA includes performing T7-MBD-seq and next-generation sequencing. The next-generation sequencing may include Illumina sequencing.

[0266] The cfDNA can be amplified and / or library prepared prior to methylation sequencing. In some embodiments, library preparation is single-stranded DNA (ssDNA) library preparation. In some embodiments, the cfDNA can be fragmented prior to methylation sequencing.

[0267] Step i) may include analyzing cell-free DNA (cfDNA) obtained from a liquid sample from the subject to obtain sequence data containing regional methylation features (“sample data”) and a mutation profile. The mutation profile may be as defined above with respect to the first aspect. In such embodiments, analyzing cfDNA may further include analyzing the mutation status of at least 10, at least 20, at least 50, at least 100, or at least 500 genes.

[0268] In some embodiments, step i) includes analyzing cell-free DNA (cfDNA) obtained from a liquid sample from the subject to obtain sequence data (“sample data”) containing regional methylation features and a tumor score (TF). The tumor score may be as defined above with respect to the first aspect. In such embodiments, the analysis of cfDNA may also include performing tumor score analysis, for example using ichorCNA.

[0269] The sequence data may also include fragment omics profiles and / or SCNA profiles, both of which may be defined as above with respect to the first aspect. In such embodiments, the analysis of cfDNA may also include taking into account the size distribution of the sequenced DNA fragments and / or the genomic background at the site where the DNA fragments are cleaved.

[0270] Considering the size distribution of sequenced DNA fragments may include filtering computer-simulated sequenced DNA fragments to a tumor-enriched size according to their length using computer simulation. Those skilled in the art will know the appropriate tumor-enriched size. Not wishing to be bound by theory, the inventors believe that sequenced DNA fragments filtered to a tumor-enriched size may include more tumor DNA, and therefore the DNA methylation status of genomic regions has improved accuracy in indicating the type of primary malignancy.

[0271] Therefore, in some implementations, sequence data is derived from DNA fragments pre-filtered from cfDNA to tumor-enriched sizes via fragmentomics.

[0272] According to a third aspect, the present invention provides a method for selecting cancer treatment for a subject with a malignant tumor of unknown origin (MUO), the method comprising:

[0273] i) Using the first approach to determine the most probable primary cancer in the subject; and

[0274] ii) Select anticancer treatment targeting the most likely primary cancer.

[0275] According to another aspect, the present invention provides a method for treating a subject with a malignant tumor of unknown origin (MUO), the method comprising:

[0276] i) Using the first approach to determine the most probable primary cancer in the subject; and

[0277] ii) Administer anticancer treatment targeting the most likely primary cancer.

[0278] As used in this article, “treatment” refers to the reduction, relief, or elimination of one or more symptoms of the cancer being treated, relative to the symptoms before treatment.

[0279] The subject can be a mammal (e.g., human, non-human primate, cat, dog, horse, donkey, sheep, pig, goat, cow, mouse, rat, rabbit, or guinea pig). The subject is preferably a human. The subject can be an adult (at least 18 years old). In some embodiments, the subject is a human child (under 18 years old).

[0280] Cancer treatment may include surgery, radiation therapy, chemotherapy, targeted therapy, or immunotherapy. In some embodiments, cancer treatment includes immunotherapy. Those skilled in the art will know of various immunotherapies, and these are commercially available. In some embodiments, immunotherapy includes pembrolizumab, atezolizumab, nivolumab, ipilimumab, or combinations thereof.

[0281] The individual may have already undergone pretreatment with chemotherapy agents.

[0282] When administered anticancer therapy to subjects, the size of the primary cancer tumor can be reduced by approximately 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, or even 100% compared to untreated tumors.

[0283] Cancer treatment can be administered via any suitable route. For example, it can be administered subcutaneously, intranasally, orally, topically, intraperitoneally, or intravenously. In some implementations, cancer treatment is administered intraperitoneally.

[0284] In implementation plans for anticancer therapy involving genetically engineered cell populations / immune cell populations, the number of cells administered to the subject should take into account the route of administration, the primary cancer being treated, the subject's weight, and / or the subject's age. Generally, approximately 1 × 10⁻⁶ cells are administered to the subject. 6 To approximately 1×10 11 One genetically engineered cell / immune cell. In some implementations, approximately 1 × 10⁻⁶ cells are administered to the subject. 7 To approximately 1×10 10 One genetically engineered cell / immune cell, or approximately 1 × 10⁻⁶ 8 To approximately 1×10 9 Genetically engineered cells / immune cells.

[0285] Systems, computer-readable media, and reagent kits

[0286] According to another aspect, the present invention provides a system comprising:

[0287] iii) the processor; and

[0288] iv) A computer-readable medium containing instructions that, when executed by the processor, cause the processor to perform the steps of the method of the first aspect.

[0289] Preferably, the system includes an Illumina sequencing platform system. The Illumina sequencing platform system will be understood as a system containing instructions that, when executed by a processor, cause the processor to perform the steps of Illumina sequencing.

[0290] Preferably, the system includes a computer-readable medium containing instructions that, when executed by the processor, cause the processor to:

[0291] a) Analyze cell-free DNA (cfDNA) obtained from liquid samples of the subject to obtain sequence data containing regional methylation features; and

[0292] b) The steps of performing the method in the first aspect.

[0293] Step a) can be as defined with respect to any of the foregoing aspects. For example, analyzing cfDNA may include performing methylation sequencing and optionally next-generation sequencing. Next-generation sequencing may include Illumina sequencing.

[0294] The present invention also provides one or more computer-readable media comprising instructions that, when executed by one or more processors, cause one or more processors to perform the steps of the method of the first aspect.

[0295]

[0296] Features disclosed in the foregoing description, in the following claims, or in the accompanying drawings, expressed in their specific form, or in the manner of performing the disclosed function, or in the form of a method or process for obtaining the disclosed result, may, as appropriate, be used alone or in any combination of such features to implement the invention in its various forms.

[0297] While the invention has been described in conjunction with the exemplary embodiments described above, many equivalent modifications and variations will be apparent to those skilled in the art in light of this disclosure. Therefore, the exemplary embodiments of the invention shown above are intended to be illustrative rather than restrictive. Various changes may be made to the described embodiments without departing from the spirit and scope of the invention.

[0298] To avoid any doubt, any theoretical explanations provided herein are offered for the purpose of improving the reader's understanding. The inventors do not wish to be bound by any of these theoretical explanations.

[0299] Any chapter headings used in this document are for organizational purposes only and should not be construed as limiting the subject matter described.

[0300] Throughout this specification (including the appended claims), unless the context otherwise requires, the words “comprising” and “including” and variations thereof shall be construed as implying the inclusion of the said integer or step or group of integers or steps but excluding any other integer or step or group of integers or steps.

[0301] It should be noted that, unless the context clearly indicates otherwise, nouns without quantifiers, as used in this specification and the appended claims, include plural pronouns. A range herein may be expressed as from “about” a particular value, and / or to “about” another particular value. When such a range is expressed, another embodiment includes from one particular value and / or to another particular value. Similarly, when a value is expressed as an approximation using the antecedent “about,” it will be understood that the particular value forms another embodiment. The term “about” in relation to numerical values ​​is optional and means, for example, + / - 10%.

[0302] Example

[0303] introduction

[0304] We previously described a robust and sensitive genome-wide circulating cell-free DNA (cfDNA) methylation profiling workflow (T7-MBD-seq) that detects circulating tumor DNA (ctDNA) from patients with early-stage small cell lung cancer (SCLC) and distinguishes molecular subtypes (Chemi et al., 2022). Furthermore, several early cancer detection studies have demonstrated that cfDNA methylation patterns can predict tissue of origin (TOO) with high accuracy (Moss et al., 2018, Liu et al., 2020, and Klein et al., 2018). Here, we generate and test a high-accuracy TOO classifier derived from methylation profiling and apply it to a pilot CUP cohort in this proof-of-concept study. We identify the potential utility of cfDNA analysis from a single blood sample (combining mutation detection with TOO prediction) in facilitating diagnostic and treatment stratification.

[0305] Materials and methods

[0306] Data availability

[0307] The T7-MBD-seq data, shallow whole-genome sequencing, and targeted NGS data supporting the findings of this study have been stored in the European Genome-Phenome Archive (EGA), accession number EGAS00001007445. Pre-normalized TCGA data were downloaded from the Xena browser (https: / / tcga-pancan-atlas-hub.s3.us-east-1.amazonaws.com / download / jhu-usc.edu_PANCAN_HumanMethylation450.betaValue_whitelisted.tsv.synapse_download_5096262.xena.gz). Previously published cholangiocarcinoma methylation arrays were downloaded from the Gene Expression Omnibus (GSE32079, GSE49656, GSE89803).

[0308] Patient recruitment and sample collection

[0309] Cancer patients were recruited through the TARGET (Tumour Characterisation to Guide Experimental Targeted Therapy) trial (ethical approval REC reference: 15 / NW / 0078) (Rothwell et al., 2019) and the CUP Biobank at Cancer Research Manchester, a UK Level 3 cancer centre under the Christie NHS Foundation Trust (ethical approval number 18_ALCO_01). Non-cancer control (NCC) samples were collected from three sources: 1. Community Lung Health Study (ethical approval REC reference: 17 / LO415); 2. University of Manchester Healthy Volunteer Study (University of Manchester Ethics Committee Approval No. 2017-2761-4606); or 3. Purchased from Cambridge Bioscience (Ethics Committee Approval No. 2019-7920-11797).

