Methods for determining cancers

Abstract

The present invention relates to a computer implemented method for determining the primary cancer of a malignancy of unknown origin (MUO) of 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 a region methylation signature ("sample data"); ii) providing the sample data as input to a machine learning model comprising a multi-class classifier trained on a training data set comprising region methylation signatures from at least two primary malignant tumour types ("training data"), each primary malignant tumour type having a distinct location and / or histological subtype; iii) receiving an output from the machine learning model, the output indicative of a primary malignant tumour type classification or an unknown type classification for 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 methods of selecting a cancer treatment for a subject having a MUO and methods of treating a subject having MUO. Systems and computer readable media are also provided.

Description

[0001] Methods for determining cancers

[0002] Field of the Invention

[0003] The present invention relates to computer implemented methods for determining the primary cancer of a malignancy of unknown origin (MUO) of a subject. The present invention also relates to methods of selecting a cancer treatment for a subject having a MUO and methods of treating a subject having MUO. Systems and computer readable media are also provided.

[0004] Background

[0005] Malignancy of unknown origin (MUO) is a clinical condition representing a heterogenous group of malignancies wherein likely cancerous cells are identified in a subject, but the primary source of the cancerous cells is unknown. This condition also encompasses cancers of unknown primary (CUP), which represent a heterogenous group of cancers where a secondary cancerous tumour has been identified, but the primary cancer remains unknown after standard clinical investigations. Malignancies of unknown origin and cancers of unknown primary represent about 1 to 3% of all cancer diagnoses.

[0006] In the absence of the identity of the primary source of the cancerous cells, therapeutic options for MUO and CUP remain limited. Many subjects with CUP receive “one size fits all” standard chemotherapy, despite a wealth of clinical, pathological and molecular heterogeneity. Since the therapy isn’t targeted to the unknown primary source, therapy often has limited, at best, efficacy.

[0007] Current guidelines identify around 20% of CUP subjects as having “favourable CUP”. This is based on clinicopathological features corresponding to more treatment-responsive cancer types. Even for subjects classified as having favourable CUP, arriving at a primary tumour diagnosis can take a considerable length of time, involving multiple investigations and pathology reviews, from a very limited quantity of tissue from an invasive tumour biopsy. The remaining 80% of CUP subjects are categorised as having “unfavourable CUP”, which is considered to have a more negative prognosis. Accordingly, CUP remains a common cause of cancer-related death. Indeed, in England, only about 15% of subjects with CUP survive for one year or more, and only about 10% of subjects with CUP survive for three years or more.

[0008] The role of biomarker-driven precision oncology and advent of immunotherapies is rapidly changing standard-of-care (SOC) treatment and improving overall survival across numerous tumour types. However, these approaches remain out-of-reach for subjects with CUP, especially subjects with unfavourable CUP. Apart from a handful of tumour agnostic treatments, most targeted therapies demonstrate tumour-type dependent efficacy, exemplified by activity of targeted inhibitors in B-RAF mutant melanoma versus inactivity in colorectal cancers. Additionally, immunotherapy is increasingly indicated by biomarker presence validated by tumour type.

[0009] Molecular characterisation approaches to predict the tissue of origin (which may otherwise be referred to as tumour of origin) for CUP subjects are considered a gateway to better treatment stratification.

[0010] However, the scarcity of tumour tissue in subjects with CUP means that detection of the tumour of origin remains challenging. Furthermore, current tissue of origin predictions do not necessarily identify a particular primary cancer, given that cancers of the same tissue of origin can differ due to clinical, pathological and molecular heterogeneity. As such, there remains a need for improved diagnosis of MUO, particularly of CUP. This, in turn, should provide clinicians with an increased number of therapeutic options with improved therapeutic efficacy.

[0011] The present invention has been devised in light of the above considerations.

[0012] Summary of the Invention

[0013] Broadly, the present inventors have developed an improved method for determining the primary cancer of a malignancy of unknown origin (MUO) of a subject. Since access to adequate tumour tissue for molecular profiling in MUO and CUP remains challenging, the present invention utilises a liquid biopsy approach to predict the primary cancer of the subject. This provides an increased sensitivity of detection, due to an increased availability and quality of tumour tissue and associated DNA for analysis from the liquid biopsy sample. The computer-implemented analysis of region methylation signature(s) in the liquid biopsy then advantageously enables the accurate stratification of subjects in a swift turnaround time. This, in turn, enables the selection of treatment options targeted to the primary cancer for the subject, enabling improved efficacy of treatment. Furthermore, the swift identification of the primary cancer of the subject allows the clinician to administer treatment earlier than would otherwise be possible, further increasing the efficacy of treatment. In certain embodiments the method may be termed “CUPiD” (Cancer of Unknown Primary identification).

[0014] Accordingly, in a first aspect, the present invention provides a computer-implemented method for determining the primary cancer of a malignancy of unknown origin (MUO) of a subject, the method comprising: i) providing sequence data of cell-free DNA (cfDNA) obtained from a liquid sample from the subject, the sequence data comprising a region methylation signature (“sample data”); ii) providing the sample data as input to a machine learning model comprising a multi-class classifier trained on a training data set comprising region methylation signatures from at least two primary malignant tumour types (“training data”), each primary malignant tumour type having a distinct location and / or histological subtype; iii) receiving an output from the machine learning model, the output indicative of a primary malignant tumour type classification or an unknown type classification for the sample data; and iv) determining the primary cancer of the subject based on the output received in step (iii).

[0015] In the context of the present invention, the term “region methylation signature” will be understood to refer to the DNA methylation status of at least one region of the genome. The region of the genome will be understood to refer to a portion of the genome, which may otherwise be referred to as genomic region. Generally, “DNA methylation status” will be understood to refer to the level of DNA methylation. DNA methylation may be DNA hypermethylation or DNA hypomethylation.

[0016] In some embodiments, the liquid sample comprises a blood, urine or plasma sample.

[0017] In some embodiments, the machine learning model comprises a decision tree, a logistic regression model, an artificial neural network, a support vector machine (SVM), naive bayes or k-nearest neighbour algorithm, optionally wherein the decision tree comprises a gradient boosting algorithm or a random forest algorithm.

[0018] The machine learning model may comprise at least 60 individual classifiers, optionally at least 100 individual classifiers.

[0019] In some embodiments, the at least two primary malignant tumour types comprise at least three primary malignant tumour types, optionally at least four primary malignant tumour types, wherein each primary malignant tumour type has a distinct location and / or histological subtype.

[0020] In some embodiments, the at least two primary malignant tumour types are selected from: a) a primary malignant hepato-pancreatobiliary tumour; b) a primary malignant female genital tract tumour; c) a primary malignant upper or lower gastrointestinal tumour; d) a primary malignant lung tumour; e) a primary malignant breast tumour; and f) a primary malignant urological tumour.

[0021] At least one of the at least two primary malignant tumour types may comprise a primary malignant hepato-pancreatobiliary tumour, optionally a primary malignant cholangiocarcinoma tumour.

[0022] In some embodiments, the at least two primary malignant tumour types comprise at least 20 primary malignant tumour types, each primary malignant tumour type having a distinct location and / or histological subtype.

[0023] In some embodiments, the at least two primary malignant tumour types are selected from: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynaelogical carcinoma (Gynae), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), acute myeloid leukaemia (LAML), brain lower grade glioma (LGG), liver hepatocellular carcinoma (LIHC), colon and rectum adenocarcinoma (Lower Gl), 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 tumours (TGCT), thyroid carcinoma (THCA), thymoma (THYM), stomach and oesophageal adenocarcinoma (Upper Gl), head and neck and oesophageal squamous cell carcinoma (UpperSq) and uveal melanoma (UVM).

[0024] In some embodiments, the training data further comprises a region methylation signature from at least one non-cancer control. It will be appreciated that a region methylation signature from a non-cancer control comprises the DNA methylation status of at least one region of the genome obtained from a sample from a non-cancer control.

[0025] In some embodiments, the output comprises a probabilistic score for each classification.

[0026] In some embodiments, each region methylation signature comprises the DNA methylation status of a plurality of genomic regions. Each genomic region may comprise a length of from about 200 to about 2000 base pairs.

[0027] It will be appreciated that the training data comprises a region methylation signature from a primary malignant tumour type and a region methylation signature from at least one other primary malignant tumour type. Typically, the region methylation signature from each primary malignant tumour type comprises the same region(s) of the genome. By same region, this will be understood to refer to a portion of the genome having the same genomic coordinates and length. It, will, however, be appreciated, that the DNA methylation status of each region may differ between each primary malignant tumour type. Therefore, the region methylation signature of each primary malignant tumour type is a distinctive “signature” for each primary malignant tumour type.

[0028] In some embodiments, the region methylation signature from each primary malignant tumour type comprises 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 (which may otherwise be referred to as DMRs) listed in Table 1 . In some embodiments, the genomic regions are selected from the genomic regions listed in Table 2. In some embodiments, the genomic regions comprise at least the top five genomic regions listed in Table 2. In some embodiments, the genomic regions comprise all genomic regions listed in Table 2.

[0029] 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 comprise at least the top five genomic regions listed in Table 6. In some embodiments, the genomic regions are selected from the genomic regions listed in Table 7.

[0030] In some embodiments, the genomic region(s) in each region methylation signature in the training data comprise genomic regions identified from a pairwise comparison of DNA methylation status of genomic regions between the at least two primary malignant tumour types.

[0031] The genomic regions in the training data may comprise at least 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 which have the greatest difference in methylation status as identified from a pairwise comparison. In some embodiments, the genomic regions are selected based on a function of similarity between each primary malignant tumour type.

[0032] The training data may comprise region methylation signatures derived from primary malignant tumour tissue samples obtained from a plurality of individuals each having a known primary malignant cancer.

[0033] In some embodiments, the region methylation signature for each primary malignant tumour type in the training data comprises a DNA methylation status of at least one genomic region derived from a primary malignant tumour tissue sample and a DNA methylation status of at least one genomic region derived from cfDNA from at least one non-cancer control.

[0034] In some embodiments, the sample data further comprises a tumour fraction (TF) for the cell-free DNA (cfDNA) obtained from the liquid sample. In some embodiments, the tumour fraction (TF) is at least about 3%.

[0035] In some embodiments, the MUO comprises a cancer of unknown primary cancer (CUP). The subject may have unfavourable CUP. In some embodiments, the subject is known to have secondary carcinoma, but the primary cancer is unknown. In some embodiments, the subject is known to have secondary adenocarcinoma, secondary squamous carcinoma or secondary poorly differentiated carcinoma, but the primary cancer is unknown.

[0036] In some embodiments, the primary cancer in step iv) is selected from: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynaelogical carcinoma (Gynae), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), acute myeloid leukaemia (LAML), brain lower grade glioma (LGG), liver hepatocellular carcinoma (LIHC), colon and rectum adenocarcinoma (Lower Gl), 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 tumours (TGCT), thyroid carcinoma (THCA), thymoma (THYM), stomach and oesophageal adenocarcinoma (Upper Gl), head and neck and oesophageal squamous cell carcinoma (UpperSq), uveal melanoma (UVM) and unclassified primary cancer.

[0037] According to a second aspect, the present invention provides a method of determining the primary cancer of a subject having a malignancy of unknown origin (MUO), the method comprising: i) analysing cell-free DNA (cfDNA) obtained from a liquid sample from the subject in order to obtain sequence data comprising a region methylation signature (“sample data”); and ii) performing the method of the first aspect on the sample data to determine the primary cancer of the subject.

[0038] According to a third aspect, the present invention provides a method of selecting a cancer treatment for a subject having a malignancy of unknown origin (MUO), the method comprising: i) performing the method of the first aspect to determine the likely primary cancer of the subject; and ii) selecting an anti-cancer treatment that targets the likely primary cancer.

[0039] According to a further aspect, the present invention provides a method of treating a subject having malignancy of unknown origin (MUO), the method comprising: i) performing the method of the first aspect to determine the likely primary cancer of the subject; and ii) administering an anti-cancer treatment that targets the likely primary cancer.

[0040] According to another aspect, the present invention provides a system comprising: i) a processor; and ii) a computer readable medium comprising instructions that, when executed by the processor, cause the processor to perform the steps of the method of the first aspect.

[0041] Also provided by the present invention is one or more computer readable media comprising instructions that, when executed by one or more processors, cause the one or more processors to performs the steps of the method of the first aspect.

[0042] The invention includes the combination of the aspects and preferred features described except where such a combination is clearly impermissible or expressly avoided.

[0043] Summary of the Figures

[0044] Embodiments and experiments illustrating the principles of the invention will now be discussed with reference to the accompanying figures in which:

[0045] Figure 1 : Flow diagram of CUP cohort diagnosis classification after retrospective review of clinical data.

[0046] Figure 2: Schematic of CUPID classifier development.

[0047] Figure 3: Diagram to show number of arrays used in each cancer class.

[0048] Figure 4: Example volcano plot showing the difference in beta values and false-discovery rate adjusted p-values for the 59,918 differentially methylated regions found between the 79 ACC and the 409 BLCA converted arrays. The highlighted 250 windows with the largest magnitude of difference in beta values between each class were selected to build the classifiers.

[0049] Figure 5: Two-dimensional UMAP using 22,179 differentially methylated regions (DMRs) selected by tumour class by class comparison, applied to 9,017 converted methylation arrays. Class labels are superimposed over the centroid of members of that class.

[0050] Figure 6: A) Area Under the Receiver Operator Curve (AUROC or AUC) values for the 100 individual classifiers, evaluated on the held-out mixture samples not used to train the classifier (10,611-11 ,508 held- out mixtures per classifier). Cancer class abbreviations are defined in Table 3. B) Receiver Operator Curves (ROCs) for each class, evaluated using the ensemble classifier applied to held-out mixture samples for each class.

[0051] Figure 7: A) Number of T7-MBD-Seq samples of each class in the independent cfDNA test cohort. B) Relative enrichment score (relH) of methylation enriched samples (MeCap) vs. Non-enriched sample across the 170 cfDNA samples in test cohort.

[0052] Figure 8: CUPiD performance in test cohort of 170 cfDNA samples, consisting of 143 from known primary tumour types and 27 non-cancer controls (‘denotes CUPiD prediction)

[0053] Figure 9: A) Tumour fraction as estimated by ichorCNA using non-enriched sample shallow whole genome sequencing, on the 170 cfDNA samples. 3% limit of detection cut-off shown. B) CUPiD prediction concordance in relation to tumour fraction.

[0054] Figure 10: Mutational profiling of a 641 gene targeted panel of cfDNA compared with matched germlines for 40 patients with CUP. Oncoplot showing mutations in genes marked as Oncogenic by oncoKB, with actionable mutations highlighted by inset stars. Top panel shows average variant allele frequency (VAF) per patient, across all alterations found. Bottom panel annotated by subsequent primary diagnoses and the CUPiD prediction. Cancer class abbreviations are defined in Table 3.

[0055] Figure 11 : Alluvial plot showing how the tumour type-enriched alterations (TTEAs) (left) and CUPiD predictions (right) correspond to clinical classifications (centre).

[0056] Figure 12: Pearsons correlation of ichorCNA estimated tumour fraction determined from non-enriched sample against the average variant allele frequency (VAF) of mutations found by targeted 641 gene panel sequencing, from cfDNA of the 41 patients with CUP.

[0057] Figure 13: Sankey plot of CUPiD predictions and correlation with subsequent primary tumour diagnosis or clinical suspicion.

[0058] Figure 14: Estimated tumour fraction (from ichorCNA) of 41 cfDNA samples from patients with CUP, grouped by CUPiD prediction status and coloured by predicted class. Dotted line denotes 3% tumour fraction. Class abbreviations are defined in Table 3.

[0059] Figure 15: Tumour type predictions made by CUPiD in CUP cohort (n=32), excluding where no prediction was made.

[0060] Figure 16: Swimmers plot of the fifteen patients with CUP with clinically resolved primary tumour diagnosis. Timeline of diagnostic investigations from point of CUP diagnosis to death or data lock. Time of final primary tumour diagnosis colour coded. Annotated by final diagnosis, CUPiD prediction and concordance.

[0061] Figure 17: A) AUC values for 100 individual sub classifiers, each built using the N most important DMRs in the full CUPiD classifier, for different values of N. Each sub classifier is evaluated on the held-out mixture samples not used to train that sub classifier. Black dots show ensemble classifier AUC. B) Application of these classifiers to the 170 cfDNA samples comprising the test cohort. Figure 18: A) AUC values for 100 individual sub classifiers, each built using the N most significant DMRs between each comparison, for different values of N. Each sub classifier is evaluated on the held-out mixture samples not used to train that sub classifier. Black dots show ensemble classifier AUC. B) Application of these classifiers to the 170 cfDNA samples comprising the test cohort.

[0062] Figure 19: A) AUC values for 100 individual sub classifiers, each built using the N most significant DMRs between each comparison, for different values of N, after filtering out the 22,179 windows present in the CUPiD classifier. Each sub classifier is evaluated on the held-out mixture samples not used to train that sub classifier. Black dots show ensemble classifier AUC. B) Application of these classifiers to the 170 cfDNA samples comprising the test cohort.

[0063] Figure 20: A) Ensemble classifier AUC performance comparison between the non-CUPiD classifier and the CUPiD classifier, wherein the non-CUPiD classifier used only windows not used in the CUPiD classifier. B) Performance comparison between non-CUPiD classifier and CUPiD classifier on the 170 sample cfDNA test cohort.

[0064] Figure 21 : CUPiD v1.1 performance in test cohort of known cancer samples (n=143, “known”), 167 cfDNA samples (“Combo”), consisting of 143 known samples from “known” cohort with additional 24 known samples from skin cutaneous melanoma and breast cancer (labelled as “Combo”), and 41 from CUP subjects (“CUP”). Known TARGET cohort sensitivity and false negative tumour rate (FNTR) = 0.84, 0.03 respectively. Combo cohort (n=167) sensitivity and FNTR = 0.86, 0.02 respectively. CUP cohort (n=41) sensitivity and FNTR = 0.78, 0.02 respectively.

[0065] Figure 22: Broader output classifications to determine likely primary cancer. Broad cancer output groups were applied to the CUPiD v1 .1 model. Each bar represents a separate subject. Y axis represents prediction score. Samples were classed as correct predictions if the accumulated cancer type prediction score of 0.5 or above was within the correct broad category. Dashed line shows the classification threshold of 0.5. The broad cancer output groups shown are (A) breast cancer: left panel shows samples correctly called in the broad class of breast cancer. Right panel shows a sample incorrectly called in the broad class of breast cancer, where the correct tumour type was lung cancer, (B) female reproductive cancer; panel shows samples correctly called in this broad class, (C) kidney cancer; panel shows samples correctly called in the broad class of kidney cancer, (D) lower abdominal cancer; panel shows samples correctly called in this class, (E) lung cancer; panel shows samples correctly called in this class, (F) samples that remain unclassified using the broad classification approach. The sum of all prediction scores within a broad class remained at less than 0.5, (G) rare cancer; panel shows samples correctly called in this class and (H) upper abdominal cancer; left panel shows samples correctly called in broad class, right panel shows a sample incorrectly called in this broad class, where the correct tumour type was lung cancer. A key showing the more specific classifications is to the right of the panels in (A) to (H). (I) shows bar graphs showing the prediction scores for the four samples that change prediction status when the broad classification approach is applied to the known cancer cohort. Top graph shows broad class of lung and bottom graph shows broad class of upper abdominal. Cancer class abbreviations are defined in Table 3. Figure 23: Broad output classifications increase the sensitivity of CUPiD v1 .1 .

[0066] Figure 24: AUC values for CUPID v1 .1 ensemble classifiers, each built using the N most significant DMRs between each pairwise comparison, for different values of N, as detailed in X axis. Evaluated on the held-out mixture samples.

[0067] Figure 25: AUC values for CUPID v1 .1 ensemble classifiers, each built using the N most significant DMRs between each pairwise comparison, for different values of N, as detailed in X axis. Evaluated on the held-out mixture samples separated by tumour content (TC) of below 3% (left graph), between 3% and 5% (middle graph) and 5% and above (right graph).

[0068] Figure 26: The proportions of correct, incorrect and no predictions for the held-out in silico mixture sets of Figure 24 for each number of DMRs per pairwise comparison.

[0069] Figure 27: Proportion of samples predicted as correct, incorrect or no prediction for each number of most significant DMRs per pairwise comparison in CUPiD v.1 .1 , when tested on the cohort of 143 cfDNA samples from TARGET and CUP (n=41).

[0070] Figure 28: Plot of DMR variable importance rankings in CUPiD v1 vs v1 .1 using all DMRs.

[0071] Detailed Description of the Invention

[0072] According to a first aspect, the present invention provides a computer-implemented method for determining the primary cancer of a malignancy of unknown origin (MUO) of a subject, the method comprising: i) providing sequence data of cell-free DNA (cfDNA) obtained from a liquid sample from the subject, the sequence data comprising a region methylation signature (“sample data”); ii) providing the sample data as input to a machine learning model comprising a multi-class classifier trained on a training data set comprising region methylation signatures from at least two primary malignant tumour types (“training data”), each primary malignant tumour type having a distinct location and / or histological subtype; iii) receiving an output from the machine learning model, the output indicative of a primary malignant tumour type classification or an unknown type classification for the sample data; and iv) determining the primary cancer of the subject based on the output received in step (iii).

[0073] As used herein, the term “primary cancer” refers to an original, or first, malignant tumour in the subject.

[0074] As used herein “subject” refers to an animal or human. The subject may be mammalian (e.g. a human, a non-human primate, a cat, dog, horse, donkey, sheep, pig, goat, cow, mouse, rat, rabbit or guinea pig). The subject is preferably a human. The subject may be a human adult (at least 18 years old). In some embodiments, the subject is a human child (less than 18 years old). By “malignancy of unknown origin (MUO)”, this will be understood to refer to a condition wherein likely cancerous cells and / or symptoms are identified in a subject, but the primary source of the cancerous cells is unknown. In some embodiments, the MUO comprises a cancer of unknown primary cancer (CUP). A cancer of unknown primary cancer (CUP) will be understood to refer to a cancerous tumour having been identified in the subject, but the primary cancer remains unknown after standard clinical investigations.

[0075] The machine learning model of the present invention comprises a multi-class classifier trained on a training data set comprising region methylation signatures from at least two primary malignant tumour types, each primary malignant tumour type having a distinct location and / or histological subtype. In some embodiments, each primary malignant tumour type has a distinct location and histological subtype. The distinct location and / or histological subtype of each primary malignant tumour type ensures that each primary malignant tumour type on which the classifier is trained is clearly defined and clearly distinct from any other primary malignant tumour types from which the classifier is trained. Advantageously, this ensures that the classifier, when receiving the sample data as input, is capable of providing an output which indicates an accurate and sensitive primary tumour type classification.

[0076] As used herein, the term “distinct location” will be understood to refer to each primary malignant tumour type having a different anatomical location, when in situ in a subject, relative to the anatomical location of each other primary malignant tumour type. The anatomical location may comprise a specific organ or grouping of organs of the subject. For example, the location may comprise the gastrointestinal tract, the head, the neck, the liver, cervix, biliary tract, bladder, breast, blood, kidney, lung, pancreas, prostate, skin, muscle, tendon, fat, lymph, blood vessel, nerves, bile duct, penis, bone, testes, thyroid, thymus, mesothelial tissue, ovary, uterus, adrenal gland, heart, spleen or gallbladder.

[0077] The head may comprise one or more of the oesophagus, brain, tongue, larynx, salivary gland, eye and cheek.

[0078] The gastrointestinal tract may comprise one or more of the stomach, oesophagus, colon, rectum, small intestine and large intestine.

[0079] In some embodiments, a distinct location comprises a specific location within an organ or area of the subject - for example, the upper gastrointestinal tract or lower gastrointestinal tract.

[0080] In the context of the present invention, the term “histological subtype”, will be understood to refer to a classification of the primary malignant tumour type based on, at least, the morphological appearance of the tumour cells. It will be appreciated that the morphological appearance of tumour cells between different histological subtypes (for example carcinoma and neuroendocrine) will typically differ. Therefore, the histological subtype may comprise the classification of the tissue type of the primary malignant tumour type. In some embodiments, the histological subtype is a classification of the primary malignant tumour type based on the morphological appearance of the tumour cells and the rate at which the tumour cells grow and divide.

[0081] The term “cell-free DNA” (cfDNA) as used herein, may otherwise be referred to as circulating free DNA. Typically, “cell-free DNA” refers to DNA which is not contained within intact cells (though the sample may contain one or more cells either intentionally or as trace impurities). Thus, cfDNA typically comprises DNA which has been released from cells into the subject’s biological fluid, typically circulatory biological fluid, such as, for example, blood, plasma, urine or cerebrospinal fluid.

[0082] The region methylation signature may comprise the DNA methylation status of a plurality of genomic regions. It will be appreciated that the term “genomic region” may be used interchangeably with the term “genomic window”. The methylation status may comprise a hyper or hypo methylation status. Generally, the methylation status comprises a measure of the degree to which hyper or hypo methylation is observed at the genomic region. The methylation status may comprise methylation sequencing reads.

[0083] In the context of the present invention, a genomic region will be understood to refer to a portion of the genome. The genomic region may have a length of 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 embodiments, a genomic region has a length of no more than about 2000 base pairs, no more than about 1900 base pairs, no more than about 1800 base pairs, no more than about 1700 base pairs, no more than about 1600 base pairs, no more than about 1500 base pairs, no more than about 1400 base pairs, no more than about 1300 base pairs, no more than about 1200 base pairs, no more than about 1100 base pairs, no more than about 1000 base pairs, no more than about 900 base pairs, no more than about 800 base pairs, no more than about 700 base pairs, no more than about 600 base pairs, no more than about 550 base pairs, no more than about 500 base pairs, no more than about 450 base pairs, no more than about 400 base pairs or no more than about 350 base pairs. In some embodiments, each genomic region comprises a length of from about 200 to about 2000 base pairs. In some embodiments, each genomic region comprises a length of from about 200 to about 1000 base pairs. In some embodiments, each genomic region comprises a length of from about 200 to about 400 base pairs. In some embodiments, each genomic region comprises a length of about 300 base pairs. The genomic regions will be discussed in further detail below.

[0084] In some embodiments, a genomic region comprises one or more CpG islands or CpG sites.

[0085] Primary Malignant Tumour Type

[0086] In some embodiments, the at least two primary malignant tumour types comprise at least three primary malignant tumour types, optionally at least four primary malignant tumour types, wherein each primary malignant tumour type has a distinct location and / or histological subtype.

[0087] In some embodiments, the at least two primary malignant tumour types comprise at least five primary malignant tumour types, at least six primary malignant tumour types, at least seven primary malignant tumour types, at least eight primary malignant tumour types, at least nine primary malignant tumour types or at least ten primary malignant tumour types. It will be appreciated that each primary malignant tumour type has a distinct location and / or histological subtype to the other primary malignant tumour types. In some embodiments the at least two primary malignant tumour types comprise at least eleven primary malignant tumour types, at least 12 primary malignant tumour types, at least 13 primary malignant tumour types, at least 14 primary malignant tumour types or at least 15 primary malignant tumour types.

[0088] In some embodiments, the at least two primary malignant tumour types comprise at least 20 primary malignant tumour types, each primary malignant tumour type having a distinct location and / or histological subtype.

[0089] In some embodiments, the at least two primary malignant tumour types comprise at least 21 primary malignant tumour types, at least 22 primary malignant tumour types, at least 23 primary malignant tumour types, at least 24 primary malignant tumour types or at least 25 primary malignant tumour types.

[0090] In some embodiments, the at least two primary malignant tumour types comprise at least 29 primary malignant tumour types.

[0091] In some embodiments, the at least two primary malignant tumour types comprise no more than 40 primary malignant tumour types, no more than 39 primary malignant tumour types, no more than 38 primary malignant tumour types, no more than 37 primary malignant tumour types, no more than 36 primary malignant tumour types or no more than 35 primary malignant tumour types.

[0092] In some embodiments the at least two primary malignant tumour types comprise no more than 30 primary malignant tumour types.

[0093] In some embodiments, the at least two primary malignant tumour types comprise of from 10 primary malignant tumour types to 40 primary malignant tumour types. The at least two primary malignant tumour types may comprise of from 20 primary malignant tumour types to 30 primary malignant tumour types. In some embodiments the at least two primary malignant tumour types comprise 29 primary malignant tumour types.