[0310] CUP Clinical, Radiological and Pathological Review and Data Collection

[0311] Clinical data for the entire CUP cohort were acquired retrospectively and anonymized. All patients were discussed at CUP-specific multidisciplinary team (MDT) meetings at the tertiary cancer center (Christie NHS Foundation Trust) in accordance with the National Institute for Health and Care Excellence (NICE) guidelines (NICE, 2010). Therefore, patients underwent histopathological and radiological reviews to confirm the initial diagnosis of CUP. All clinical, pathological, and radiological investigations were reviewed along with any MDT meeting. Subsequent primary tumor diagnosis status was determined as follows: a "primary tumor diagnosis" was confirmed if a primary tumor diagnosis was prospectively documented at any point from recruitment to death or data lockout (August 2022) and the patient received subsequent treatment based on that diagnosis. In cases where a single primary tumor diagnosis was prospectively or retrospectively suspected, but the patient did not receive treatment based on that suspicion due to remaining uncertainty, the patient was classified as having a "highly suspected tumor type." The "differential diagnosis" category includes patients whose clinicopathological features can narrow down the possible diagnoses to two or more different tumor types. Figure 1 ).

[0312] Blood sample collection

[0313] Blood samples were collected in up to 4 × 10 mL cell-free DNA BCT tubes (Streck, Omaha, NE) for cfDNA analysis. Plasma was separated from whole blood by two consecutive centrifugations (2,000 g, 10 min) within 4 hours of collection and stored at -80°C before cfDNA processing. Additionally, according to the study protocol, 1 × 10 mL BD vacuum blood collection tubes were collected for K₂ ethylenediaminetetraacetic acid (K₂EDTA) sampling for germline DNA extraction.

[0314] Cell-free DNA extraction and quantification

[0315] According to the manufacturer's instructions, cfDNA can be isolated from up to 20 mL of plasma using one of the following three isolation techniques: 1. QIAmp MinElute ccfDNA MIDI Kit (Qiagen, catalog number 55284); 2. QIAsymphony with Circulating DNA Kit (Qiagen, catalog number 1091063); or 3. QIAmp Circulating Nucleic Acid Kit (Qiagen, catalog number 55154). The cfDNA yield is quantified using the TaqMan RNase P Assay Kit (Life Technologies, catalog number 4316831).

[0316] T7-MBD-seq library preparation and next-generation sequencing (NGS)

[0317] All cfDNA samples (1 to 35 ng cfDNA) were processed using the T7-MBDSeq method as previously described (Chemi et al., 2022), and the libraries were sequenced at both ends on an Illumina NextSeq 500 or NovaSeq 6000. For each sample, both the methylation enriched fraction and the non-enriched fraction were sequenced for whole-genome methylation profiling and copy number analysis, respectively.

[0318] T7-MBD-seq read segment comparison

[0319] The Nextflow (v22.04.5) DSL2 process (Ewels et al. 2020), built using tools and modules provided by the nf-core community, is used to process FASTQ files and generate QSEA objects, as detailed below.

[0320] All reads were trimmed to ensure that the initial lengths of R1 and R2 were uniformly 91 and 61 base pairs (bp) respectively (including the 26 bp T7-MBD-seq construct containing the R1 start position). Unique molecular identifiers (UMIs) were extracted using umitools (Smith et al., 2017, v1.1, 2), and demultiplexing and adapter trimming were performed on the samples using cutadapt (Martin et al., 2011, v3.4). Reads were aligned to the GRCh38 reference genome using bwa mem (Li, 2013, v0.7.17), and repeat removal was performed by combining the R1 start position with the UMI using umi-tools (Smith et al., 2017, v1.1.2). Pairing quality scores were assigned using samtools fixmate (Li et al., 2009, v1.15.1) to generate the final BAM file.

[0321] Methylation enrichment analysis

[0322] The QSEA package (Lienhard et al., 2017) (v1.16) was used to analyze BAM files, and a custom R package (v4.2.0) (MESA, Methylation Enrichment Sequencing Analysis, v0.2.1, available at www.github.com / crukmi / mesa) was built to extend QSEA. The genome was divided into 300 bp non-overlapping windows, with over-represented windows removed. Following the methodology used by the ENCODE consortium (Amemiya et al., 2019), over-represented windows were determined by analyzing 168 non-enriched NCC samples: a window was identified as over-represented if the fragment count within it was in the top 0.1% of the total fragment count, and if adjacent windows had a count in the top 1%. Here and elsewhere, fragments represent genomic locations between the two paired ends of a sequencing read.

[0323] Fragments were filtered to be paired reads only, where the mapping quality (MAPQ) score at either end of the pair was at least 10, the fragment length was 90 to 1000 bp, and the distance along the reference genome was at least 30 bp. Fragments were then uniquely assigned to windows based on their midpoint location. For use in QSEA, copy number variations (CNVs) were calculated per sample from non-enriched samples using HMMcopy (v1.32 Shah et al., 2006) with baseline parameters on a 1 Mbp window. Normalized reads per million (NRPM) were generated using CNVs and the effective fragment count in the sample, without applying trimmed mean of Mvalues ​​(TMM) normalization. In QSEA, β values ​​(a standardized measure of methylation between 0 and 1) were calculated for each window in each sample using a “blind calibration” method (Lienhard et al., 2017).

[0324] IchorCNA

[0325] Tumor fraction (TF) was estimated for each sample using non-enriched input cfDNA samples processed with IchorCNA (Adalsteinsson et al., 2017) (v0.32). The 79 NCC cfDNA samples used to generate CUPids were used as a non-cancer sample panel, and a 1 Mbp window size was applied without estimating the subclonal population. An estimated TF below 0.03 was considered below the limit of detection (Adalsteinsson et al., 2017).

[0326] Quality control

[0327] NGSCheckMate (Lee et al., 2017) (v1.0.0) was used to verify that all samples from individual patients matched as expected in the tool output. All four modalities (cfDNA and germline targeted sequencing, T7-MBD-seq methylated and unmethylated components) for each patient were examined, where applicable.

[0328] To calculate the relative enrichment score (relH) of the T7-MBD-seq samples, the method of the MEDIPS R package (Lienhard et al., 2014) (v1.42) was used. This method calculates the total CG density contained in the DNA location mapped on the reference sequence and divides it by the total CG density of the entire genome.

[0329] An additional QC metric, termed the “hyperstability score,” was calculated for adequate methylation capture. The proportion of samples with a β value of 0.8 or higher was calculated using 805 windows (Edgar et al., 2014) corresponding to persistently hypermethylated CpG sites in methylation array data from both cancer and non-cancer samples. This metric is a combination of the number of effective fragments and the global enrichment profile. For the validation set, samples with a relH value below 2.5 or a hyperstability score below 0.4 were excluded. This process removed 5 patient samples and 3 non-cancer control samples from the test set, while no samples were removed from the CUP cohort.

[0330] Convert methylation array data to qseaSets

[0331] Download the preprocessed beta value table of the TCGA pancanine methylation array dataset from the following address: https: / / tcga-pancan-atlas-hub.s3.us-east-1.amazonaws.com / download / jhu-usc.edu_PANCAN_HumanMethylation450.betaValue_whitelisted.tsv.synapse_download_5096262.xena.gz. This dataset consists of 9,639 arrays (including 721 adjacent normal tissues) spanning 33 tumor types.

[0332] Due to the insufficient representativeness of cholangiocarcinoma in the TCGA dataset, additional cholangiocarcinoma array data were obtained. Additional preprocessed β values ​​(accession numbers GSE89803 (Jusakul et al., 2017), GSE32079 (Wang et al., 2013), and GSE49656 (Chan-On et al., 2013)) were downloaded from the gene expression compilation using GEOquery (v2.66.0, Davis and Meltzer 2007), resulting in a total of 256 cholangiocarcinoma arrays.

[0333] Grouping of tumor samples

[0334] The TCGA pan-cancer methylation array dataset was filtered to exclude samples classified as recurrent, redacted, metastatic, extra tumors, or adjacent normal tissue (except liver, see below); for pancreatic ductal adenocarcinoma (PDAC), we removed 28 samples that might have been misclassified due to reasons such as extremely low tumor content or belonging to pancreatic tumor types different from PDAC [Peran et al. 2018]. This resulted in 8,797 tumor samples and 49 normal liver tissue samples. Arrays from all 33 TCGA categories (and additional cholangiocarcinoma samples from GEO) were regrouped into categories distinguished by both anatomical location and histological subtype (Table 3). This allowed the classifier to be trained on more histologically distinct categories and to merge categories with fewer but highly similar samples. For example, the TCGA esophageal cancer category (ESCA) is split by histological subtype: esophageal adenocarcinoma (n=88) samples are merged with gastric adenocarcinoma (STAD, n=393) into the upper gastrointestinal tract (upper GI) category, while esophageal squamous cell carcinoma (n=96) and head and neck squamous cell carcinoma (HNSC, n=525) are merged into the upper squamous cell category (upper Sq). When mixed histological tumor types exist within a category, these are excluded from training: seven hepatocellular carcinoma arrays labeled "hepatobiliary carcinoma (mixed)" are excluded, as are seven cervical and endocervical arrays labeled "adenosquamous carcinoma". Table 3 summarizes the categories and tumor subtypes grouped within this category.

[0335] The process generates 30 general categories, including a non-cancer category, which is incorporated to provide a neutral category for the classifier construction process, used to assign samples with low tumor scores, rather than forcing potentially incorrect predictions.