[0094] In some embodiments, the at least two primary malignant tumour types comprise at least one rare primary cancer tumour type. In some embodiments, the at least two primary malignant tumour types comprise at least two rare primary cancer tumour types. By “rare cancer” or “rare primary cancer”, this is typically defined as a cancer having an incidence of <6 / 100,000 general members of the population. The rare cancer may comprise a paediatric rare cancer or a teenage and young adult cancer. Further information on rare cancer 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 herein incorporated by reference in its entirety.

[0095] In some embodiments, the at least two primary malignant tumour types are selected from: a) a primary malignant hepato-pancreatobiliary tumour; b) a primary malignant female genital tract tumour; c) a primary malignant upper or lower gastrointestinal tumour; d) a primary malignant lung tumour; e) a primary malignant breast tumour; and f) a primary malignant urological tumour.

[0096] At least one of the at least two primary malignant tumour types may comprise a primary malignant hepato-pancreatobiliary tumour, optionally a primary malignant cholangiocarcinoma tumour.

[0097] In some embodiments, the at least two primary malignant tumour types are selected from: adrenal cancer, anal cancer, bile duct cancer, bladder cancer, bone cancer, brain / cns tumours, breast cancer, Castleman disease, cervical cancer, colon / rectum cancer, endometrial cancer, esophagus cancer, ewing family of tumours, eye cancer, gallbladder cancer, gastrointestinal carcinoid tumours, gastrointestinal stromal tumour (gist), gestational trophoblastic disease, Hodgkin disease, Kaposi sarcoma, kidney cancer, laryngeal and hypopharyngeal cancer, leukaemia (acute lymphocytic, acute myeloid, chronic lymphocytic, chronic myeloid, chronic myelomonocytic), liver cancer, lung cancer (non-small cell, small cell, lung carcinoid tumour), lymphoma, lymphoma of the skin, malignant mesothelioma, multiple myeloma, myelodysplastic syndrome, nasal cavity and paranasal sinus cancer, nasopharyngeal cancer, neuroblastoma, non-Hodgkin lymphoma, oral cavity and oropharyngeal cancer, osteosarcoma, ovarian cancer, penile cancer, pituitary tumours, prostate cancer, retinoblastoma, rhabdomyosarcoma, salivary gland cancer, sarcoma — adult soft tissue cancer, skin cancer (basal and squamous cell, melanoma, merkel cell), small intestine cancer, stomach cancer, testicular cancer, thymus cancer, thyroid cancer, uterine sarcoma, vaginal cancer, vulvar cancer, waldenstrom macroglobulinemia and Wilms tumour.

[0098] The at least two primary malignant tumour types may comprise breast cancer and at least one other primary malignant tumour type. The breast cancer may comprise hormone positive breast cancer or human epithelial growth factor-receptor 2 (HER2) positive breast cancer. Hormone positive breast cancer may comprise estrogen receptor positive (ER+) breast cancer, progesterone receptor positive (PR+) breast cancer or ER+PR+ breast cancer. In some embodiments, the breast cancer comprises triplenegative breast cancer.

[0099] In some embodiments, the at least two primary malignant tumour types are selected from: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynaelogical carcinoma (Gynae), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), acute myeloid leukaemia (LAML), brain lower grade glioma (LGG), liver hepatocellular carcinoma (LIHC), colon and rectum adenocarcinoma (Lower Gl), 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 tumours (TGCT), thyroid carcinoma (THCA), thymoma (THYM), stomach and oesophageal adenocarcinoma (Upper Gl), head and neck and oesophageal squamous cell carcinoma (UpperSq) and uveal melanoma (UVM).

[0100] For example, the at least two primary malignant tumour types may comprise at least two of the following primary malignant tumour types: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynaelogical carcinoma (Gynae), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), acute myeloid leukaemia (LAML), brain lower grade glioma (LGG), liver hepatocellular carcinoma (LIHC), colon and rectum adenocarcinoma (Lower Gl), 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 tumours (TGCT), thyroid carcinoma (THCA), thymoma (THYM), stomach and oesophageal adenocarcinoma (Upper Gl), head and neck and oesophageal squamous cell carcinoma (UpperSq) and uveal melanoma (UVM).

[0101] In some embodiments, the at least two primary malignant tumour types comprise at least three of the following primary malignant tumour types: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynaelogical carcinoma (Gynae), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), acute myeloid leukaemia (LAML), brain lower grade glioma (LGG), liver hepatocellular carcinoma (LIHC), colon and rectum adenocarcinoma (Lower Gl), 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 tumours (TGCT), thyroid carcinoma (THCA), thymoma (THYM), stomach and oesophageal adenocarcinoma (Upper Gl), head and neck and oesophageal squamous cell carcinoma (UpperSq) and uveal melanoma (UVM).

[0102] In some embodiments, the at least two primary malignant tumour types comprise at least four of the following primary malignant tumour types: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynaelogical carcinoma (Gynae), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), acute myeloid leukaemia (LAML), brain lower grade glioma (LGG), liver hepatocellular carcinoma (LIHC), colon and rectum adenocarcinoma (Lower Gl), 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 tumours (TGCT), thyroid carcinoma (THCA), thymoma (THYM), stomach and oesophageal adenocarcinoma (Upper Gl), head and neck and oesophageal squamous cell carcinoma (UpperSq) and uveal melanoma (UVM).

[0103] In some embodiments, the at least two primary malignant tumour types comprise at least five of the following primary malignant tumour types: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynaelogical carcinoma (Gynae), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), acute myeloid leukaemia (LAML), brain lower grade glioma (LGG), liver hepatocellular carcinoma (LIHC), colon and rectum adenocarcinoma (Lower Gl), 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 tumours (TGCT), thyroid carcinoma (THCA), thymoma (THYM), stomach and oesophageal adenocarcinoma (Upper Gl), head and neck and oesophageal squamous cell carcinoma (UpperSq) and uveal melanoma (UVM).

[0104] In the context of the present invention, the acronym “UpperSq” will be understood to refer to a grouping of head, neck and oesophageal squamous cell carcinomas. These different tumour types were grouped together in the group “UpperSq” by the inventors due to their similar anatomical and histological subtype. It will, therefore, be appreciated, that in the context of the present invention, the term “UpperSq” can encompass one or a combination of head and neck and oesophageal squamous cell carcinoma.

[0105] In some embodiments, at least one of the at least two primary malignant tumour types comprises UpperSq primary malignant tumour type. The UpperSq primary malignant tumour type may comprise a primary malignant head squamous cell carcinoma, a primary malignant neck squamous cell carcinoma or a primary malignant oesophageal squamous cell carcinoma. In some embodiments, the UpperSq primary malignant tumour type may comprise a primary malignant head squamous cell carcinoma and a primary malignant neck squamous cell carcinoma.

[0106] The term “non-squamous gynaelogical carcinoma”, as used herein, may refer to one or more of endocervical adenocarcinoma, ovarian cystadenocarcinoma, endometrial carcinoma and uterine carcinosarcomas. In some embodiments, at least one of the at least two primary malignant tumour types comprises a non-squamous gynaelogical carcinoma. The non-squamous gynaelogical carcinoma may comprise endocervical adenocarcinoma, ovarian cystadenocarcinoma, endometrial carcinoma and uterine carcinosarcoma. In some embodiments, the non-squamous gynaelogical carcinoma comprises endocervical adenocarcinoma. In some embodiments, the non-squamous gynaelogical carcinoma comprises ovarian cystadenocarcinoma. In some embodiments, the non-squamous gynaelogical carcinoma comprises endometrial carcinoma. In some embodiments, the non-squamous gynaelogical carcinoma comprises uterine carcinosarcoma.

[0107] The at least two primary malignant tumour types may comprise adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynaelogical carcinoma (Gynae), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), acute myeloid leukaemia (LAML), brain lower grade glioma (LGG), liver hepatocellular carcinoma (LIHC), colon and rectum adenocarcinoma (Lower Gl), 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 tumours (TGCT), thyroid carcinoma (THCA), thymoma (THYM), stomach and oesophageal adenocarcinoma (Upper Gl), head and neck and oesophageal squamous cell carcinoma (UpperSq) and uveal melanoma (UVM).

[0108] Thus, in some embodiments, the at least two primary malignant tumour types comprise the following 29 primary malignant tumour types: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynaelogical carcinoma (Gynae), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), acute myeloid leukaemia (LAML), brain lower grade glioma (LGG), liver hepatocellular carcinoma (LIHC), colon and rectum adenocarcinoma (Lower Gl), 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 tumours (TGCT), thyroid carcinoma (THCA), thymoma (THYM), stomach and oesophageal adenocarcinoma (Upper Gl), head and neck and oesophageal squamous cell carcinoma (UpperSq) and uveal melanoma (UVM).

[0109] In some embodiments, the training data further comprises a region methylation signature from at least one non-cancer control. In the context of the present invention, the term “non-cancer control” refers to an individual known not to have cancer. The “non-cancer control” individual may therefore be referred to as healthy.

[0110] The training data may comprise a region methylation signature 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 comprises a region methylation signature from at least 70 non- cancer controls, optionally from at least 75 non-cancer controls. In some embodiments, the training data comprises a region methylation signature from about 79 non-cancer controls. Advantageously, the provision of a region methylation signature from at least one non-cancer control avoids false positives. In particular, the inclusion of a region methylation signature from at least one non-cancer control enables an “unknown type classification” output for sample data wherein the region methylation signature does not align enough with any of the region methylation signatures of the primary malignant tumour types in the training data. An unknown type classification output may thus occur where the subject does not have cancer or has a cancer substantially different to those of the primary malignant tumour types in the training data. Alternatively, an unknown type classification may occur where the amount of tumour DNA present in the liquid biopsy is too low for detection and / or analysis.

[0111] In some embodiments, the region methylation signature from the at least one non-cancer control comprises a region methylation signature obtained from cfDNA from at least one non-cancer control. In some embodiments, the region methylation signature from the at least one non-cancer control comprises a region methylation signature obtained from tissue DNA from at least one non-cancer control, or from an organ or tissue which is known not to comprise cancer, which may otherwise be referred to as non- cancerous tissue, for example non-cancerous liver tissue. In some embodiments, the region methylation signature from the at least one non-cancer control comprises a region methylation signature obtained from cfDNA and a region methylation signature obtained from liver tissue DNA from at least one noncancer control. The present inventors have advantageously found that the inclusion of a region methylation signature obtained from liver tissue DNA from at least one non-cancer control increases the accuracy of the present method, in particular by reducing the chance of the output from the model erroneously indicating that a subject not having cancer has a primary liver cancer tumour.

[0112] The subject may have unfavourable CUP. By “unfavourable CUP”, this will be understood to refer to CUP which is not favourable CUP and hence is considered to be a cancer likely to be less responsive to treatment. In some embodiments, the subject is known to have secondary carcinoma, but the primary cancer is unknown. In some embodiments, the subject is known to have secondary adenocarcinoma, secondary squamous carcinoma or secondary poorly differentiated carcinoma, but the primary cancer is unknown.

[0113] In some embodiments, the primary cancer in step iv) is selected from: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynaelogical carcinoma (Gynae), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), acute myeloid leukaemia (LAML), brain lower grade glioma (LGG), liver hepatocellular carcinoma (LIHC), colon and rectum adenocarcinoma (Lower Gl), 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 tumours (TGCT), thyroid carcinoma (THCA), thymoma (THYM), stomach and oesophageal adenocarcinoma (Upper Gl), head and neck and oesophageal squamous cell carcinoma (UpperSq), uveal melanoma (UVM) and unclassified primary cancer.

[0114] In some embodiments, the primary cancer in step iv) is selected from breast cancer, female reproductive cancer, haematological cancer, kidney cancer, lower abdominal cancer, lung cancer, upper abdominal cancer, brain cancer, male reproductive cancer, NETs (neuroendocrine tumours), skin cancer, thyroid cancer, rare cancer and unclassified primary cancer.

[0115] Training Data

[0116] The training data comprises a region methylation signature from a primary malignant tumour type and a region methylation signature from at least one other primary malignant tumour type. Each primary malignant tumour type has a distinct location and / or histological subtype.

[0117] As described herein, each region methylation signature comprises the DNA methylation status of at least one genomic region. Typically, region methylation signatures of different primary malignant tumour types comprise the DNA methylation status of the same region(s) of the genome. It, will, however, be appreciated, that the DNA methylation status of the region(s) may differ between each primary malignant tumour type. Therefore, the region methylation signature of each primary malignant tumour type will be distinct.

[0118] Each region methylation signature in the training data may 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. In some embodiments, each region methylation signature in the training data comprises the DNA methylation status of at least 50, at least 100, at least 150, at least 200 or at least 250 genomic regions. In some embodiments, each region methylation signature in the training data comprises the DNA methylation status of at least 1000, at least 5000, at least 10000, at least 15000, at least 20000 or at least 25000 genomic regions. In some embodiments, each region methylation signature in the training data comprises the DNA methylation status of at least 5000 genomic regions. In some embodiments, each region methylation signature in the training data comprises the DNA methylation status of at least 1 x 106, at least 2 x 106, at least 3 x 106, at least 4 x 106, at least 5 x 106, at least 6 x 106, at least 7 x 106, at least 8 x 106or at least 9 x 106genomic regions.

[0119] The present inventors have observed that in embodiments comprising a plurality of genomic regions, a substantial proportion of these genomic regions can be close to one another. This may suggest that certain areas of the genome are especially indicative of cancer status. Thus, when the region methylation signature comprises the DNA methylation status of a plurality of genomic regions, at least 10%, at least 20%, at least 30%, at least 40% or at least 50% of the genomic regions in each region methylation signature may be no more than 1000 base pairs distance from at least one other genomic region in the region 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 region methylation signature is no more than 500 base pairs distance from at least one other genomic region in the region methylation signature. In some embodiments, at least 60%, optionally at least 70% of the genomic regions in each region methylation signature is no more than 1000 base pairs distance from at least one other genomic region in the region methylation signature. In some embodiments, at least 60%, optionally at least 70% of the genomic regions in each region methylation signature is no more than 500 base pairs distance from at least one other genomic region in the region methylation signature. In some embodiments, at least 80% of the genomic regions in each region methylation signature is no more than 1000 base pairs, optionally no more than 500 base pairs distance from at least one other genomic region in the region methylation signature.

[0120] 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 region methylation signature is adjacent to or overlapping with at least one other genomic region in the region methylation signature. For example, at least 60%, optionally at least 70% of the genomic regions in each region methylation signature may be adjacent to or overlapping with at least one other genomic region in the region methylation signature. In some embodiments, at least 80% of the genomic regions in each region methylation signature is adjacent to or overlapping with at least one other genomic region in the region methylation signature.

[0121] In some embodiments, each region methylation signature comprises the DNA methylation status of all genomic regions in the genome. This may otherwise be referred to as whole genome DNA methylation status.

[0122] In some embodiments, the genomic region(s) in each region methylation signature in the training data comprise genomic regions identified from a pairwise comparison of DNA methylation status of genomic regions between the at least two primary malignant tumour types. In some embodiments, the genomic region(s) in each region methylation signature in the training data comprise genomic regions identified from pairwise comparisons of DNA methylation status of genomic regions between the at least two primary malignant tumour types and at least one non-cancer control. In some embodiments, the genomic region(s) in each region methylation signature in the training data comprise genomic regions identified from pairwise comparisons of DNA methylation status of genomic regions between the at least two primary malignant tumour types and non-cancerous tissue. In the context of the present invention, a pairwise comparison of DNA methylation status will be understood to refer to a comparison of the DNA methylation status of genomic regions between two different types, for example between two different primary malignant tumour types, between a primary malignant tumour type and a non-cancer control, or between a primary malignant tumour type and non-cancerous tissue. The genomic region(s) identified from the pairwise comparison may comprise the genomic region(s) identified as having the greatest difference in DNA methylation status between the two types, for example between the two primary malignant tumour types.

[0123] The genomic regions in the training data may comprise the genomic regions, which, when compared using a pairwise comparison (between, for example, two different primary malignant tumour types or between a primary malignant tumour type and a non-cancer control), have the greatest difference in methylation status. Thus, the genomic regions in the training data may comprise at least one, at least 5, at least 10, at least 20, at least, 30, at least 50, at least 100, at least 150 or at least 200 genomic regions which have the greatest difference in methylation status as identified from a pairwise comparison. In some embodiments, the training data comprises about 250 genomic regions which have the greatest difference in methylation status as identified from a pairwise comparison. Difference in methylation status may be measured by statistical analysis. In such embodiments, the greatest difference may be the most statistically significant difference in methylation status. Statistical analysis may comprise placing a threshold on the adjusted p-values and ranking the resulting regions by difference in beta values.

[0124] The genomic regions in the training data may comprise the genomic regions, which, when compared using a pairwise comparison (between for example, two different primary malignant tumour types or between a primary malignant tumour type and a non-cancer control), have the greatest difference in hypomethylation. Alternatively, the genomic regions in the training data may comprise the genomic regions, which, when compared using a pairwise comparison (between for example, two different primary malignant tumour types or between a primary malignant tumour type and a non-cancer control), have the greatest difference in hypermethylation.

[0125] In some embodiments, the genomic regions in the training data comprise the genomic regions, which, when compared using a pairwise comparison (between for example, two different primary malignant tumour types or between a primary malignant tumour type and a non-cancer control), have the greatest difference in hypomethylation and the greatest difference in hypermethylation. Such a pairwise comparison may otherwise be referred to as an “up and down” pairwise comparison.

[0126] For example, in a pairwise comparison between primary malignant tumour type A and primary malignant tumour type B, of at least three genomic regions compared, primary malignant tumour type A may be hypermethylated in two genomic regions and hypomethylated at one genomic region. Therefore, primary malignant tumour type B will be hypomethylated in two genomic regions and hypermethylated at one genomic region.

[0127] The genomic region(s) in each region methylation signature in the training data may comprise 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 which have the greatest difference in hypermethylation. The genomic region(s) in each region methylation signature in the training data may comprise 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 which have the greatest difference in hypomethylation.

[0128] In some embodiments, the training data comprises about 250 genomic regions which have the greatest difference in hypermethylation as identified from a pairwise comparison.

[0129] In some embodiments, the training data comprises about 250 genomic regions which have the greatest difference in hypomethylation as identified from a pairwise comparison.

[0130] In some embodiments, the training data comprises about 250 genomic regions which have the greatest difference in hypermethylation and the 250 genomic regions which have the greatest difference in hypomethylation as identified from a pairwise comparison.

[0131] The genomic region(s) in each region methylation signature in the training data may comprise 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 region(s) identified as having the greatest differential methylation in one or more pairwise comparisons.

[0132] The one or more pairwise comparisons may comprise two or more pairwise comparisons, preferably three or more pairwise comparisons if the at least two primary malignant tumour types comprise at least three primary malignant tumour types or at least two primary malignant tumour types and one non-cancer control.

[0133] In some embodiments, the number of pairwise comparisons is calculated from the formula n(n-1) / 2, where n comprises the number of primary malignant tumour types. In embodiments where the training data comprises a region methylation signature from at least one non-cancer control, n may comprise the number of primary malignant tumour types + 1 , wherein + 1 represents the non-cancer control.

[0134] In embodiments comprising at least four primary malignant tumour types, the one or more pairwise comparisons may comprise at least six pairwise comparisons. In embodiments comprising at least five primary malignant tumour types, the one or more pairwise comparisons may comprise at least ten pairwise comparisons.

[0135] For example, when the at least two primary malignant tumour types comprise at least three primary malignant tumour types (which may otherwise be referred to as a first, second and third primary malignant tumour type), the pairwise comparisons may comprise: a) a pairwise comparison of methylation status of genomic regions between the first and the second primary malignant tumour type; b) a pairwise comparison of methylation status of genomic regions between the first and the third primary malignant tumour type; and c) a pairwise comparison of methylation status of genomic regions between the second and the third primary malignant tumour type.

[0136] When the at least two primary malignant tumour types comprise at least four primary malignant tumour types, the pairwise comparisons may comprise: a) a pairwise comparison of methylation status of genomic regions between the first and the fourth primary malignant tumour type; b) a pairwise comparison of methylation status of genomic regions between the first and the third primary malignant tumour type; c) a pairwise comparison of methylation status of genomic regions between the first and the second primary malignant tumour type; d) a pairwise comparison of methylation status of genomic regions between the second and fourth primary malignant tumour type; e) a pairwise comparison of methylation status of genomic regions between the third and fourth primary malignant tumour type; and f) a pairwise comparison of methylation status of genomic regions between the second and the third primary malignant tumour type.

[0137] The genomic region(s) in each region methylation signature in the training data may comprise 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. By “top 5”, this will be understood to refer to the five genomic regions identified as having the greatest differential methylation in the at least one pairwise comparison. The greatest differential methylation may comprise greatest hypermethylation or greatest hypomethylation. In some embodiments the greatest differential methylation comprises greatest hypermethylation and greatest hypomethylation.

[0138] In some embodiments, the plurality of genomic regions comprises no more than about 500,000 genomic regions, no more than about 450000 genomic regions, no more than about 400000 genomic regions, no more than about 350000 genomic regions, no more than about 300000 genomic regions, or no more than about 250000 genomic regions. In some embodiments, the plurality of genomic regions comprises no more than about 50000 genomic regions, no more than about 40000 genomic regions, no more than about 30000 genomic regions, or no more than about 25000 genomic regions.

[0139] In embodiments where the genomic regions were identified from a plurality of pairwise comparisons, the genomic regions may comprise 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 each pairwise comparison. The genomic regions may comprise 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 as having the greatest differential hypermethylation from each pairwise comparison. In some embodiments, the genomic regions comprise 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 as having the greatest differential hypomethylation from each pairwise comparison. In some embodiments, the genomic regions comprise 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 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 top 250 genomic regions identified as having the greatest differential hypermethylation.

[0140] For example, the genomic regions may comprise about 250 genomic regions identified per pairwise comparison. The genomic regions may comprise about 250 genomic regions identified as having the greatest hypermethylation and about 250 genomic regions identified as having the greatest hypermethylation. Thus, in embodiments concerning, for example, six pairwise comparisons, the genomic regions may comprise about 500 genomic regions identified from each pairwise comparison (250 having the greatest hypermethylation and 250 having the greatest hypomethylation, thus a total of about 3000 genomic regions).

[0141] In some embodiments, a non-cancer control in the pairwise comparison is a non-cancer liver control, which may otherwise be referred to as a normal liver class. The non-cancer control in the pairwise comparison may not comprise any non-liver control.

[0142] Prior to the pairwise comparisons, genomic regions from non-cancer controls having a median methylation beta value of at least 0.05 may be excluded from the training data. For example, prior to the pairwise comparisons, genomic regions from non-cancer controls having a median methylation beta value of at least about 0.25 may be excluded from the training data. In some embodiments, prior to the pairwise comparisons, genomic regions from non-cancer controls having a median methylation beta value of at least about 0.5 are excluded from the training data. Alternatively, genomic regions from non-cancer controls having the methylation beta value described above may be excluded from the training data following the pairwise comparisons.

[0143] The inventors have identified a plurality of genomic regions which exhibit the greatest difference in methylation status in pairwise comparisons of different primary malignant tumour types. Advantageously, the inventors used the DNA methylation status of these specific genomic regions as training data on which to train the present machine model. This has resulted in a machine model which is accurately able to distinguish between and allocate primary malignant tumour types. In particular, the inventors have identified the top 250 genomic regions which exhibit the greatest difference in methylation status when analysed in a pairwise comparison between primary malignant tumour types. Increased hypermethylation at a genomic region for one primary malignant tumour type, compared to another primary malignant tumour type may be indicative of an association with the one primary malignant tumour type. Alternatively, increased hypomethylation at a genomic region for one primary malignant tumour type, compared to another primary malignant tumour type may be indicative of an association with the one primary malignant tumour type. Therefore, the regions identified as having the greatest difference in methylation status in pairwise comparisons of different primary malignant tumour types are the regions which the inventors have found especially useful training data. The inventors have advantageously used such training data to generate an accurate and effective machine learning model which can then indicate the primary malignant tumour type of the subject.

[0144] Table 1 comprises the top 250 genomic regions identified by the inventors from each of 30 pairwise comparisons. These genomic regions were used in the training data of Examples 1 to 6 in “CUPiD v1”. CUPiD v1 is a machine model which was trained on a training data set comprising region methylation signatures from 29 primary malignant tumour types, a general non-cancer control and a normal liver class. The first three columns of Table 1 indicate the genomic coordinates of each genomic region. The fourth column lists the primary malignant tumour type classes where the specific genomic region is hypermethylated (i.e. higher than the other class) in at least one pairwise comparison. The fifth column shows the pairwise comparisons where the genomic region is in the top 250 most hypermethylated genomic region. For example, TGCT>Gynae means that that particular genomic region is in the top 250 genomic regions having the greatest difference in methylation at that region between TGCT and Gynae (with higher methylation in TGCT).

[0145] Table 3 of the present application provides the full definition of each acronym used in the fourth and fifth columns of Table 1. For example, TGCT is used to refer to testicular germ cell tumours, while Gynae is used to refer to non-squamous gynaelogical carcinomas (endocervical adenocarcinoma, ovarian cystadenocarcinoma, endometrial carcinoma, uterine carcinosarcomas).

[0146] Thus, in some embodiments, the genomic regions are selected from the genomic regions listed in Table 1.

[0147] In embodiments wherein a primary malignant tumour type is adrenocortical carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “ACC >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is bladder urothelial carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “BLCA >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is breast invasive carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “BRCA >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is cervical squamous cell carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “CervSq >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is cholangiocarcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “CHOL >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is diffuse large B-cell lymphoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “DLBC >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is glioblastoma multiforme, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “GBM >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is non-squamous gynaelogical carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “Gynae >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is kidney chromophobe, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “KICH >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is kidney renal clear cell carcinoma , the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “KIRC >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is kidney renal papillary cell carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “KIRP >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is acute myeloid leukaemia, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “LAML >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is brain lower grade glioma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “LGG >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is liver hepatocellular carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “LIHC >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is colon and rectum adenocarcinomas, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “LowerGI >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is lung adenocarcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “LUAD >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is lung squamous cell carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “LUSC >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is mesothelioma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “MESO >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is pancreatic adenocarcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “PAAD >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is pheochromocytoma and paraganglioma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “PCPG >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is prostate adenocarcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “PRAD >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is sarcoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “SARC >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is skin cutaneous melanoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “SKCM >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is testicular germ cell tumours, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “TGCT >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is thyroid carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “THCA >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is thymoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “THYM >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is stomach and oesophageal adenocarcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “UpperGI >” is presented in the fourth column of Table 1. In embodiments wherein a primary malignant tumour type is head and neck and oesophageal squamous cell carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “UpperSq >” is presented in the fourth column of Table 1 . In embodiments wherein a primary malignant tumour type is uveal melanoma, the genomic regions may be selected from any of the genomic regions listed in Table 1 wherein “UVM >” is presented in the fourth column of Table 1.

[0148] In some embodiments, the genomic regions are selected from the genomic regions listed in Table 2. The genomic regions listed in Table 2 have been identified by the inventors as the most informative genomic regions for discriminating between primary malignant tumour types, as determined by variable importance score for pairwise comparison between different primary malignant tumour class types. The first column identifies the chromosome on which the genomic region is found. The start and end point of the genomic region is then listed in columns 2 and 3 of Table 2, respectively. These genomic regions have been ranked in Table 2 by variable importance score- with those at the start of the table deemed more important for the present model than those at the bottom of the table (see column 4, detailing an average importance). In some embodiments, the genomic regions comprise at least five, at least 10, at least 20, at least 30, at least 40 or at least 50 of the genomic regions listed in Table 2. In some embodiments, the genomic regions comprise at least the top five, the top 10, the top 20, the top 30, the top 40, the top 50, the top 60, the top 70, the top 80 or the top 90 genomic regions in Table 2. By “top 10”, this will be understood to refer to the first 10 genomic regions listed in Table 2, for example. In some embodiments, the genomic regions comprise the top 100 genomic regions listed in Table 2. In some embodiments, the genomic regions comprise the top 150 genomic regions listed in Table 2. In some embodiments, the genomic regions comprise the top 200 genomic regions listed in Table 2. In some embodiments, the genomic regions comprise all 250 genomic regions listed in Table 2.