[0336] We then converted the methylation array probe level data into a window-based read format compatible with T7-MBD-seq read-based enrichment sequencing, as previously described (Chemi et al., 2022). First, all 79 NCC cfDNA samples were combined to calculate a typical average enrichment profile for our T7-MBD-Seq method, generating a lookup table that determines how many counts are required for each β value and each CG density and can be used to convert array β values ​​into estimated T7-MBD-seq counts. The largest β value was used when multiple probes were located within a single window. The resulting qseaSet estimates the expected counts when performing T7-MBD-seq on the sample within a window of overlapping array probes.

[0337] Uniform manifold approximation and projection

[0338] Uniform manifold approximation and projection were computed on the transformed tumor array (UMAP, McInnes et al., 2018) using the uwot package (v0.1.14, Melville et al., 2022) with parameters n_neighbors = 15 and min_dist = 1.

[0339] Classifier construction

[0340] To train the classifier, computer-simulated synthetic mixtures, “qseaSets,” were generated by mixing processed fragment counts between samples: array samples (converted to qseaSets as described above) were mixed with NCC cfDNA T7-MBD-seq samples at a ratio of 1% to 10%, or between two NCC T7-MBD-seq samples at a ratio of 15% to 50%, all with different fragment numbers (between 1 million and 10 million NCC fragments, whichever is lower). This was done by mixing each array sample (or NCC) once with each NCC, with the ratio and fragment number randomized. Early iterations of classifier development showed that some NCC mixtures were predicted as hepatocellular carcinoma. This was presumably due to the possible presence of high levels of normal liver tissue signaling in the normal cfDNA component of these samples. To address this, we included 49 neighboring normal liver arrays from TCGA into the NCC mixture set used for training, and these mixtures were assigned to NCC categories.

[0341] Using the QSEA package (Lienhard et al., 2017), differentially methylated regions (DMRs) were computed pairwise between classes on a transformed array of qseaSets with 100% tumor scores. A false discovery rate (FDR) of 0.001 was applied. The difference in mean β values ​​between classes in the comparisons, Δβ, was calculated and used to rank the DMRs according to effect magnitude. DMR groups were generated between each class, and the top 250 DMRs in each pairwise comparison were selected. After reducing DMRs appearing in multiple comparisons, 22,179 distinct windows were used for classifier development.

[0342] Up to 10,000 mixtures of qseaSets were sampled for each class, resulting in a total of 276,108 unique samples for classifier training, leading to approximately equal class distributions (3,460 to 10,000 samples per group, depending on the number of available arrays). For each sample, the normalized reads per million (NRPM) was calculated for each of the 22,179 regions; this was used as input data for each classifier. An ensemble of 100 classifiers was then built, each individual classifier consisting of a mixture of 80% array samples and 80% NCC samples. These individual classifiers were constructed within the R tidymodels (v0.2.0) framework using extreme gradient boosting trees (Chen and Guestrin, 2016) (xgboost R package, v1.6.0.1), with default parameters except for the number of trees (trees) = 200, sample size = 0.5, and mtry = 2135. These parameter choices resulted in 200 constructed sequential trees, each using a randomly selected 50% mixture (within 80% stratification) and 10% region. These parameters provide a significant amount of variation within the tree population for each classifier, as well as variation between individual classifiers based on the mixtures they encounter.

[0343] Model performance

[0344] Each individual sub-classifier is tested on the remaining mixture set, which contains samples that the sub-classifier did not see during model training (representing 4% (20%) of the total mixture set). (20%). The subclassifier is applied to the retained mixture set, producing a predicted score for each of the 30 classes. These class-specific predicted scores for all retained mixture sets are compared to the ground truth to determine the area under the one-vs-all subject operating characteristic curve (AUROC) value for each subclassifier. For this, we use the multiclass roc_auc function from the R package yardstick [v1.0.0, Kuhn and Vaughan, 2022], which is based on the method in Hand and Till: A Simple Generalization of the Area Under the ROC Curve for Multiple Class Classification Problems. Machine Learning 45, 171-186 (2001).

[0345] For ensemble classifier performance, we take each individual mixture set and apply all subclassifiers trained without using either half of the mixture. We then calculate the average predicted score for each of the 30 classes (averaging the outputs of the relevant subclassifiers). The overall AUROC value for the ensemble is then calculated again using the multi-class Hand-Till method.

[0346] Application of CUPiD in cfDNA samples treated by T7-MBD-Seq

[0347] The trained classifier ensemble (CUPiD) was then applied to 143 cfDNA samples from patients with known tumor types, 27 NCC samples not used to generate CUPID, and 41 samples from CUP patients. The average of each prediction from the 100 individual classifiers was used as the final prediction score for each of the 30 categories. When applying CUPID, an average prediction score higher than 0.5 is required for a single category to be predicted. This threshold ensures that the predicted value for an assigned category is higher than that for the remaining categories in the ensemble, as they sum to one. When the average prediction value is <0.5 or the prediction is for the NCC category, it is reported as an unclassified prediction.

[0348] Targeted library preparation and sequencing

[0349] cfDNA and germline DNA from all CUP samples were processed according to the previously described TARGET laboratory protocol (Rothwell et al., 2019, TARGET patients up to TAR00286) or via the updated methods described below (patients and biobank samples from TAR00287 onwards). NEBNext® Ultra was used. TMDNA repair and dA tail addition were performed using the II-end repair / dA tail addition module (New England Biolabs, catalog number E7546L). Connector ligation and indexing were performed using the KAPA HyperPrep kit (Roche, catalog number 07962355001) and NEBNext® Multiplex Oligos for Illumina (New England Biolabs, catalog number E7335L). Whole-genome libraries of cfDNA and corresponding germline DNA were targeted by NGS using SureSelect custom DNA target enrichment probes (Agilent, catalog number 5190-4822). Target enrichment was performed using the SureSelect XT HS target enrichment system (Agilent, catalog number G9703A) and a 641-gene hybridization set, with 0.5 to 1.0 µg of each DNA library (paired cfDNA and gDNA libraries from the same patient were merged in each pull-down capture) for each target enrichment. The captured library was amplified using the KAPA HiFi HotStart PCR Kit (Roche, catalog number 07958897001) and quantified using the KAPA Library Quantification qPCR Kit (Roche, catalog number 07960140001). The library was sequenced at 2 × 150 bp using a NextSeq 500 or NovaSeq 6000 (Illumina).

[0350] Targeted library alignment and mutation determination (call)

[0351] FASTQ files were aligned to GRCh38 using bwa mem (Li, H 2013, v0.7.17) and deduplicated using samtools (Li et al., 2009, v1.9). Mutect was determined using Mutect2 (v4.2.5.0, following GATK best practices (Vander Auwera and O'Connor, 2020), with default parameters except for an f-score β of 5, a log 10 odds threshold of 1.0, and a minimum variant allele fraction of 1%) and QIAGEN CLC Genomics Workbench (v20.0.2build 200002) to determine variants between cfDNA samples and matched germline controls derived from whole blood. Mutations identified by both tools were designated as high confidence and annotated using VEP (v193.1, McLaren et al., 2016) and OncoKB (v3.17, Chakravarty et al., 2017). VCF files are converted to MAF files using vcf2maf (v1.6.21, Kandoth, C. mskcc / vcf2maf: vcf2maf v1.6.19. (2020)), and the Variant_Classification field is restricted to one of the following: frame shift missing (Frame_Shift_Del), frame shift insert (Frame_Shift_Ins), splice site mutation (Splice_Site), translation start site mutation (Translation_Start_Site), nonsense mutation (Nonsense_Mutation), nonstop codon mutation (Nonstop_Mutation), in-frame missing (In_Frame_Del), in-frame insert (In_Frame_Ins), or missense mutation (Missense_Mutation), and any mutations with a population frequency higher than 1% present in the gnomAD database are filtered out (Karczewski et al., 2020, v2.1.1). Use maftools (v2.8.05, Mayakonda et al., 2018) to generate an oncoplot and select those genes that are marked as oncogenic by OncoKB (v3.17).

[0352] Analysis of mutations that are carcinogenic and modifiable

[0353] Each high-confidence genomic alteration that passed the mutation filtering step was annotated using OncoKB to determine its oncogenic potential and potential actionability. Only alterations with level 3 or higher evidence of actionability were reported.

[0354] Annotation of mutant genes enriched by genomic alterations in a single tumor type

[0355] Based on AACR Project GENIE mutation data (Posner et al., 2022), genes were annotated according to tumor type enrichment using previously described genomes (which showed statistically significant enrichment of genomic alterations in a single tumor type compared to all other cancers). For patients with “clinically resolved” or “highly suspected” diagnoses, mutations were considered supportive if at least one gene was marked as statistically enriched in that tumor type. For patients with “differential diagnoses,” mutations were considered supportive if only one potential differential diagnosis was present; mutations enriched for more than one potential diagnosis were not considered supportive.