[0149] In some embodiments, the genomic regions are selected from the genomic regions 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 for discriminating between primary malignant tumour types, as determined by variable importance score for pairwise comparison between different primary malignant tumour class types in CUPiD v1 . The first column identifies the chromosome on which the genomic region is found. The start and end point of the genomic region is then listed in columns 2 and 3 of Table 9, respectively. These genomic regions have been ranked in Table 9 by variable importance score- with those at the start of the table deemed more important for the present model than those at the bottom of the table (see column 4, detailing an average importance).Thus, Table 9 comprises all 250 genomic regions listed in Table 2 and a further 250 genomic regions. It will be appreciated that these further 250 genomic regions are also listed in Table 1 . In some embodiments, the genomic regions comprise at least five, at least 10, at least 20, at least 30, at least 40 or at least 50 of the genomic regions listed in Table 9. In some embodiments, the genomic regions comprise at least the top 300, the top 350, the top 400, or the top 450 genomic regions in Table 9. In some embodiments, the genomic regions comprise all 500 genomic regions listed in Table 9.

[0150] In some embodiments, the genomic regions are selected from the genomic regions listed in Table 5. Table 5 comprises the top 250 genomic regions identified by the inventors from pairwise comparisons in a later version of the model (“CUPiD v1 .1”). These pairwise comparisons were between the same 29 primary malignant tumour classes as CUPiD v1 and a normal liver class.

[0151] In contrast to Table 1 , the pairwise comparisons of Table 5 did not include any comparison to non-cancer control samples, other than a normal liver class. The genomic regions of Table 5 were used in the training data of Examples 8 to 10. 11 ,809 of the same genomic regions are found in Tables 1 and 5. Thus, in some embodiments, the genomic regions are selected from genomic regions found in both Tables 1 and 5.

[0152] In embodiments wherein a primary malignant tumour type is adrenocortical carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “ACC >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is bladder urothelial carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “BLCA >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is breast invasive carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “BRCA >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is cervical squamous cell carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “CervSq >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is cholangiocarcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “CHOL >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is diffuse large B-cell lymphoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “DLBC >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is glioblastoma multiforme, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “GBM >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is non-squamous gynaelogical carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “Gynae >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is kidney chromophobe, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “KICH >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is kidney renal clear cell carcinoma , the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “KIRC >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is kidney renal papillary cell carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “KIRP >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is acute myeloid leukaemia, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “LAML >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is brain lower grade glioma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “LGG >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is liver hepatocellular carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “LIHC >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is colon and rectum adenocarcinomas, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “LowerGI >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is lung adenocarcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “LUAD >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is lung squamous cell carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “LUSC >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is mesothelioma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “MESO >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is pancreatic adenocarcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “PAAD >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is pheochromocytoma and paraganglioma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “PCPG >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is prostate adenocarcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “PRAD >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is sarcoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “SARC >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is skin cutaneous melanoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “SKCM >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is testicular germ cell tumours, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “TGCT >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is thyroid carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “THCA >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is thymoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “THYM >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is stomach and oesophageal adenocarcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “UpperGI >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is head and neck and oesophageal squamous cell carcinoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “UpperSq >” is presented in the fourth column of Table 5. In embodiments wherein a primary malignant tumour type is uveal melanoma, the genomic regions may be selected from any of the genomic regions listed in Table 5 wherein “UVM >” is presented in the fourth column of Table 5.

[0153] In some embodiments, the genomic regions are selected from the genomic regions listed in Table 6. The genomic regions listed in Table 6 have been identified by the inventors as the top 250 genomic regions overall for discriminating between primary malignant tumour types, as determined by variable importance score in CUPiD v1 .1 . The first column identifies the chromosome on which the genomic region is found. The start and end point of the genomic region is then listed in columns 2 and 3 of Table 6, respectively. These genomic regions have been ranked in Table 6 by variable importance score- with those at the start of the table deemed more important than those at the bottom of the table (see column 4, detailing an average importance). In some embodiments, the genomic regions comprise at least five, at least 10, at least 20, at least 30, at least 40 or at least 50 of the genomic regions listed in Table 6. In some embodiments, the genomic regions comprise at least the top five, the top 10, the top 20, the top 30, the top 40, the top 50, the top 60, the top 70, the top 80 or the top 90 genomic regions in Table 6. By “top 10”, this will be understood to refer to the first 10 genomic regions listed in Table 6, for example. In some embodiments, the genomic regions comprise the top 100 genomic regions listed in Table 6. In some embodiments, the genomic regions comprise the top 150 genomic regions listed in Table 6. In some embodiments, the genomic regions comprise the top 200 genomic regions listed in Table 6. In some embodiments, the genomic regions comprise all 250 genomic regions listed in Table 6.

[0154] Many of the genomic regions of Table 2 are also common to Table 6. In total, 174 of the genomic regions listed in Table 2 are also present in Table 6. These are listed in Table 7. It will be appreciated that these 174 genomic regions are also found in Tables 1 and 5. Thus, in some embodiments, the genomic regions comprise a plurality of regions common to both Tables 2 and 6. In other words, the genomic regions may comprise a plurality of regions selected from Table 7. For example, the genomic regions may comprise at least five, at least 10, at least 20, at least 30, at least 40 or at least 50 of the genomic regions common to Tables 2 and Table 6 (as found in Table 7). Of the top 5 genomic regions of Table 2, four of these are also in the top 7 genomic regions of Table 6. Furthermore, out of the top 10 genomic regions of Table 2, six of these are also in the top 10 of Table 6. Thus, in some embodiments, the genomic regions comprise at least the top 2, 3, 4, and 5 listed genomic regions of Table 2, as also found in Table 6. In some embodiments, the genomic regions comprise at least five of the top 10 genomic regions of Table 2, optionally at least 6, at least 7, at least 8 or at least 9 of the top 10 genomic regions of Table 2. Of the top 10 genomic regions of Table 6, 8 (genomic regions ranked 2-9) are also found in Table 2. Thus, in some embodiments, the genomic regions comprise the genomic regions ranked 2-9 in Table 6.

[0155] In some embodiments, the genomic regions comprise all genomic regions listed in Table 7.

[0156] In some embodiments, the genomic regions are selected from the genomic regions listed in Table 8. The genomic regions listed in Table 8 have been identified by the inventors as top 500 genomic regions overall for discriminating between primary malignant tumour types, as determined by variable importance score in CUPiD v1 .1 . The first column identifies the chromosome on which the genomic region is found. The start and end point of the genomic region is then listed in columns 2 and 3 of Table 8, respectively. These genomic regions have been ranked in Table 8 by variable importance score- with those at the start of the table deemed more important than those at the bottom of the table (see column 4, detailing an average importance). In some embodiments, the genomic regions comprise at least five, at least 10, at least 20, at least 30, at least 40 or at least 50 of the genomic regions listed in Table 8. In some embodiments, the genomic regions comprises the top 300 genomic regions listed in Table 8. In some embodiments, the genomic regions comprises the top 350 genomic regions listed in Table 8. In some embodiments, the genomic regions comprises the top 400 genomic regions listed in Table 8. In some embodiments, the genomic regions comprises the top 450 genomic regions listed in Table 8. In some embodiments, the genomic regions comprise all genomic regions listed in Table 8.

[0157] 224 of the genomic regions listed in Table 8 are also listed in Table 2 (the top 250 genomic regions for CUPiD v1). Thus, in some embodiments, the genomic regions comprise genomic regions common to Table 2 and Table 8. For example, the genomic regions may comprise at least 150, at least 200 or all 224 of the genomic regions common to Tables 8 and Table 2.

[0158] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75% or at least 90% of the genomic regions employed in the method of any aspect of the invention are selected from those set forth in any of Tables 1 , 2, 5, 6, 7, 8 and 9.

[0159] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75% or at least 90% of the genomic regions employed in the method of any aspect of the invention are selected from those set forth in Table 1 or Table 2. In some embodiments, at least 10%, at least 25%, at least 50%, at least 75% or at least 90% of the genomic regions employed in the method of any aspect of the invention are selected from those set forth in Table 5, Table 6, Table 7, Table 8 or Table 9.

[0160] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75% or at least 90% of the genomic regions employed in the method of any aspect of the invention are selected from those set forth in Table 6, Table 7, Table 8 or Table 9.

[0161] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75% or at least 90% of the genomic regions employed in the method of any aspect of the invention are selected from those set forth in Table 6.

[0162] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75% or at least 90% of the genomic regions employed in the method of any aspect of the invention are selected from those set forth in Table 7.

[0163] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75% or at least 90% of the genomic regions employed in the method of any aspect of the invention are selected from those set forth in Table 1 .

[0164] In some embodiments, at least 10%, at least 25%, at least 50%, at least 75% or at least 90% of the genomic regions employed in the method of any aspect of the invention are selected from those set forth in Table 5.

[0165] In some embodiments, the genomic regions do not comprise any of the genomic regions of Table 7. In some embodiments, the genomic regions do not comprise any of the genomic regions of Table 2. In some embodiments, the genomic regions do not comprise any of the genomic regions of Table 6. In some embodiments, the genomic regions do not comprise any of the genomic regions of Tables 2 and 6.

[0166] In some embodiments, the genomic regions do not comprise any of the genomic regions of Table 8.

[0167] In some embodiments, the genomic regions do not comprise any of the genomic regions of Table 9.

[0168] In some embodiments, the genomic regions are selected from genomic regions listed in Tables 1 and / or 5 but are not listed in Tables 2, 6, 7, 8 and 9. In some embodiments, the genomic regions are selected from genomic regions listed in Tables 1 and / or 5 but are not listed in Table 7. In some embodiments, the genomic regions are selected from genomic regions listed in Tables 1 and / or 5 but are not listed in Table 2. In some embodiments, the genomic regions are selected from genomic regions listed in Tables 1 and / or 5 but are not listed in Table 6. In some embodiments, the genomic regions are selected from genomic regions listed in Tables 1 and / or 5 but are not listed in Table 8. In some embodiments, the genomic regions are selected from genomic regions listed in Tables 1 and / or 5 but are not listed in Tables 2 and 6. In some embodiments, the genomic regions are selected from genomic regions listed in Tables 1 and / or 5 but are not listed in Table 9.

[0169] In some embodiments, less than 50% of the genomic regions are selected from those set forth in Table 2.

[0170] In some embodiments, less than 50% of the genomic regions are selected from those set forth in Table 6. In some embodiments, less than 50% of the genomic regions are selected from those set forth in Table 7. In some embodiments, less than 50% of the genomic regions are selected from those set forth in Tables 2 and 6. In some embodiments, less than 50% of the genomic regions are selected from those set forth in Table 8. In some embodiments, less than 50% of the genomic regions are selected from those set forth in Table 9.

[0171] In some embodiments, less than 40%, optionally less than 30% of the genomic regions are selected from those set forth in Table 2. In some embodiments, less than 40%, optionally less than 30% of the genomic regions are selected from those set forth in Table 6. In some embodiments, less than 40%, optionally less than 30% of the genomic regions are selected from those set forth in Table 7. In some embodiments, less than 40%, optionally less than 30% of the genomic regions are selected from those set forth in Tables 2 and 6. In some embodiments, less than 40%, optionally less than 30% of the genomic regions are selected from those set forth in Table 8. In some embodiments, less than 40%, optionally less than 30% of the genomic regions are selected from those set forth in Table 9.

[0172] In some embodiments, less than 20% of the genomic regions are selected from those set forth in Table 2.

[0173] In some embodiments, less than 20% of the genomic regions are selected from those set forth in Table 6.

[0174] In some embodiments, less than 20% of the genomic regions are selected from those set forth in Table 7.

[0175] In some embodiments, less than 20% of the genomic regions are selected from those set forth in Tables

[0176] 2 and 6. In some embodiments, less than 20% of the genomic regions are selected from those set forth in Table 8. In some embodiments, less than 20% of the genomic regions are selected from those set forth in Table 9.

[0177] In some embodiments, the genomic regions are selected based on a function of similarity between each primary malignant tumour type. For example, if the primary malignant tumour types are from a similar location or histological subtype, a greater number of genomic regions may be selected to improve accuracy in distinguishing between the two similar primary malignant tumour types.

[0178] In some embodiments, the training data comprises region methylation signatures derived from primary malignant tumour tissue samples obtained from a plurality of individuals known to have a primary malignant cancer. In other words, the DNA methylation status of genomic regions in the training data may comprise the DNA methylation status of genomic regions obtained from primary malignant tumour tissue samples. The primary malignant tumour tissue samples may have been obtained from a plurality of individuals diagnosed with a specific primary malignant cancer. In other embodiments, the training data comprises region methylation signatures derived from cfDNA obtained from a plurality of individuals diagnosed with a specific primary malignant cancer. Alternatively, the training data may comprise region methylation signatures derived from cfDNA obtained from a plurality of individuals previously labelled as having CUP but subsequently have had their primary cancer identified. It will be appreciated that in such embodiments, the training data comprises region methylation signatures derived from cfDNA obtained from liquid samples from the plurality of individuals. In some embodiments, the region methylation signature of each primary malignant tumour type in the training data comprises the DNA methylation status of different genomic regions from the region methylation signature of each other primary malignant tumour type.

[0179] In the context of the present invention, it will be appreciated that the training data comprises a region methylation signature for each primary malignant tumour type. The region methylation signature for each primary malignant tumour type may comprise a DNA methylation status of at least one genomic region derived from a primary malignant tumour tissue sample and a DNA methylation status of at least one genomic region derived from cfDNA from at least one non-cancer control. This may otherwise be referred to as an “in silico mix”. Advantageously, an in silico mix mimics the proportion of circulating tumour (ct) DNA which would normally be expected in cfDNA but by using more readily available data from tumour tissue samples. In addition, an in silico mix enables the generation of a large dataset on which the model can be trained from an initial small volume of DNA data from primary malignant tumour tissue samples. Furthermore, it can be difficult to determine how much DNA from a primary malignant tumour tissue sample is likely to be shed into the biological fluid of a subject to form ctDNA. Therefore, by “diluting” the DNA methylation status of genomic region(s) derived from a primary malignant tumour tissue sample with a DNA methylation status derived from cfDNA from at least one non-cancer control, this gives greater accuracy to the training data and thus the model.

[0180] The methylation status of the genomic regions may be obtained from MBD-Seq Reads. In some embodiments, the MBD-Seq Reads are equivalent MBD-Seq Reads obtained by the conversion of beta values into equivalent MBD-Seq Reads. As the skilled person will appreciate, beta values are known forms of quantification for methylation sequencing. Beta values are methylation estimates as measured by a probe. The beta values may be obtained from methylation arrays, bisulphite sequencing or enzymatic methyl sequencing.