[0356] Example 1: Building a robust multi-class TOO classifier

[0357] To construct a robust multi-class TOO classifier applicable to cfDNA samples, we must address the significant challenge of the high variability in ctDNA content (tumor fraction, TF) within cfDNA, which leads to the dilution of tumor-specific signals by the dominant non-cancer cfDNA signal, even in metastatic cancer. To address this without performing profiling on thousands of cfDNA samples, we apply bioinformatics methods to simulate cfDNA samples with different TFs (Chemi et al. 2022). Figure 2 First, we used publicly available DNA methylation data from tumor tissues (using an Infinium 450K methylation microarray, primarily from The Cancer Genome Atlas (TCGA)), representing 29 tumor categories (9,017 tumors (Hoadley et al. 2018)). Figure 3 The methylation β value of each array probe was converted into an estimated T7-MBD-seq read count. For controls, DNA methylation data were determined from cfDNA in blood samples from 79 non-cancer control (NCC) subjects. We then created a computer-simulated mixture by mixing the estimated cancer category counts with previously sequenced cfDNA samples from patients. Figure 2 (Methods). A series of TFs (0.5% to 10%) and NCC clinical characteristics (age, sex, race, smoking status, comorbidities) were distributed for each tumor category. Preliminary classifier development showed that the presence of normal liver tissue components in NCC cfDNA may incorrectly lead to hepatocellular carcinoma category prediction, which was remedied by adding an array of non-cancerous liver tissues (n=49) to the non-cancer categories. A total of 276,108 mixtures were created across 30 categories.

[0358] Tumor-specific genomic regions were calculated using differentially methylated regions (DMRs), comparing each of 30 categories (100% tumor content), and selecting the 250 DMRs (22,179 unique regions) with the greatest differences between each pairwise comparison. Figure 4 Using these DMRs, dimensionality reduction was applied to 9,017 tumor samples using Uniform Manifold Approximation and Projection (UMAP), reproducing the class clustering observed when using all regions. Figure 5 An ensemble classifier named CUPID was constructed, consisting of 100 individual gradient boosting tree sub-classifiers, each trained on a computer-simulated mixture of 80% array and NCC. Figure 2 The subclassifiers were then tested on a “reserved” dataset containing the 20% array excluded from the training set and the NCC mixture set. These subclassifiers performed accurately across 30 classes, with a mean area under the multi-class receiver operating characteristic (AUROC) of 0.980 (standard deviation 0.00521). Figure 6 A). The resulting ensemble classifier, taking the average prediction score of each sub-classifier, has an AUROC of 0.984. Figure 6 (B) When a score for a single category is >0.5, it is considered a tumor prediction, ensuring that the score of the assigned category is higher than the remaining categories in all combinations. When all scores are <0.5 or the prediction is for a non-cancer category, it is reported as an unclassified prediction.

[0359] Example 2: Testing CUPID on an independent test cohort with known primary cancer types

[0360] We tested CUPiD on an independent testing cohort of 170 cfDNA samples, including 143 cancer patient samples and 27 NCCs from 13 different tumor types, and performed profiling analysis using T7-MBD-seq. Figure 7 (AB). In cancer cfDNA samples, CUPID correctly predicted the tumor type in 121 / 143 patients (overall sensitivity, 84.6%), made no classification prediction in 18 / 143 patients (12.6%), and made incorrect predictions in 4 / 143 patients (2.8%). Figure 8 Of the 27 NCC samples, none were predicted to be a tumor category. Figure 8Of the 125 samples with tumor predictions, correct predictions were made in 121 out of 125 cases (TOO accuracy of 96.8%). These results are comparable to other cfDNA TOO classifiers under development for early cancer detection; for example, TOO accuracy using DNA methylation in 1393 samples with “cancer-like” methylation signals across 21 different tumor categories has recently been reported to be 87.0% to 90.2% (Klein et al., 2021).

[0361] In this study, the tumor fraction (TF) for each sample was estimated using ichorCNA (Adalsteinsson et al., 2017) via shallow whole-genome sequencing, and the tumor fraction varied by tumor type, ranging from 0 to 60.3%. Figure 9 A). For 54 cancer cfDNA samples, the estimated TF was <3% of the ichorCNA detection limit. Of these cases, 37 were correctly predicted (67.9%), 16 were unclassified (30%), and 1 was incorrectly predicted (…). Figure 9 B). This indicates that although the estimated TF was low, methylation profiling detected tumor signals and accurately predicted TOO; only two samples with estimated TF > 3% were unclassified ( Figure 9 B).

[0362] Of the four cases where CUPID incorrectly predicted tumor types, three were lung adenocarcinoma patients, two-thirds of whom were predicted to have lung cancer but were actually squamous cell carcinoma, and one-third were breast cancer. The remaining incorrect prediction was bile duct carcinoma being predicted as pancreatic adenocarcinoma.

[0363] Example 3: cfDNA mutation profile analysis in CUP patients

[0364] Then, a profiling analysis of CUP patients was investigated in a pilot study of 41 patients. In particular, the feasibility of combining cfDNA methylation and mutation profiling analysis with TOO prediction was evaluated.

[0365] Prior to the spectrum analysis, most CUP cases had been classified as adenocarcinoma (25 / 41, 61.0%) or poorly differentiated carcinoma (11 / 41, 26.8%). At this stage, standard clinical investigation was unable to identify the primary carcinoma more specifically.

[0366] Unsurprisingly, validating TOO predictions is challenging given the inherent nature of CUP. Retrospectively, we reviewed clinical data, including clinical history, pathology, radiology, and discussions at CUP-specific multidisciplinary team (MDT) meetings, including additional investigations following the initial diagnosis. In 15 / 41 (36.6%) patients, a subsequent diagnosis of the primary tumor was made, and the patient received treatment for that tumor type (“clinical resolution”).

[0367] Of the 26 patients (n=26) who remained confirmed with CUP (cCUP) throughout their cancer journey, 18 / 26 had a suspected tumor diagnosis based on their clinical data; a highly suspected primary tumor diagnosis was made (n=11, "highly suspected"), or multiple suspected primary tumor diagnoses were made (n=7, "differential diagnosis"). The remaining 8 / 41 patients had no clinical or radiological indications for a potential primary tumor diagnosis ("no clinical suspicion").

[0368] Initially, the cfDNA mutation profiles of 41 CUP patients were assessed using a comprehensive 641-genome analysis (see Methods). cfDNA mutation profile analysis was unsuccessful in one patient. This patient exhibited numerous mutations common in population variant genomic databases, suggesting potential contamination. However, cfDNA mutation profile analysis was successful in the remaining 40 patients. Among these patients, we identified 345 mutations in 33 patients (82.5% with at least one mutation). The median number of mutations per patient was 5 (range 0–77), and the median variant allele frequency (VAF) was 10.4% (range 0–61.3%). Figure 10 The OncoKB (Chakravarty et al., 2017) analysis predicted 60 alterations in 26 patients as carcinogenic or probable carcinogenic. Figure 10 Of these, 7 out of 26 patients carried potentially interventional alterations with grade 3 or higher evidence. This included 3 patients with PIK3CA mutations, 2 patients with non-V600E BRAF mutations, and 1 patient with an IDH1 mutation, all of whom were targeted with FDA-approved drugs in their respective tumor types.

[0369] To determine whether any identified alterations in the CUP cohort could indicate tumor type, we considered previously described genetic alterations significantly enriched in specific tumor types (Posner et al. 2022) and inquired whether these tumor type enriched alterations (TTEAs) supported clinical characteristics. No TTEAs supporting only one tumor type were found in any case, demonstrating the limitations of mutation analysis. At least one TTEA was present and supportive in 14 / 40 cases (35.0%). Figure 11 In 14 (35.0%) cases, TTEA was present but not supportive, or it vaguely supported several tumor types in differential diagnosis cases. The remaining cases had no TTEA (n=12). Figure 11 ).

[0370] Example 4: Application of CUPID in identifying primary cancer in a CUP queue

[0371] cfDNA methylation profiling was successfully performed in our cohort of 41 CUP patients (41 / 41 passed QC). The estimated ichorCNA TF ranged from 0 to 53.4%, with 27 / 41 samples (65.8%) >3%, which correlated well with the mean VAF calculated by mutation analysis (r=0.74, p=<0.001). Figure 12 When CUPID was applied to this cohort, it produced tumor prediction in 32 / 41 cases (78.0%). Figure 13 (Table 3). For 9 / 41 patients (22.0%), no prediction was made, of which 7 / 9 unclassified cases showed no copy number change (estimated TF <3%), and 5 / 7 of these patients also had undetectable mutations, indicating low ctDNA content ( Figure 10 , Figure 14 Of the 32 tumor predictions, 26 patients (81%) were predicted to fall within five broad cancer categories. Figure 15 The most common predictions were hepatobiliary and pancreatic tumors (7 / 32, 20%, of which 6 were predicted to be cholangiocarcinoma) and female reproductive tract tumors (6 / 32, 18.75%). Figure 15 Although cholangiocarcinoma is rare, it is increasingly recognized in the CUP cohort, and diagnostic biomarkers are limited (Saha et al. 2016, Conway et al. 2022). Upper and lower gastrointestinal cancers (4 / 32, 12.5%), lung cancer (5 / 32, 15.2%), and urinary cancers (4 / 32, 12.5%) are also frequently predicted by CUPID. These tumor type predictions are consistent with historical data on primary tumor types found in autopsies of CUP patients (Le Chevalier 1988, Mayordomo et al. 1993) and are frequently predicted in other large TOO studies based on tumor tissue profile analysis (Fizaki et al. 2019, Hainsworth et al. 2013, Lu et al. 2021, Hayashi et al. 2019, and Moran et al. 2016). Interestingly, all tumor types predicted by CUP employed fundamentally different treatment strategies compared to SOC chemotherapy for CUP. With the exception of single tumors, all tumor types predicted warrant consideration of immunotherapy or targeted therapy as first- or second-line SOC treatment options.

[0372] Of the 33 patients with "clinical resolution" (n=15) or suspected diagnosis (n=18), 26 had CUPID tumor type predictions, and 23 of these predictions (88.5%) were consistent with either the identified primary tumor type or the suspected differential diagnosis. Figure 11 and 13Three cases showed discrepancies between CUPID predictions and clinical data, including two "clinically resolved" cases: the first was predicted to be bladder cancer, but was a rare yolk sac tumor not present in the CUPID classifier's training data; the second was predicted to be bile duct cancer, but the patient was ultimately diagnosed with gastric cancer; the last patient was predicted to have an upper GI malignancy, but still presented with clinical features suggestive of primary breast cancer (cCUP). Figure 13 CUPiD predictions were also made in 6 out of 8 patients with no suspected or confirmed primary tumors, demonstrating the potential of TOO molecular profiling analysis in the most uncertain cases.

[0373] Fifteen patients with “clinically resolved” primary tumors experienced prolonged diagnostic uncertainty, and most had received suboptimal empirical chemotherapy before the primary tumor was identified. Figure 16 The median time to diagnosis was 7.1 months (range 0.4 to 47.2 months), and invasive repeat biopsies were performed in 6 patients to arrive at a definitive diagnosis. With sufficient TF, a liquid biopsy-based TOO classifier, measured at suspected cancer diagnosis, could have significantly accelerated diagnosis in most patients and potentially avoided the need for repeat invasive biopsies. The experimental turnaround time for CUPiD is currently less than three weeks and is expected to be shortened through assay automation.

[0374] Example 5: CUPID application using fewer genomic regions

[0375] The decision was then made to study the effect of using the CUPID classifier and changing the DMR set used.

[0376] First, the variable importance of the genomic regions used in the original CUPID classifier was considered. Then, for different N values ​​of N, the top N of these genomic regions were selected, and the same ensemble classifier (using the same computer-simulated mixture and parameters) was fitted to each reduced DMR set. Variable importance was generated using the vip R package v0.4.1.

[0377] To determine performance, two metrics were considered. First, the area under the curve (AUC) on the retention metric was analyzed for the performance of a single 100 subclassifiers and the average ensemble (although on average this only spans 4 subclassifiers, not the 100 used on unseen data). This included retained samples from all 30 classes and therefore represents performance across the entire tumor type set. The results are shown in... Figure 17 A.

[0378] Then, the 170 cfDNA samples used in Example 2 were considered. These included 143 cancer samples from 13 categories and 27 non-cancer controls, thus excluding those fully validated for all categories, but all being actual patient cfDNA samples. Then... Figure 17 Model A was applied to 170 cfDNA samples in the test cohort. Figure 17 B shows the results, indicating how many genomic regions might be needed in this model for accurate predictions.

[0379] On both metrics, these classifiers achieved maximum performance with 5,000 windows compared to the larger total of 22,179 windows used in Example 4, indicating that removing non-informative windows allows the classifier to focus on the most important windows. It should be noted that these classifiers utilize window importance knowledge provided by the larger CUPID classifier; therefore, that larger classifier needs to be developed first.

[0380] Example 6: Different numbers of paired DMRs

[0381] Another way to change the number of windows entering the classifier is to take the first N windows of differential methylation in each pairwise comparison. For CUPID, the first 250 are used in each comparison, resulting in a total of 22,179 distinct windows, as many windows are in multiple comparisons.

[0382] The AUC for different numbers of windows was analyzed, and the results are shown in the figure. Figure 18 A. Figure 18 A shows that as the DMR quantity increases, the AUC steadily rises. Then... Figure 18 Model A was applied to 170 cfDNA samples in the test set. Figure 18 B shows the results.

[0383] The analysis showed that the ROC AUC continued to increase with the addition of additional DMRs, but for cfDNA samples, the performance using the first 25 paired DMRs was slightly better than using the first 250 paired DMRs. This is likely due to the distribution of the cfDNA sample cohort considered in this analysis.

[0384] Example 7: Constructing a classifier using a different genomic region than that used in CUPID

[0385] A classifier was then constructed without using the 22,179 windows used in CUPID. Pairwise DMRs were calculated in the same manner as before, and then the top N DMRs in each pairwise comparison were selected after filtering out the windows used in CUPID. Therefore, these are at least the 251st DMR in each pairwise comparison. The AUC of the classifier using the top 1, top 2, top 3, top 5, top 10, top 20, and top 25 pairwise DMRs (after filtering out the windows used in CUPID) is shown below. Figure 19 A, which shows that the AUC steadily increases with increasing DMR levels. These models were then applied to 170 cfDNA samples in the test cohort. Figure 19 B shows its results.

[0386] The performance of these classifiers was then compared with that of CUPID. Both the AUC and cfDNA scores of these classifiers were lower than the top N classifiers using the CUPID region. The AUC of the ensemble classifier is shown below. Figure 20 In A, the performance on the actual cfDNA cohort is shown in Figure 20 B in.

[0387] Example 8: CUPID v1.1 in an independent test team with improved quantity and diversity of known primary cancer types Performance on columns

[0388] A robust multi-class TOO classifier was constructed according to Example 1, except that pairwise comparisons did not include a non-cancer control class, except for the normal liver class. This resulted in a total of 30 classes: 29 primary malignant tumor classes and 1 normal liver class. The training data still included NCC data (excluding the primary malignant tumor data), where genomic regions were identified using pairwise comparisons between tumor classes. However, the partitioning of the NCC cfDNA sample set used between the training and test sets differed from v1. Some additional NCC samples were also used. After pairwise comparisons, genomic regions with a median β value > 0.25 in the non-cancer control data were excluded from the training data before selecting the top 250 upregulated and downregulated regions from each pairwise comparison. This classifier may also be referred to as CUPID v1.1.

[0389] Table 5 lists all the genomic regions used in CUPID v1.1. Table 6 lists 250 genomic regions of CUPID v1.1, which the inventors identified as the most informative genomic regions distinguishing between primary malignant tumor types, as determined by variable importance scores for pairwise comparisons between different primary malignant tumor categories. 174 of the 250 genomic regions in Table 6 also appeared in the first 250 genomic regions of earlier CUPID models (Table 2). These 174 specific genomic regions are shown in Table 7. Table 8 lists 500 genomic regions of CUPID v1.1, which the inventors identified as the most informative genomic regions distinguishing between primary malignant tumor types, as determined by variable importance scores for pairwise comparisons between different primary malignant tumor categories.

[0390] The performance of Cupido v1.1 was evaluated. Figure 21The percentages of samples correctly, incorrectly, or not predicted by CUPID v1.1 when applied to the following samples are shown: 143 cfDNA known cancer samples (“Known” cohort) and 41 cfDNA CUP samples (“CUP” cohort) from the same validation cohort used in Examples 2-4. An additional 6 skin melanoma samples and 18 breast cancer samples (hormone-positive or HER2-positive subtypes) were added to the original 143 “Known” samples, for a total of 167 cfDNA known cancer samples (“Combo” cohort).

[0391] For the CUP cohort, clinical relevance informs the likelihood of a correct prediction: a prediction is considered correct if a primary tumor diagnosis is subsequently made or if there is a high degree of clinical suspicion, or if the predicted tumor type is one of the potential differential diagnoses made clinically. No suspicion represents samples from the CUP cohort that do not have clinical suspicion of a primary tumor or potential differential diagnosis. Predictions were made for these samples, but it cannot be confirmed whether these predictions were correct or incorrect.

[0392] The known cohort had a sensitivity of 84% and a false negative tumor rate (FNTR) of 0.03. The combined cohort (n=167) had a sensitivity of 86% and an FNTR of 0.02. The CUP cohort (n=41) had a sensitivity of 78% and an FNTR of 0.02. Figure 21 ).

[0393] Example 9: Using a broader output classification to identify possible primary cancers

[0394] Next, it was decided to use a broader cancer category for the primary malignancy classification in the CUPID v1.1 output. The results are shown in... Figure 22 The figure shows the results of applying eight broad cancer categories as possible outputs to a validation cohort of 143 cfDNA samples with known cancers.

[0395] To achieve this, the model's training (using the original number of categories) remains unchanged, and the model still outputs a predicted score for each category. To obtain broader predicted scores, a wide range of cancer categories are predefined. The predicted scores for each subcategory are then summed to form a cumulative score within each broad category. If the cumulative score > a threshold of 0.5, it is considered a prediction for that particular broad category.

[0396] For example, the broad category for lung is the sum of the predicted scores from LUAD and LUSC. This might mean that a sample with a LUAD score of 0.35 and a LUSC score of 0.25 would normally result in an unclassified prediction (both below 0.5), but adding them together gives 0.6, thus predicting a broad lung category. Table 10 below details how each specific cancer category is broadly defined.

[0397]

[0398]

[0399] Table 10: Broad Cancer Grouping for 29 Tumor Categories

[0400] exist Figure 22 At the top of each of the graphs A through H are broad cancer groups. The prediction scores for each cancer type within each broad category are summed. The dashed lines represent a 0.5 prediction threshold. Only eight broad reporting categories are shown in the graphs because these are the categories from which we had test samples. If the cumulative cancer type prediction score of 0.5 or higher falls within the correct broad category, the sample is classified as correctly predicted.