[0181] Conversion of beta values from methylation arrays into equivalent MBD-Seq Reads may be as described in Chemi et al., 2022, which is herein incorporated by reference in its entirety. For example, conversion of beta values into equivalent MBD-Seq Reads may comprise the calculation of beta values for MBD-Seq windows, using a suitable program such as qsea. The beta values for fitting may be obtained by calculating an average enrichment profile typical of a T7-MBD-Seq method. In particular, the beta values for fitting may be obtained by calculating an average enrichment profile from a plurality of non-cancer control cfDNA samples. The average enrichment profile may comprise fitted beta values for every window CG density in values rounded to one decimal place, for varying numbers of counts in each window. This profile can then be used as a comparison to determine how many counts are required for each beta value for each CG density and can be used to convert from beta values to estimated T7-MBD-seq counts.

[0182] Machine learning model

[0183] In some embodiments, the machine learning model comprises one or more of a decision tree, a logistic regression model, an artificial neural network, a support vector machine (SVM), naive bayes, linear discriminant analysis or k-nearest neighbour algorithm, optionally wherein the decision tree comprises a gradient boosting algorithm or a random forest algorithm.

[0184] In some embodiments the machine learning model comprises one or more of a decision tree or a logistic regression model.

[0185] The logistic regression model may comprise a penalised logistic regression model. A penalised logistic regression model may comprise a penalty term of L1 (lasso logistic regression), L2 (ridge logistic regression) or a combination of L1 and L2 penalties (elastic net logistic regression).

[0186] In some embodiments, the machine learning model comprises a decision tree comprising a gradient boosting algorithm.

[0187] In some embodiments, the machine learning model comprises 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 comprise at least 60 individual classifiers, optionally at least 100 individual classifiers. In some embodiments, the machine learning model comprises 100 individual classifiers.

[0188] In some embodiments, each individual classifier comprises an Extreme Gradient Boosting Tree.

[0189] In some embodiments, the output comprises a probabilistic score for each classification. In some embodiments, the output comprises ranked probabilistic scores for each classification.

[0190] The output may comprise a primary malignant tumour type classification for the sample data when the probabilistic score for said primary malignant tumour type classification is greater than 0.5. The output may comprise the unknown type classification for the sample data when the probabilistic score for the unknown type classification is greater than 0.5, or when the probabilistic score for each classification is less than 0.5.

[0191] In some embodiments, the output comprises a primary malignant tumour type classification for the sample data when the probabilistic score for said primary malignant tumour type classification is greater than 0.3. In some embodiments, the output comprises a primary malignant tumour type classification for the sample data when the probabilistic score for said primary malignant tumour type classification is greater than 0.4.

[0192] In some embodiments, the output comprises ranked probabilistic scores for each classification and the primary cancer of the subject is determined to be the primary malignant tumour type classification having the highest ranked probabilistic score. In embodiments wherein the output comprises ranked probabilistic scores for each classification, the primary cancer of the subject may be determined to be the top ranked primary malignant tumour type classification if the difference in probabilistic score between the top ranked and second ranked primary malignant tumour type classification is greater than at least about 0.2. In embodiments wherein the output comprises ranked probabilistic scores for each classification, the primary cancer of the subject may be determined to be an unknown or unclassified cancer if the difference in probabilistic score between the top ranked and second ranked primary malignant tumour type classification is less than about 0.05, optionally less than about 0.04, further optionally less than about 0.03. In other embodiments, the primary cancer of the subject may be determined to be the first or the second ranked primary malignant tumour type classification, if the difference in probabilistic score between the top ranked and second ranked primary malignant tumour type classification is less than about 0.05, optionally less than about 0.04, further optionally less than about 0.03. In some embodiments, the primary cancer of the subject may be determined to be the first or the second ranked primary malignant tumour type classification, if the difference in probabilistic score between the top ranked and second ranked primary malignant tumour type classification is less than about 0.3, optionally less than about 0.2, further optionally less than about 0.1 .

[0193] For example, if: a) the difference in probabilistic score between the top ranked and second ranked primary malignant tumour type classification is less than about 0.05, optionally less than about 0.04, further optionally less than about 0.03; and b) the top ranked and second ranked primary malignant tumour type classification are primary malignant tumour types having the same or similar location, then the primary cancer of the subject may be determined to be the first or the second ranked primary malignant tumour type classification.

[0194] An example of two primary malignant tumour types having the same or similar anatomical location is cholangiocarcinoma and pancreatic adenocarcinoma. Cholangiocarcinoma may otherwise be referred to as cancer of the bile duct. Thus, cholangiocarcinoma and pancreatic adenocarcinoma have the same location (pancreatic region).

[0195] In some embodiments, the primary malignant tumour type classification(s) of the output correspond to the primary malignant tumour types on which the model is trained. For example, in embodiments where the output comprises a probabilistic score for each classification, the classifications may correspond to the primary malignant tumour types on which the model is trained. In embodiments where the output comprises one primary malignant tumour type classification, this classification may correspond to one of the primary malignant tumour types on which the model is trained.

[0196] Alternatively, the primary malignant tumour type classification(s) of the output may comprise broader classification(s) than the primary malignant tumour type on which the model is trained. The broader classifications may comprise classifications which specify a broader anatomical area. The broader classifications may not distinguish between histological subtypes. Exemplary broader classifications may include, but not necessarily be limited to brain cancer, breast cancer, female reproductive cancer, haematological cancer, kidney cancer, lower abdominal cancer, lung cancer, male reproductive cancer, neuroendocrine tumours, rare cancer, skin cancer, thyroid cancer, upper abdominal cancer and unclassified primary cancer. The broader classifications may not comprise a non-cancer class. While the present invention has utility in accurately determining the anatomical location and histological subtype of the primary cancer, the present inventors have also found that there is also utility in a broad classification of cancer. Such broad classifications can be helpful to initially stratify subjects, in order to determine initial treatment strategies or further tests which might be helpful to specify the particular primary malignant tumour type.

[0197] In some embodiments, probabilistic scores for each classification are accumulated to form an accumulated score for one or more broader cancer classifications. The primary cancer of the subject may be determined to be the broader cancer classification having the highest ranked accumulated score. In some embodiments, the primary cancer of the subject is determined to be the broader cancer classification having an accumulated score of at least about 0.5. In some embodiments, the primary cancer of the subject is determined to be the broader cancer classification having an accumulated score of more than 0.5.

[0198] Determining a broader cancer classification may comprise the last step in the first layer of a two-layer computer-implemented model, where steps (i) to (iv) represent the first layer. Thus, in embodiments where the primary cancer of the subject is determined to be a broader cancer classification, the method may further comprise step (v) inputting the output from step (iii) into a second machine learning model comprising a multi-class classifier trained on a training data as described herein. The training data may be identical to or different from the training data on which the machine learning model of step (ii) was trained. The method may further comprise step (vi) comprising receiving an output from the second machine learning model, the output indicative of a primary malignant tumour type classification for the sample data and determining the primary cancer of the subject based on the output received in step (vi), wherein the primary cancer is a cancer subtype with a distinct anatomical and histological subtype.

[0199] Sample Data

[0200] It will be appreciated that the region methylation signature of the sample data comprises the DNA methylation status of at least one region of the genome of the cfDNA. As used herein, the “DNA methylation status” will be understood to refer to the level of DNA methylation. DNA methylation may be DNA hypermethylation or DNA hypomethylation.

[0201] In some embodiments, the region methylation signature of the sample data comprises the DNA methylation status of a plurality of genomic regions.

[0202] The region methylation signature of the sample data may be as defined for the region methylation signature of the training data. In some embodiments, the region methylation signature of the sample data comprises the DNA methylation status of the same genomic regions as for the training data. As used herein, “same genomic region” will be understood to refer to two or more genomic regions having the same length and genomic coordinates, but originating from different samples, for example from a subject liquid biopsy sample and a primary malignant tumour sample, or from samples from two or more different primary malignant tumour samples. Therefore, while the “same” genomic regions may have the same length and genomic coordinates, they may each have a different DNA methylation status.

[0203] Alternatively, the region methylation signature of the sample data may comprise the DNA methylation status of a subset of the genomic regions employed in the training data. While predictive performance may be reduced where the sample data is "incomplete" in terms of genomic regions, the impact is expected to be modest where the sample data comprises the majority of genomic regions of the training data. For example, the region methylation signature of the sample data may comprise 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 employed in the training data.

[0204] The liquid sample may comprise any biological fluid. For example, the liquid sample may comprise a blood sample, plasma sample, serum, lymphatic fluid, synovial fluid, ascites fluid, interstitial or extracellular fluid, cerebrospinal fluid, saliva, mucus, semen, sweat, urine or any other bodily fluids. In some embodiments, the liquid sample comprises a blood, urine or plasma sample. In some embodiments, the liquid sample comprises a blood or plasma sample. In some embodiments, the liquid sample comprises a blood sample. In some embodiments, the liquid sample comprises a plasma sample. The sample may be one which has been freshly obtained from a subject or may be one which has been processed and / or stored prior to making a determination (e.g. frozen, fixed or subjected to one or more purification, enrichment or extractions steps). The sample is preferably from a mammalian (such as e.g. a mammalian cell sample or a sample from a mammalian subject, including in particular a model animal such as mouse, rat, etc.), preferably from a human (such as e.g. a human cell sample or a sample from a human subject). Further, the sample may be transported and / or stored, and collection may take place at a location remote from the genomic sequence data acquisition (e.g. sequencing) location, and / or the computer- implemented method steps may take place at a location remote from the sample collection location and / or remote from the genomic data acquisition (e.g. sequencing) location (e.g. the computer-implemented method steps may be performed by means of a networked computer, such as by means of a “cloud” provider).

[0205] In some embodiments, the sample data further comprises sequence data of DNA obtained from a tissue sample from the subject. In such embodiments, it will be appreciated that the sample comprises a liquid sample and a tissue sample. Any biological tissue can be obtained forthe tissue sample. The tissue sample, may for, example, comprise liver tissue, lymph node tissue, bone tissue, lung tissue, skin tissue, gastrointestinal tract tissue, or tissue from any other suitable organ or other site. The skilled person will be aware of suitable organs / sites from which tissue samples can be obtained. In some embodiments, the tissue sample comprises a tumour sample. A “tumour sample” refers to a sample that contains tumour cells or genetic material derived therefrom. The tumour sample may be a cell or tissue sample (e.g. a biopsy) obtained directly from a tumour.

[0206] The tissue sample may comprise a fresh tissue sample, a frozen tissue sample (for example frozen at - 20°C or at -80°C) or a formalin-fixed paraffin embedded tissue sample.

[0207] In some embodiments, the sample data further comprises sequence data of DNA obtained from circulating tumour cells in the liquid sample.

[0208] In some embodiments the liquid sample comprises 1 pg (i.e. picogram) to 1 pg, such as 1 pg to 200ng or 1 ng to 200 ng of cfDNA and / or wherein the cfDNA sample comprises less than 10000, less than 1000, less than 300 or less than 100 haploid human genome equivalents. In some embodiments, the sample data further comprises a tumour fraction (TF) for the cell-free DNA (cfDNA) obtained from the liquid sample. The sample data may comprise a TF of 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 sample data comprises a TF of no more than about 80%, no more than about 75%, or no more than about 70%. In some embodiments, the sample data comprises a TF of from about 1 % to about 75%.

[0209] As used herein, the term “tumour fraction (TF)”, will be understood to refer to the proportion of the cell- free DNA which is DNA originating from cancer.

[0210] The present inventors have advantageously found that while determination of the primary cancer of the subject based on the output from the machine learning model can be achieved from a sample data comprising very low TFs (for example, a TF of more than 0 but less than about 2%), the proportion of primary cancers determined to be an unknown type classification in this group were greater than from sample data comprising higher TFs.

[0211] In some embodiments, the sample data comprises a TF of at least about 3%. The inventors have advantageously found that a TF of at least about 3% reduces the risk of a primary cancer being classified as an unclassified and hence unknown primary cancer.

[0212] In some embodiments, the sample data further comprises a mutation profile. In some embodiments, the training data further comprises mutation profiles from at least two primary malignant tumour types. In some embodiments, the training data further comprises mutation profiles from at least one non-cancer control.

[0213] The mutation profile may comprise the mutation status of a plurality of genes.

[0214] Various genes are already known to be mutated in certain cancers. Therefore, the skilled person will readily be able to select the genes. For example, the genes may include, but not necessarily be limited to ALK, APC, AR, ARID1A, ATM, ATRX, GRIN2A, KDR, KMT2C, KMT2D, KRAS, LRP1 B, MED12, NF1 , NOTCH1 , PIK3CA, RB1 , SMAD4, SPEN, TP53, BRAF, BRCA1 / 2, EGFR and IDH1 .

[0215] In some embodiments, the genes comprise PIK3CA, BRAF and IDH1 .

[0216] In other embodiments, the genes are selected from ALK, APC, AR, ARID1 A, ATM, ATRX, GRIN2A, KDR, KMT2C, KMT2D, KRAS, LRP1 B, MED12, NF1 , NOTCH1 , PIK3CA, RB1 , SMAD4, SPEN and TP53, For example, the genes may comprise ALK, APC, AR, ARID1 A, ATM, ATRX, GRIN2A, KDR, KMT2C, KMT2D, KRAS, LRP1 B, MED12, NF1 , NOTCH1 , PIK3CA, RB1 , SMAD4, SPEN and TP53.

[0217] Mutation status may comprise a mutated or non-mutated status. A mutated status may comprise at least one non-synonymous mutation. A mutated status may comprise a frame shift deletion, frame shift insertion, splice site, translation start site, nonsense mutation, nonstop mutation, in frame deletion and / or missense mutation.

[0218] In some embodiments, the sample data further comprises a somatic copy number alteration (SCNA) profile. As the skilled person will appreciate, somatic copy number alteration is known to occur in certain cancers. By “somatic copy number alteration”, this will be understood to refer to an altered dosage of certain genes, chromosome arms or whole chromosomes, relative to a non-cancer control. SCNA profiles for various tumour samples are publicly available and known for example, from the Cancer Genome Atlas and the International Cancer Genome Consortium. In some embodiments, the training data further comprises SCNA profiles from at least two primary malignant tumour types.

[0219] In some embodiments, the sample data further comprises a fragmentomic profile. As the skilled person will be aware, fragmentomics is a term used to describe the study of the fragment size of cfDNA in subjects with cancer. In particular, recent research has indicated that the fragment size of cfDNA in cancer patients is much more varied than the fragment size of cfDNA in non-cancer control subjects. Thus, in some embodiments, the training data further comprises fragmentomic profiles from at least two primary malignant tumour types. In some embodiments, the training data further comprises fragmentomic profiles from at least one non-cancer control.

[0220] Methods of determining the primary cancer of a subject, methods of selecting a cancer treatment and methods of treatment

[0221] According to a second aspect, the present invention provides a method of determining the primary cancer of a subject having a malignancy of unknown origin (MUO), the method comprising: i) analysing cell-free DNA (cfDNA) obtained from a liquid sample from the subject in order to obtain sequence data comprising a region methylation signature (“sample data”); and ii) performing the method of the first aspect on the sample data to determine the primary cancer of the subject.

[0222] In some embodiments, analysing cfDNA obtained from a liquid sample comprises performing methylation sequencing of a plurality of genomic regions of the cfDNA. The plurality of genomic regions may be as defined above in relation to the first aspect.

[0223] In some embodiments, methylation sequencing comprises non-destructive methylation sequencing, for example a non-destructive enzyme-based conversion method to differentiate unmethylated cytosine from 5-methylcytosine and / or 5-hydroxymethylcytosine. This may otherwise be referred to as enzymatic methyl sequencing (EM-seq). In some embodiments, methylation sequencing comprises a step of bisulphite conversion (sometimes known as “bisulphite sequencing”). In some embodiments, methylation sequencing comprises direct measurement of base modifications, e.g., on a single molecule sequencing platform, such as via nanopore sequencing platforms or PacBio sequencing platforms.

[0224] In some embodiments, methylation sequencing comprises methylated DNA immunoprecipitation sequencing (MeDIP-seq). As the skilled person will appreciate, MeDIP-seq utilises a 5-methylcytosine monoclonal antibody to capture methylated fragments from single-stranded DNA. In some embodiments, methylation sequencing comprises methyl-binding-domain-seq (MBD-seq). MBD- seq, as used herein, will be understood to refer to the capture of double-stranded methylated DNA fragments using a methyl-binding-domain from MECP2.

[0225] In preferred embodiments, methylation sequencing comprises T7-MBD-seq, as described in Chemi et al. 2022, which is herein incorporated by reference in its entirety. Briefly, T7-MBD-seq comprises barcoding and pooling of fragment DNA, followed by incubation with MBD2. Next-generation sequencing (NGS) libraries can then be generated for methylated enriched fractions.

[0226] Thus, in some embodiments, analysing cfDNA comprises performing methylation sequencing and performing next-generation sequencing. Preferably, analysing cfDNA comprises performing T7-MBD-seq and performing next-generation sequencing. The next-generation sequencing may comprise Illumina sequencing.

[0227] The cfDNA may be subjected to amplification and / or library preparation prior to methylation sequencing. In some embodiments, the library preparation is single stranded DNA (ssDNA) library preparation. In some embodiments, the cfDNA may be fragmented prior to methylation sequencing.

[0228] Step i) may comprise analysing cell-free DNA (cfDNA) obtained from a liquid sample from the subject in order to obtain sequence data comprising a region methylation signature (“sample data”) and a mutation profile. The mutation profile may be as defined above in relation to the first aspect. In such embodiments, analysing cfDNA may further comprise analysing the mutation status of at least 10, at least 20, at least 50, at least 100 or at least 500 genes.

[0229] In some embodiments, step i) comprises analysing cell-free DNA (cfDNA) obtained from a liquid sample from the subject in order to obtain sequence data comprising a region methylation signature (“sample data”) and a tumour fraction (TF). The tumour fraction may be as defined above in relation to the first aspect. In such embodiments, analysing cfDNA may further comprising performing tumour fraction analysis, such as, for example, using ichorCNA.

[0230] The sequence data may further comprise a fragmentomic profile and / or an SCNA profile, either of which may be as defined above in relation to the first aspect. In such embodiments, analysing cfDNA may further comprise considering the size distribution of sequenced DNA fragments and / or the genomic context at which the DNA fragments have been cut.

[0231] Considering the size distribution of sequenced DNA fragments may comprise filtering in silico sequenced DNA fragments by their length in silico to tumour enriched sizes. The skilled person will be aware of suitable tumour enriched sizes. Without wishing to be bound by theory, the present inventors believe that sequenced DNA fragments filtered to tumour enriched sizes may comprise more tumour DNA and hence an DNA methylation status of genomic region(s) which have improved accuracy for indicating primary malignant tumour type.

[0232] Therefore, in some embodiments, the sequence data is derived from DNA fragments pre-filtered from the cfDNA by fragmentomics to tumour-enriched sizes.

[0233] According to a third aspect, the present invention provides a method of selecting a cancer treatment for a subject having a malignancy of unknown origin (MUO), the method comprising: i) performing the method of the first aspect to determine the likely primary cancer of the subject; and ii) selecting an anti-cancer treatment that targets the likely primary cancer.

[0234] According to a further aspect, the present invention provides a method of treating a subject having malignancy of unknown origin (MUO), the method comprising: i) performing the method of the first aspect to determine the likely primary cancer of the subject; and ii) administering an anti-cancer treatment that targets the likely primary cancer.

[0235] As used herein "treatment" refers to reducing, alleviating or eliminating one or more symptoms of the cancer that is being treated, relative to the symptoms prior to treatment.

[0236] The subject may be mammalian (e.g. a human, a non-human primate, a cat, dog, horse, donkey, sheep, pig, goat, cow, mouse, rat, rabbit or guinea pig). The subject is preferably a human. The subject may be a human adult (at least 18 years old). In some embodiments, the subject is a human child (less than 18 years old).

[0237] The anti-cancer treatment may comprise surgery, radiotherapy, chemotherapy, targeted therapy or immunotherapy. In some embodiments, the anti-cancer treatment comprises immunotherapy. Various immunotherapies will be known to the skilled person and commercially available. In some embodiments, the immunotherapy comprises pembrolizumab, atezolizumab, nivolumab, ipilimumab or a combination thereof.

[0238] The subject may have been pre-treated with a chemotherapeutic agent.

[0239] Administration of the anti-cancer treatment to the subject may result in a decrease in primary cancer tumour size of about 10%, about 20%, about 30%, about 40%, about 50%, about 60%, about 70%, about 80%, about 90%, or even about 100%, when compared to an untreated tumour.

[0240] The anti-cancer treatment may be administered using any suitable route. For example the anti-cancer treatment may be administered subcutaneously, intranasally, orally, topically, intraperitoneally or intravenously. In some embodiments, the anti-cancer treatment is intraperitoneally administered.

[0241] In embodiments wherein the anti-cancer treatment comprises a population of genetically engineered cells / immune cells, the number of the cells administered to the subject should take into account the route of administration, the primary cancer being treated, the weight of the subject and / or the age of the subject. In general, from about 1 x 106to about 1 x 1011genetically engineered cells / immune cells are administered to the subject. In some embodiments, from about 1 x 107to about 1 x 1010genetically engineered cells / immune cells, or from about 1 x 108to about 1 x 109genetically engineered cells / immune cells are administered to the subject. Systems, computer readable media and kits

[0242] According to another aspect, the present invention provides a system comprising: iii) a processor; and iv) a computer readable medium comprising instructions that, when executed by the processor, cause the processor to perform the steps of the method of the first aspect.

[0243] Preferably, the system comprises an Illumina sequencing platform system. An Illumina sequencing platform system will be understood to refer to a system which comprises instructions that, when executed by the processor, cause the processor to perform the steps of Illumina sequencing.

[0244] Preferably, the system comprises a computer readable medium comprising instructions that, when executed by the processor, cause the processor to: a) analyse cell-free DNA (cfDNA) obtained from a liquid sample from the subject in order to obtain sequence data comprising a region methylation signature; and b) perform the steps of the method of the first aspect.

[0245] Step a) may be as defined in relation to any of the previous aspects. For example, analysing cfDNA may comprise performing methylation sequencing and optionally performing next generation sequencing. The next generation sequencing may comprise Illumina sequencing.

[0246] Also provided by the present invention is one or more computer readable media comprising instructions that, when executed by one or more processors, cause the one or more processors to performs the steps of the method of the first aspect.

[0247] ***

[0248] The features disclosed in the foregoing description, or in the following claims, or in the accompanying drawings, expressed in their specific forms or in terms of a means for performing the disclosed function, or a method or process for obtaining the disclosed results, as appropriate, may, separately, or in any combination of such features, be utilised for realising the invention in diverse forms thereof.

[0249] 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 when given this disclosure. Accordingly, the exemplary embodiments of the invention set forth above are considered to be illustrative and not limiting. Various changes to the described embodiments may be made without departing from the spirit and scope of the invention.

[0250] For the avoidance of any doubt, any theoretical explanations provided herein are provided for the purposes of improving the understanding of a reader. The inventors do not wish to be bound by any of these theoretical explanations.

[0251] Any section headings used herein are for organizational purposes only and are not to be construed as limiting the subject matter described. Throughout this specification, including the claims which follow, unless the context requires otherwise, the word “comprise” and “include”, and variations such as “comprises”, “comprising”, and “including” will be understood to imply the inclusion of a stated integer or step or group of integers or steps but not the exclusion of any other integer or step or group of integers or steps.

[0252] It must be noted that, as used in the specification and the appended claims, the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. Ranges may be expressed herein as from “about” one particular value, and / or to “about” another particular value. When such a range is expressed, another embodiment includes from the one particular value and / or to the other particular value. Similarly, when values are expressed as approximations, by the use of the antecedent “about,” it will be understood that the particular value forms another embodiment. The term “about” in relation to a numerical value is optional and means for example + / - 10%.

[0253] Examples

[0254] INTRODUCTION

[0255] We previously described a robust and sensitive whole genome circulating-free DNA (cfDNA) methylation profiling workflow (T7-MBD-seq) that detects circulating tumour DNA (ctDNA) from patients with early stage small cell lung cancer (SCLC) and discriminates molecular subtypes (Chemi et al., 2022). In addition, several cancer early detection studies have demonstrated cfDNA methylation patterns predict tissue of origin (TOO) with high accuracy (Moss et al., 2018, Liu et al., 2020 and Klein et al., 2018). Here we generated and tested a highly-accurate TOO classifier derived from methylation profiles and applied it to a pilot CUP cohort in this proof-of-concept study. We demonstrate potential utility from a single blood draw of cfDNA analysis to facilitate diagnosis and treatment stratification, combining mutation detection with TOO predictions.

[0256] MATERIALS AND METHODS

[0257] Data availability

[0258] T7-MBD-seq data, shallow whole genome sequencing and targeted NGS data that support the findings of this study have been deposited in the European Genome-Phenome Archive (EGA) under the accession number EGAS00001007445. Pre-normalised TCGA data was downloaded from 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). Patient recruitment and sample collection

[0259] Patients with cancer were recruited through the TARGET (Tumour Characterisation to Guide Experimental Targeted Therapy) trial (ethically approved study REC reference: 15 / NW / 0078) (Rothwell et al., 2019) and the Manchester Cancer Research Centre CUP Biobank (ethic approval number 18_ALCO_01) at The Christie NHS Foundation Trust, a UK tertiary cancer centre. Non-cancer-control (NCC) samples were collected from three sources: 1 . The Community Lung Health Study (ethically approved study REC reference: 17 / LO415); 2. The University of Manchester healthy normal 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).

[0260] CUP clinical, radiological, and pathological review and data collection

[0261] Clinical data for the entire CUP cohort was retrospectively obtained and anonymised. All patients were discussed at a CUP dedicated Multi-Disciplinary Team (MDT) meeting within a tertiary cancer centre (The Christie NHS Foundation Trust) in accordance with National Institute for Health and Care Excellence (NICE) guidelines (NICE, 2010). Patients therefore had histopathological and radiological review to confirm the initial diagnosis of CUP. All clinical, pathological, and radiological investigations were reviewed alongside any MDT meetings. Subseguent primary tumour diagnosis status was determined as follows: A ‘Primary tumour diagnosis’ was confirmed if at any point from recruitment to death or data lock (August 2022) a primary tumour diagnosis was documented prospectively and the patient had subseguent treatment based on that diagnosis. A ‘highly suspected tumour type’ was categorised for patients where a single primary tumour diagnosis was suspected either prospectively or retrospectively, but patient was not treated based on this suspicion as uncertainty remained. A ‘differential diagnosis’ category comprised patients where clinicopathological features could narrow potential diagnosis to two or more different tumour types (Figure 1).

[0262] Blood sample collection

[0263] Blood samples were collected in up to 4 x 10 mL Cell-Free DNA BCT tubes (Streck, Omaha, NE), for cfDNA analysis. Plasma was separated from whole blood within 4 hours of blood draw by two seguential centrifugations (2,000 g, 10 min) and stored at -80 °C before cfDNA processing. In addition, 1 x 10 ML BD Vacutainer K2 ethylenediaminetetraacetic acid (K2EDTA) sample was taken for germline DNA extraction according to study protocols.

[0264] Circulating cell-free DNA extraction and guantification cfDNA was isolated from up to 20 mL of plasma by using one of three isolation technigues according to the manufacturer’s instructions: 1. QIAmp MinElute ccfDNA MIDI Kit (Qiagen, catalog number 55284) 2. Using the QIAsymphony with the Circulating DNA Kit (Qiagen, catalog number 1091063); or 3. QIAmp Circulating Nucleic Acid Kit (Qiagen, catalog number 55154). cfDNA yields were guantified by using the TagMan RNase P Detection Reagents Kit (Life Technologies, catalog number 4316831). T7-MBD-seq library preparation and next generation sequencing (NGS)

[0265] For all cfDNA samples between 1 -35 ng of cfDNA was processed via the T7-MBDSeq method as previously described (Chemi et al., 2022) and libraries were paired-end sequenced on an Illumina NextSeq 500 or NovaSeq 6000. For each sample both a methylation enriched fraction and non-enriched fraction were sequenced for genome-wide methylation profiling and copy number analysis, respectively.