[0401] Figure 22 I shows that, in the known cancer cohort, two samples that were incorrectly predicted when reported individually were correctly predicted using the broad reporting method. One LUAD sample (TAR00338) had an incorrect LUSC prediction score of 0.57 but was correctly predicted in the broader lung category, and one cholangiocarcinoma (CHOL) sample (TAR00248) had an incorrect pancreatic adenocarcinoma (PAAD) prediction score of 0.99 but was correctly predicted in the broader upper abdominal category. Furthermore, two samples that were not predicted using the individual reporting method were correctly predicted at the broad cancer category level: one sample was predicted as lung cancer (TAR00279), and the other as upper abdominal cancer (TAR00203).

[0402] Compared to outputs that could include any one or more of 29 primary malignancy classifications, this broader output classification improves the sensitivity of CUPiD, with each primary malignancy classification categorized according to different anatomical and / or histological tumor subtypes. Specifically, in known cohorts, the broader output classification increased CUPiD sensitivity from 84% to 87%. Figure 23 In the combined queue, a broader output classification improved the sensitivity of CUPiD from 86% to 89%, while in the CUP queue, it improved the sensitivity of CUPiD from 70% (23 / 33) to 76% (25 / 33). Figure 23 The two incorrect broad predictions in the CUP cohort were samples where the primary tumor type was not clinically confirmed but was narrowed down to several differential clinical diagnoses, and therefore the true diagnosis remained uncertain.

[0403] Example 10: Benchmarking of the CUPID v1.1 Model

[0404] To benchmark the CUPID v1.1 model, the AUC (which may also be referred to as AUROC) for different numbers of windows / DMRs for each pairwise comparison was analyzed. The total number of DMRs in the model, depending on the number of DMRs for each pairwise comparison, is shown in Table 11 below. Table 11 also shows the number of DMRs excluding duplicate regions. This means the total number of unique DMRs in the model, since some DMRs are among the top DMRs in multiple pairwise comparisons (or among the top 2, 5, etc.).

[0405]

[0406] Table 11: The total number of DMRs in the model, depending on the number of DMRs compared for each cancer type.

[0407] Figure 24 The integrated AUC of the retained computer simulation mixture set for each number of DMRs for each pairwise comparison is shown. The figure shows that the AUC steadily increases as the number of DMRs for each pairwise comparison increases.

[0408] Next, it was decided to evaluate the performance of the CUPID v1.1 model when considering different numbers of windows / DMR for each pairwise comparison, grouped by tumor content (TC). Tumor content was divided into groups below 3%, 3% to 5%, and above 5%. Figure 25 The integrated AUC of the computer-simulated mixture set with preserved DMR numbers for each pairwise comparison is shown. This indicates that increasing the DMR number for each pairwise comparison is most beneficial for the AUC of the mixture set with low tumor content. It also shows that a lower DMR number for each pairwise comparison is still effective, especially when the tumor content is at least 3%.

[0409] Next, we examined the prediction performance for each pairwise comparison of different DMR numbers. Figure 26 The graphs show the proportions of correct, incorrect, and no predictions for the retained computer simulation mixture set for each pairwise comparison with each number of DMRs. This graph excludes the NCC-only mixture set. These graphs show that AUC and CUPID performance remain relatively high with only 1 DMR per pairwise comparison, and begin to plateau after 50 DMRs per pairwise comparison.

[0410] Then, different numbers of windows / DMRs in CUPID v1.1 were tested on known cfDNA samples used in earlier embodiments. Figure 27In short, the known (n=143) predictions are from a cohort of 143 cfDNA samples with known primary cancer, and the CUP (n=41) predictions are from a cohort of 41 cfDNA CUP samples. The no-suspect ratio represents samples with no clinically suspected TOO. CUPid made predictions for these samples but could not classify them as correct or incorrect. When only one DMR was used per comparison, CUPid performed well in predicting within the known sample cohort. A plateau was formed after 50 DMRs per comparison. Increasing the number of DMRs resulted in incorrect predictions in the CUP sample cohort becoming no predictions, thus indicating improved accuracy.

[0411] Example 11: Comparison of the top 250 DMRs in CUPID v1 and CUPID v1.1

[0412] Based on the ranking, the top 250 DMRs for each model were compared (Table 2 for CUPID v1 and Table 6 for CUPID v1.1). Figure 28 A graph showing the variable importance ranking of all DMRs in CUPID v1 relative to all DMRs in CUPID v1.1 is presented. The graph shows that 174 of the top 250 DMRs in CUPID v1.1 also appear in the top 250 DMRs in CUPID v1. The graph also shows that while the highest-ranking DMR in CUPID v1.1 does not appear in CUPID v1, the next four highest-ranking DMRs in CUPID v1.1 (actually the next nine highest-ranking DMRs) also appear in CUPID v1.

[0413] discuss

[0414] In summary, we developed CUPiD, an accurate TOO liquid biopsy with encouraging sensitivity and accuracy across known tumor types and clinically consistent prediction in CUP patients with adequate TF in cfDNA. Because cfDNA mutation and methylation profiling can be performed from the same blood sample, this method is able to identify potential modifiable alterations to assist in treatment stratification while simultaneously predicting TOO. A pilot cohort of 41 CUP patients was applied to CUPiD, generating predicted tumor types in 32 out of 41 patients. All 32 patients have the potential to benefit from a tumor-specific treatment strategy fundamentally different from CUP SOC chemotherapy.

[0415] References

[0416] Numerous publications have been cited above to more fully describe and disclose the invention and the technical field to which it pertains. Full citations of these references are provided below. The entire contents of each of these references are incorporated herein by reference.

[0417] Amemiya HM, Kundaje A, Boyle AP. The ENCODE Blacklist: Identificationof Problematic Regions of the Genome. Sci Rep. 2019 Jun 27;9(1):9354. doi:10.1038 / s41598-019-45839-z. PMID: 31249361; PMCID: PMC6597582.

[0418] Adalsteinsson VA, Ha G, Freeman SS, Choudhury AD, Stover DG, ParsonsHA, et al. Scalable whole-exome sequencing of cell-free DNA reveals highconcordance with metastatic tumors. Nature Communications. 2017;8(1):1324.

[0419] Chakravarty D, Gao J, Phillips SM, Kundra R, Zhang H, Wang J, et al.OncoKB: A Precision Oncology Knowledge Base. JCO Precis Oncol. 2017;2017.

[0420] Chan-On W, Nairismägi ML, Ong CK, Lim WK, Dima S, Pairojkul C, LimKH, McPherson JR, Cutcutache I, Heng HL, Ooi L, Chung A, Chow P, Cheow PC,Lee SY, Choo SP, Tan IB, Duda D, Nastase A, Myint SS, Wong BH, Gan A, Rajasegaran V, Ng CC, Nagarajan S, Jusakul A, Zhang S, Vohra P, Yu W, Huang D, Sithithaworn P, Yongvanit P, Wongkham S, Khuntikeo N, Bhudhisawasdi V, Popescu I, Rozen SG, Tan P, Teh BT. Exome sequencing identifies distinctmutational patterns in liver fluke-related and non-infection-related bileduct cancers. Night Genet. 2013 Dec;45(12):1474-8. doi: 10.1038 / ng.2806. Epub2013 Nov 3. PMID:24185513.

[0421] Chemi, F., Pearce, S. P., Clipson, A., Hill, S. M., Conway, A.-M.,Richardson, S. A., Kamieniecka, K., Caeser, R., White, D. J., Mohan, S., Foy,V., Simpson, K. L., Galvin, M., Frese, K. K., Priest, L., Egger, J., Kerr,A., Massion, P. P., Poirier, J. T., Brady, G., Blackhall, F., Rothwell, D.G., Rudin, C. M. & Dive, C. 2022. cfDNA methylome profiling for detection andsubtyping of small cell lung cancers. Nature Cancer, 3, 1260-1270.

[0422] Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A Scalable TreeBoosting System. In Proceedings of the 22nd ACM SIGKDD InternationalConference on Knowledge Discovery and Data Mining (KDD '16). Association forComputing Machinery, New York, NY, USA, 785–794. https: / / doi.org / 10.1145 / 2939672.2939785

[0423] Conway A-M, Morris GC, Smith S, Manoharan P, Mitchell C, Backen A, etal. Intrahepatic cholangiocarcinoma hidden within cancer of unknown primaryBritish Journal of Cancer. 2022;in press.

[0424] Davis S, Meltzer PS. GEOquery: a bridge between the Gene ExpressionOmnibus (GEO) and BioConductor. Bioinformatics. 2007 Jul 15;23(14):1846-7.doi: 10.1093 / bioinformatics / btm254. Epub 2007 May 12. PMID: 17496320.

[0425] Edgar, R., Tan, P. P. C., Portales-Casamar, E. & Pavlidis, P. 2014.Meta-analysis of human methylomes reveals stably methylated sequencessurrounding CpG islands associated with high gene expression. Epigenetics &Chromatin, 7, 28.

[0426] Ewels, P.A., Peltzer, A., Fillinger, S. et al. The nf-core frameworkfor community-curated bioinformatics pipelines. Nat Biotechnol 38, 276–278(2020). https: / / doi.org / 10.1038 / s41587-020-0439-x

[0427] Fizazi K, Maillard A, Penel N, Baciarello G, Allouache D, Daugaard G,et al. A phase III trial of empiric chemotherapy with cisplatin andgemcitabine or systemic treatment tailored by molecular gene expressionanalysis in patients with carcinomas of an unknown primary (CUP) site(GEFCAPI 04). Annals of Oncology. 2019;30:v851-v.

[0428] Hainsworth JD, Rubin MS, Spigel DR, Boccia RV, Raby S, Quinn R, etal. Molecular gene expression profiling to predict the tissue of origin anddirect site-specific therapy in patients with carcinoma of unknown primarysite: A prospective trial of the Sarah cannon research institute. 2013. p.217-23.

[0429] Hand, D.J., Till, R.J. A Simple Generalisation of the Area Under theROC Curve for Multiple Class Classification Problems. Machine Learning 45,171–186 (2001). https: / / doi.org / 10.1023 / A:1010920819831

[0430] Hayashi H, Kurata T, Takiguchi Y, Arai M, Takeda K, Akiyoshi K, etal. Randomized phase II trial comparing site-specific treatment based on geneexpression profiling with carboplatin and paclitaxel for patients with cancerof unknown primary site. Journal of Clinical Oncology. 2019;37(7):570-9.

[0431] Hoadley KA, Yau C, Hinoue T, Wolf DM, Lazar AJ, Drill E, et al. Cell-of-Origin Patterns Dominate the Molecular Classification of 10,000 Tumorsfrom 33 Types of Cancer. Cell. 2018;173(2):291-304.e6.

[0432] Jusakul, A., et al. Whole-Genome and Epigenomic Landscapes ofEtiologically Distinct Subtypes of Cholangiocarcinoma. Cancer Discov 7, 1116-1135 (2017).

[0433] Karczewski, K.J., et al. The mutational constraint spectrumquantified from variation in 141,456 humans. Nature 581, 434-443 (2020).

[0434] Klein EA, Hubbell E, Maddala T, Aravanis A, Beausang JF, Filippova D,et al. Development of a comprehensive cell-free DNA (cfDNA) assay for earlydetection of multiple tumor types: The Circulating Cell-free Genome Atlas(CCGA) study. Journal of Clinical Oncology. 2018;36(15_suppl):12021-.

[0435] Klein, E.A., et al. Clinical validation of a targeted methylation-based multi-cancer early detection test using an independent validation set.Annals of Oncology 32, 1167-1177 (2021).

[0436] Kuhn, M. & Vaughan, D. yardstick: Tidy characterizations of modelperformance. R package version 1.0.0. (2022).

[0437] Le Chevalier T, Cvitkovic E, Caille P, Harvey J, Contesso G,Spielmann M, et al. Early metastatic cancer of unknown primary origin atpresentation. A clinical study of 302 consecutive autopsied patients.Archives of internal medicine. 1988;148(9):2035-9.

[0438] Lee, S., Lee, S., Ouellette, S., Park, W. Y., Lee, E. A. & Park, P.J. 2017. NGSCheckMate: software for validating sample identity in next-generation sequencing studies within and across data types. Nucleic AcidsRes, 45, e103.

[0439] Li, H. Aligning sequence reads, clone sequences and assembly contigswith BWA-MEM. arXiv preprint arXiv:1303.3997 (2013).

[0440] Li, H., et al. The Sequence Alignment / Map format and SAMtools.Bioinformatics (Oxford, England) 25, 2078-2079 (2009).

[0441] Lienhard, M., Grasse, S., Rolff, J., Frese, S., Schirmer, U., Becker,M., Börno, S., Timmermann, B., Chavez, L., Sültmann, H., Leschber, G.,Fichtner, I., Schweiger, M. R. & Herwig, R. 2017. QSEA-modelling of genome-wide DNA methylation from sequencing enrichment experiments. Nucleic AcidsResearch, 45.

[0442] Lienhard, M., Grimm, C., Morkel, M., Herwig, R. & Chavez, L. 2014.MEDIPS: Genome-wide differential coverage analysis of sequencing data derivedfrom DNA enrichment experiments. Bioinformatics, 30, 284-286.

[0443] Liu MC, Oxnard GR, Klein EA, Swanton C, Seiden MV, Cummings SR, etal. Sensitive and specific multi-cancer detection and localization usingmethylation signatures in cell-free DNA. Annals of Oncology. 2020.

[0444] Lu MY, Chen TY, Williamson DFK, Zhao M, Shady M, Lipkova J, et al.AI-based pathology predicts origins for cancers of unknown primary. Nature.2021;594(7861):106-10.

[0445] Martin, M. Cutadapt removes adapter sequences from high-throughputsequencing reads. EMBnet. journal 17, 10-12 (2011).

[0446] Mayakonda, A., Lin, D.-C., Assenov, Y., Plass, C. & Koeffler, H.P.Maftools: efficient and comprehensive analysis of somatic variants in cancer.Genome research 28, 1747-1756 (2018).

[0447] Mayordomo JI, Guerra JM, Guijarro C, Garcia-Prats MD, Gomez A, Lopez-Brea M, et al. Neoplasms of unknown primary site: A clinicopathological studyof autopsied patients. Tumori. 1993;79(5):321-4.

[0448] McInnes, L, Healy, J, UMAP: Uniform Manifold Approximation andProjection for Dimension Reduction, ArXiv e-prints 1802.03426, 2018

[0449] McLaren, W., et al. The ensembl variant effect predictor. Genomebiology 17, 1-14 (2016).

[0450] Melville, J. uwot: the Uniform Manifold Approximation and Projection(UMAP) method for dimensionality reduction. R package version 0.1.14. (CRAN,2022).

[0451] Moran S, Martínez-Cardús A, Sayols S, Musulén E, Balañá C, Estival-Gonzalez A, et al. Epigenetic profiling to classify cancer of unknownprimary: a multicentre, retrospective analysis. 2016. p. 1386-95.

[0452] Moss J, Magenheim J, Neiman D, Zemmour H, Loyfer N, Korach A, et al.Comprehensive human cell-type methylation atlas reveals origins ofcirculating cell-free DNA in health and disease. Nature Communications. 2018;9(1).

[0453] Nice 2010. Metastatic malignant disease of unknown primary origin inadults: diagnosis and management. In: EXCELLENCE, N. I. F. H. A. C. (ed.).https: / / www.nice.org.uk / guidance / cg104.

[0454] Peran I, Madhavan S, Byers SW, McCoy MD. Curation of the PancreaticDuctal Adenocarcinoma Subset of the Cancer Genome Atlas Is Essential forAccurate Conclusions about Survival-Related Molecular Mechanisms. Clin CancerRes. 2018 Aug 15;24(16):3813-3819. doi: 10.1158 / 1078-0432.CCR-18-0290. Epub2018 May 8. PMID: 29739787.

[0455] Posner, A., Prall, O. W., Sivakumaran, T., Etemadamoghadam, D., Thio,N., Pattison, A., Balachander, S., Fisher, K., Webb, S., Wood, C., Defazio,A., Wilcken, N., Gao, B., Karapetis, C. S., Singh, M., Collins, I. M.,Richardson, G., Steer, C., Warren, M., Karanth, N., Wright, G., Williams, S.,George, J., Hicks, R. J., Boussioutas, A., Gill, A. J., Solomon, B. J., Xu,H., Fellowes, A., Fox, S. B., Schofield, P., Bowtell, D., Mileshkin, L. &Tothill, R. W. 2022. A comparison of DNA sequencing and gene expressionprofiling to assist tissue of origin diagnosis in cancer of unknown primary.J Pathol.

[0456] Rothwell, D., Ayub, M., Cook, N., Thistlethwaite, F., Carter, L.,Dean, E., Smith, N., Villa, S., Dransfield, J., Clipson, A., White, D.,Nessa, K., Ferdous, S., Howell, M., Gupta, A., Kilerci, B., Mohan, S., Frese,K., Gulati, S., Miller, C., Jordan, A., Eaton, H., Hickson, N., O’brien, C.,Graham, D., Kelly, C., Aruketty, S., Metcalf, R., Chiramel, J., Tinsley, N.,Vickers, A., Kurup, R., Frost, H., Stevenson, J., Southam, S., Landers, D.,Wallace, A., Marais, R., Hughes, A., Brady, G., Dive, C. & Krebs, M. 2019.Utility of ctDNA to support patient selection for early phase clinicaltrials: the TARGET study.

[0457] Saha SK, Zhu AX, Fuchs CS, Brooks GA. Forty‐Year Trends inCholangiocarcinoma Incidence in the U.S.: Intrahepatic Disease on the Rise.The Oncologist. 2016;21(5):594-9.

[0458] Shah, S.P., et al. Integrating copy number polymorphisms into arrayCGH analysis using a robust HMM. Bioinformatics 22, e431-439 (2006).

[0459] Smith, T., Heger, A. & Sudbery, I. UMI-tools: modeling sequencingerrors in Unique Molecular Identifiers to improve quantification accuracy.Genome Res 27, 491-499 (2017).

[0460] Van der Auwera, G.A. & O'Connor, B.D. Genomics in the cloud: usingDocker, GATK, and WDL in Terra, (O'Reilly Media, 2020).

[0461] Wang P, Dong Q, Zhang C, Kuan PF, Liu Y, Jeck WR, Andersen JB, JiangW, Savich GL, Tan TX, Auman JT, Hoskins JM, Misher AD, Moser CD, YourstoneSM, Kim JW, Cibulskis K, Getz G, Hunt HV, Thorgeirsson SS, Roberts LR, Ye D,Guan KL, Xiong Y, Qin LX, Chiang DY. Mutations in isocitrate dehydrogenase 1and 2 occur frequently in intrahepatic cholangiocarcinomas and sharehypermethylation targets with glioblastomas. Oncogene. 2013 Jun 20;32(25):3091-100. doi: 10.1038 / onc.2012.315. Epub 2012 Jul 23. PMID: 22824796; PMCID:PMC3500578.

[0462] For standard molecular biology techniques, see Sambrook, J., Russell, DW Molecular Cloning, A Laboratory Manual. 3 ed. 2001, Cold Spring Harbor, New York: Cold Spring Harbor Laboratory Press.

[0463] Table 1: The top 250 genomic regions identified by the inventors from each of 30 pairwise comparisons. DMRs may be those disclosed at https: / / zenodo.org / doi / 10.5281 / zenodo.4071360 (as of November 10, 2023). The entire contents of https: / / zenodo.org / doi / 10.5281 / zenodo.4071360 are explicitly incorporated herein by reference as “Table 1”.

[0464]

[0465] Table 2: Top 250 genomic regions ranked by average variable importance in the model.

[0466]

[0467]

[0468]

[0469]

[0470] Table 3: List of categories used in CUPiD, and the number of methylated arrays for each category. All array samples were derived from Cancer Genome Atlas (TCGA) or Gene Expression Compilation (GEO), some of which are regroupings of CUPiD v1.

[0471]

[0472]

[0473] Table 4: CUPD prediction in 41 patients in the CUP cohort

[0474] Table 5: The top 250 genomic regions identified by the inventors from each of 30 pairwise comparisons in CUPID v1.1. Genomic regions may be those disclosed at https: / / doi.org / 10.5281 / zenodo.14054859 (as of November 8, 2024). The entire contents of https: / / doi.org / 10.5281 / zenodo.14054859 are explicitly incorporated herein by reference as “Table 5”.

[0475]

[0476] Table 6: Top 250 genomic regions ranked by average variable importance in the CUPID v1.1 model.

[0477]

[0478]

[0479]

[0480]

[0481]

[0482] Table 7: Genomic regions shared by Tables 2 and 6.

[0483] Table 8: Top 500 genomic regions ranked by average variable importance in the CUPID v1.1 model. Genomic regions can be those publicly available at https: / / doi.org / 10.5281 / zenodo.14055157 (as of November 8, 2024). The entire contents of https: / / doi.org / 10.5281 / zenodo.14055157 are explicitly incorporated herein by reference as “Table 8”.

[0484] Table 9: Top 500 genomic regions ranked by average variable importance in the CUPiD v1 model. Genomic regions can be those publicly available at https: / / doi.org / 10.5281 / zenodo.14055306 (as of November 8, 2024). The entire contents of https: / / doi.org / 10.5281 / zenodo.14055306 are explicitly incorporated herein by reference as “Table 9”.

Claims

1. A computer-implemented method for determining the primary carcinoma of a malignant tumor of unknown origin (MUO) in a subject, the method comprising: (i) Provide sequence data of cell-free DNA (cfDNA) obtained from a liquid sample from the object, the sequence data containing regional methylation features ("sample data"); (ii) The sample data is provided as input to a machine learning model containing a multi-class classifier trained on a training dataset, the training dataset containing regional methylation features ("training data") from at least two primary malignant tumor types, each primary malignant tumor type having a different location and / or histological subtype; (iii) Receiving output from the machine learning model, the output indicating the primary malignant tumor type classification or unknown type classification of the sample data; and (iv) Determine the primary cancer of the object based on the output received in step (iii).

2. The method of claim 1, wherein the liquid sample comprises a blood, urine, or plasma sample.

3. The method of claim 1 or claim 2, wherein the machine learning model comprises a decision tree, a logistic regression model, an artificial neural network, a support vector machine (SVM), a Naive Bayes algorithm, or a k-nearest neighbor algorithm, and optionally wherein the decision tree comprises a gradient boosting algorithm or a random forest algorithm.

4. The method of any one of claims 1 to 3, wherein the machine learning model comprises at least 60 individual classifiers, optionally at least 100 individual classifiers.

5. The method of any one of the preceding claims, wherein the at least two primary malignant tumor types comprise at least three primary malignant tumor types, optionally at least four primary malignant tumor types, wherein each primary malignant tumor type has a different location and / or histological subtype.

6. The method of any one of the preceding claims, wherein the at least two types of primary malignant tumors are selected from: a) Primary malignant hepatobiliary tumors; b) Primary malignant female reproductive tract tumors; c) Primary malignant upper or lower gastrointestinal tumors; d) Primary malignant lung tumor; e) Primary malignant breast tumor; and f) Primary malignant urological tumors.

7. The method of any one of the preceding claims, wherein the at least two primary malignant tumor types comprise at least 20 primary malignant tumor types, each primary malignant tumor type having a different location and / or histological subtype.

8. The method of any one of the preceding claims, wherein the at least two types of primary malignant tumors are selected from: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), invasive breast carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynecological carcinoma (Gynae), chromophobe renal cell carcinoma (KICH), clear cell renal carcinoma (KIRC), papillary renal cell carcinoma (KIRP), acute myeloid leukemia (LAML), and low-grade brain cancer. Glioma (LGG), hepatocellular carcinoma (LIHC), colorectal adenocarcinoma (lower GI), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), mesothelioma (MESO), pancreatic adenocarcinoma (PAAD), pheochromocytoma and paraganglioma (PCPG), prostate adenocarcinoma (PRAD), sarcoma (SARC), cutaneous melanoma (SKCM), testicular germ cell tumor (TGCT), thyroid cancer (THCA), thymoma (THYM), gastric and esophageal adenocarcinoma (upper GI), head and neck and esophageal squamous cell carcinoma (upper Sq), and uveal melanoma (UVM).

9. The method of any one of the preceding claims, wherein the training data further comprises regional methylation features from at least one non-cancer control.

10. The method of any one of the preceding claims, wherein the output comprises a probability score for each category.

11. The method of any one of the preceding claims, wherein each regional methylation feature comprises the DNA methylation state of at least one genomic region, optionally, the DNA methylation state of multiple genomic regions.

12. The method of claim 11, wherein each genomic region comprises a length of about 200 to about 2000 base pairs.

13. The method of claim 11 or claim 12, wherein the regional methylation features from each primary malignant tumor type comprise the DNA methylation status of at least 5, at least 10, at least 20, at least 30, at least 50, or at least 100 genomic regions.

14. The method of any one of claims 11 to 13, wherein the genomic region is selected from the genomic regions listed in Table 1.

15. The method of any one of claims 11 to 13, wherein the genomic region is selected from the genomic regions listed in Table 5.

16. The method of any one of claims 11 to 14, wherein the genomic region is selected from the genomic regions listed in Table 2.

17. The method of any one of claims 11 to 14 and 16, wherein the genomic region comprises at least the first five genomic regions listed in Table 2.

18. The method of any one of claims 11 to 13 and 15, wherein the genomic region is selected from the list in Table 6.

19. The method of any one of claims 11 to 13, 15 and 18, wherein the genomic region comprises at least the first five genomic regions listed in Table 6.

20. The method of any one of claims 11 to 16 and 18, wherein the genomic region is selected from the genomic regions listed in Table 7.

21. The method of any one of claims 11 to 20, wherein the genomic region comprises a genomic region identified from pairwise comparisons of DNA methylation states of genomic regions between the at least two primary malignant tumor types, optionally, wherein the genomic region comprises at least one, at least five, at least ten, at least twenty, at least thirty, at least fifty, at least one hundred, at least one hundred, at least two hundred, at least two hundred, or at least two hundred genomic regions with the greatest differences in methylation states identified from pairwise comparisons.

22. The method of any one of claims 11 to 21, wherein the genomic region is selected based on a function of similarity between each type of primary malignant tumor.

23. The method of any of the preceding claims, wherein the training data comprises regional methylation features derived from primary malignant tumor tissue samples obtained from a plurality of individuals, each of whom has a known primary malignant cancer.

24. The method of any one of the preceding claims, wherein the regional methylation features of each primary malignant tumor type in the training data comprise the DNA methylation status of at least one genomic region derived from a primary malignant tumor tissue sample and the DNA methylation status of at least one genomic region derived from cfDNA of at least one non-cancer control.

25. The method of any of the preceding claims, wherein the MUO comprises cancer of unknown origin (CUP).

26. The method of any of the preceding claims, wherein the sample data further comprises a tumor fraction (TF) of cell-free DNA (cfDNA) obtained from the liquid sample.

27. The method of claim 26, wherein the tumor fraction (TF) is at least about 3%.

28. A method for identifying the primary cancer in a subject with a malignant tumor of unknown origin (MUO), the method comprising: i) Analyze cell-free DNA (cfDNA) obtained from liquid samples from said object to obtain sequence data ("sample data") containing regional methylation features; and ii) Perform the method of any one of claims 1 to 27 on the sample data to determine the primary cancer of the object.

29. A method for selecting cancer treatment for a subject with a malignant tumor of unknown origin (MUO), the method comprising: i) Performing the method of any one of claims 1 to 27 to determine the most probable primary cancer of the object; as well as ii) Select an anticancer therapy that targets the most likely primary cancer.

30. A method for treating a subject with a malignant tumor of unknown origin (MUO), the method comprising: i) Performing the method of any one of claims 1 to 27 to determine the most probable primary cancer of the object; as well as ii) Administer anticancer therapy targeting the most likely primary cancer.

31. A system, which includes: Processor; and A computer-readable medium comprising instructions that, when executed by the processor, cause the processor to perform the steps of the method according to any one of claims 1 to 27.

32. A computer-readable medium comprising instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the method according to any one of claims 1 to 27.

Images