[0266] T7-MBD-seq read alignment

[0267] A Nextflow (v22.04.5) DSL2 pipeline, built using the tools and modules provided by the nf-core community organisation (Ewels et al. 2020) was used to process FASTQ files and produce QSEA objects as detailed below.

[0268] All reads were trimmed to have the same initial length of 91 / 61 basepairs (bp) for R1 s / R2s respectively (including the 26bp T7-MBD-seq construct at the start of R1), the unique molecular identifier (UMI) extracted using umitools (Smith et al., 2017, v1 .1 ,2) and the samples demultiplexed and adapter-trimmed using cutadapt (Martin et al., 2011 , v3.4). Reads were aligned to the GRCh38 reference genome using bwa mem (Li, 2013, vO.7.17), deduplicated by the combination of R1 start position and UMI using umitools (Smith et al., 2017, v1 .1 .2) and mate quality scores assigned using samtools fixmate (Li et al., 2009, v1 .15.1), to produce final bam files.

[0269] Methylation Enrichment Analysis

[0270] The QSEA package (Lienhard et al., 2017) (v1.16) was used to analyse bam files, with a custom R (v4.2.0) package (MESA, Methylation Enrichment Sequencing Analysis, vO.2.1 , available from www.github.com / crukmi / mesa) built to extend QSEA. The genome was tiled into 300 bp non-overlapping windows, with the removal of over-represented windows. Over-represented windows were identified from analysis of 168 non-enriched NCC samples if the number of fragment counts within that window were in the top 0.1 % of fragment counts, as well as adjacent windows with counts in the top 1 %, following the method used by the ENCODE consortium (Amemiya et al. 2019). Here and elsewhere a fragment represents the genomic position within the two paired ends of the sequencing read.

[0271] Fragments were filtered to only paired reads where either end of the pair mapped with a Mapping Quality (MAPQ) score of at least 10, had a fragment length between 90-1000 bp and a distance along the reference genome of at least 30 bp. Fragments were then uniquely assigned into windows according to the location of their midpoint. For use within QSEA, Copy Number Variations (CNV) were calculated for each sample from the non-enriched sample, using HMMcopy (v1 .32 Shah et al., 2006) with base parameters over 1 Mbp windows. Normalised reads per million (NRPM) were generated using the CNV and the number of valid fragments in the sample, without applying trimmed mean of M values (TMM) normalisation. Beta values (a scaled measure of methylation between 0 and 1) for each window in each sample were calculated within QSEA using the “blind calibration” method (Lienhard et al., 2017).

[0272] IchorCNA

[0273] An estimate of Tumour Fraction (TF) for each sample was made using the non-enriched input cfDNA samples processed through IchorCNA (Adalsteinsson et al., 2017) (v0.32). The 79 NCC cfDNA samples used in generation of CUPiD were used as a panel of non-cancer samples and a 1 Mbp window size applied without estimating subclonal populations. Estimated TF below 0.03 are considered below the limit of detection (Adalsteinsson et al., 2017).

[0274] Quality Controls

[0275] NGSCheckMate (Lee et al., 2017)(v1 .0.0) was applied to verify that all samples from the sample individual matched as expected in the tool output. All four modalities for each patient (targeted sequencing of cfDNA and germline, T7-MBD-seq methylated and unmethylated fractions) were checked, where available.

[0276] To calculate the relative enrichment scores (relH) for the T7-MBD-seq samples, the method of the MEDIPS R package (Lienhard et al., 2014)(v1 .42) was used. This calculates the total density of CGs contained within the mapped DNA positions on the reference sequence and divides by the total density of CGs across the whole genome.

[0277] An additional QC metric for adequate methylation capture was calculated, termed the ‘hyperstable fraction’. Using 805 windows that correspond to CpG sites shown to be consistently hypermethylated (Edgar et al. 2014) in methylation array data from cancer and non-cancer samples, the fraction of these windows with a beta value of 0.8 or above was calculated. This metric is a combination of the number of valid fragments and the global enrichment profile. For the validation set, samples with relH below 2.5 or hyperstable fraction below 0.4 are excluded. This process removed 5 patient samples and 3 non-cancer control samples from the test set, and none of the samples from the CUP cohort.

[0278] Converting methylation array data to qseaSets

[0279] A pre-processed table of beta values from the TCGA Pan-Cancer methylation array dataset was downloaded from 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 consists of 9,639 arrays (including 721 adjacent normal tissues) across 33 tumour types.

[0280] Additional array data for the cholangiocarcinoma were obtained due to underrepresentation in the TCGA dataset. Using GEOquery (v2.66.0, Davis and Meltzer 2007) we downloaded additional pre-processed beta-values from the Gene Expression Omnibus (accession numbers GSE89803 (Jusakul et al. 2017), GSE32079 (Wang et al. 2013), GSE49656 (Chan-On et al. 2013), resulting in 256 cholangiocarcinoma arrays in total.

[0281] Grouping of tumour samples

[0282] The TCGA Pan-Cancer methylation array dataset was filtered to exclude samples that were categorised as recurrent, redacted, metastatic, additional tumour or adjacent normal tissue (apart from liver, see below); for pancreatic ductal adenocarcinoma (PDAC) we removed 28 samples potentially misclassified, for reasons such as extremely low tumour content or being a different type of pancreatic tumour than PDAC [Peran et al. 2018], This results in 8797 tumour samples and 49 normal liver tissue samples.

[0283] Arrays from all of the 33 TCGA categories (and the additional cholangiocarcinoma samples from GEO), were re-grouped into classes distinguished by both anatomical location and histological subtype (Table 3). This enabled the classifier to be trained on more histologically distinct classes and combined classes with low numbers of samples but high similarity. For example, the TCGA oesophageal carcinoma class (ESCA) was separated by histological subtype: the oesophagus adenocarcinoma (n=88) samples were combined into an upper gastro-intestinal (UpperGI) class with the stomach adenocarcinomas (STAD, n=393), while the oesophagus squamous cell carcinoma (n=96) was combined with the head and neck squamous cell carcinomas (HNSC, n=525) to form an upper squamous class (UpperSq). Where mixed histological tumour types existed within a class, these were excluded from the training: the 7 liver hepatocellular carcinoma arrays marked as “Hepatocholangiocarcinoma (Mixed)” were excluded, as were the 7 cervical & endocervical arrays denoted as “Adenosquamous”. Table 3 summarises the classes and tumour subtypes grouped within them.

[0284] This process resulted in 30 total classes, including a non-cancer class, which was included to give the classifier building process a neutral class to assign samples with low tumour fraction, rather than forcing a potentially incorrect prediction.

[0285] We then converted methylation array probe level data to 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 an average enrichment profile typical of our T7-MBD- Seq method to generate a lookup table that determines how many counts are required for each beta value and each CG density and can be used to convert from array beta values to estimated T7-MBD-seq counts. The maximum beta value was used where multiple probes lie within a single window. The resulting qseaSet estimates the counts we might expect from performing T7-MBD-seq on the sample, in the windows overlapping the array probes.

[0286] Uniform manifold approximation and projection

[0287] A Uniform manifold approximation and projection (UMAP, Mclnnes et al. 2018) was calculated on the converted tumour arrays, using the uwot package (v0.1 .14, Melville et al., 2022) with parameters n_neighbors = 15, min_dist = 1 .

[0288] Classifier building

[0289] To train the classifier, in silico synthetic mixture “qseaSets” were generated by mixing processed fragment counts between samples, either between an array sample (converted into qseaSets as detailed above) with an NCC cfDNA T7-MBD-seq sample at proportions between 1 - 10%, or between mixtures of two NCC T7-MBD-seq samples at proportions between 15-50%, all at varying numbers of fragments (between 1 million and either the number of NCC fragments or 10 million, whichever was lower). This was performed mixing every array sample (or NCC) once with each NCC, with a random proportion and random number of fragments. Early iterations of classifier development revealed some NCC mixtures were predicted to be liver cancer. It was hypothesised this was due to potential high levels of normal liver tissue signal within the normal cfDNA component for these samples. To overcome this, we included 49 adjacent normal liver arrays from TCGA into the NCC mixture sets fortraining, these mixtures were assigned the NCC class. Differentially Methylated Regions (DMRs) were calculated pairwise between classes using the QSEA package (Lienhard et al., 2017) on the converted array qseaSets at 100% tumour fraction. A false discovery rate (FDR) of 0.001 was applied. The difference between the average beta values for each class in the comparison, Ap, was calculated and used to sort DMRs by effect magnitude. A set of DMRs were generated between each class and the top 250 DMRs between each pairwise comparison were selected. After reduction of DMRs occurring in multiple comparisons, 22,179 distinct windows were taken forward in classifier development.

[0290] Sampling up to 10,000 mixture qseaSets for each class resulted in 276,108 unique samples in total for classifier training, resulting in an approximately equal class distribution (3,460-10,000 samples per group, depending on the number of arrays available). For each sample, the number of normalised reads per million (NRPM) were calculated for each of the 22,179 regions; this was used as the input data for each classifier. An ensemble comprising of 100 classifiers was then built, with each individual classifier including mixtures built from 80% of the array samples and 80% of the NCC samples. These individual classifiers were built as Extreme Gradient Boosting Trees (Chen and Guestrin 2016) (xgboost R package, v1 .6.0.1) within the R tidymodels (v0.2.0) framework, with default parameters except trees = 200, sample_size = 0.5, mtry = 2135. These parameter choices resulted in 200 sequential trees built, with each tree using a randomly selected 50% of the mixtures (within the 80% stratification) and 10% of the regions. These parameters provide a large amount of variation across the population of trees in each classifier, as well as the variation between each individual classifier based on the mixtures it is exposed to.

[0291] Model Performance

[0292] Each individual sub-classifier was tested on the remaining mixture sets comprising of samples not seen by that sub-classifier during model training (comprising 4% (20% * 20%) of the total mixture sets). Applying a sub-classifier to a held-out mixture set resulted in a prediction score for each of the 30 classes. These class-specific prediction scores for all the held-out mixture sets were compared against the ground truth to determine a one-vs-all area under the receiver operating curve (AUROC) value for each sub-classifier. To do this, we used the multiclass roc_auc function in the R package yardstick [v1 .0.0, Kuhn and Vaughan 2022], which is based on the method of Hand and Till: A Simple Generalisation of the Area Under the ROC Curve for Multiple Class Classification Problems. Machine Learning 45, 171-186 (2001).

[0293] For the ensemble classifier performance, we took each individual mixture set and applied all the subclassifiers that did not use either half of the mixture fortraining. We then calculated the mean prediction score for each of the 30 classes (averaging over the relevant sub-classifier outputs). An overall AUROC value for the ensemble was again calculated using the multiclass Hand-Till method.

[0294] Application of CUPiD to cfDNA samples processed through T7-MBD-Seq

[0295] The ensemble of trained classifiers (CUPiD) was then applied to the 143 cfDNA samples from patients with known tumour types, 27 NCC samples not used in the generation of CUPiD and the 41 samples from patients with CUP. The mean of each prediction value over the 100 individual classifiers was used as the final prediction score for each of the 30 classes. When applying CUPiD, a mean prediction score above 0.5 for a single class was required for a prediction of that class. This threshold ensures that the prediction value for the assigned class was higher than the remaining classes combined, as they sum to one. An unclassified prediction was reported where the mean prediction values were <0.5 or the NCC class was predicted.

[0296] Targeted library preparation and sequencing cfDNA and germline DNA from all CUP samples were processed in accordance with TARGET trial laboratory protocol as previously described (Rothwell et al., 2019, TARGET patients up to TAR00286) or by the following updated method (TAR00287 onward and Biobank samples). DNA repair and dA-tailing were performed using the NEBNext® Ultra™ II End Repair / dA-tailing Module (New England Biolabs, catalog number E7546L). Adapter ligation and indexing was carried out using KAPA HyperPrep Kit (Roche, catalog number 07962355001) with NEBNext® Multiplex Oligos for Illumina (New England Biolabs, catalog number E7335L). Targeted NGS for whole genome libraries from cfDNA and corresponding germline DNA was carried out using SureSelect Custom DNA Target Enrichment Probes (Agilent, catalog number 5190- 4822). Target enrichment of 0.5 - 1 .0 pg of each DNA library (paired cfDNA and gDNA libraries from the same patient pooled per pull down) was performed using SureSelect XT HS Target Enrichment System (Agilent, catalog number G9703A) with a 641 gene hybridisation panel. Captured libraries were amplified using KAPA HiFi HotStart PCR Kits (Roche, catalog number 07958897001) and quantified using the KAPA Library Quantification qPCR Kit (Roche, catalog number 07960140001). Libraries were paired end sequenced at 2 x 150 bps on a NextSeq 500 or NovaSeq 6000 (Illumina).

[0297] Targeted library alignment and mutation calling

[0298] FASTQ files were aligned to GRCh38 using bwa mem (Li, H 2013, vO.7.17) and deduplicated using samtools (Li et al., 2009, v1 .9). Mutations were called using Mutect2 (v4.2.5.0, following GATK bestpractices (Van der Auwera and O’Connor, 2020) with default parameters apart from a f-score beta of 5, log 10 odds threshold of 1 .0 and a minimum variant allele fraction of 1 %) as well as QIAGEN CLC Genomics Workbench (v20.0.2 build 200002), calling variants between the cfDNA sample and a matched germline control derived from whole blood. Those mutations called by both tools were denoted as high confidence, and annotated by VEP (v193.1 , McLaren et al., 2016) and OncoKB (v3.17, Chakravarty et al., 2017). VCF files were converted to MAF files using vcf2maf (v1 .6.21 , Kandoth, C. mskcc / vcf2maf: vcf2maf v1 .6.19. (2020)) and restricted to those whose Variant_Classification field was one of Frame_Shift_Del, Frame_Shift_lns, Splice_Site, Translation_Start_Site, Nonsense_Mutation, Nonstop_Mutation, ln_Frame_Del, ln_Frame_lns or Missense_Mutation, as well as filtering out any mutations present in the gnomAD database ( Karczewski, et al., 2020, v2.1 .1) at a population frequency above 1 %. The oncoplot was generated using maftools (v2.8.05, Mayakonda et al. 2018 ), selecting for those genes noted as being oncogenic by OncoKB (v3.17). Mutation analysis for oncogenicity and actionability

[0299] Each high confidence genomic alteration that passed mutation filtering steps was annotated using OncoKB to determine likelihood of being oncogenic and potential actionability. Only alterations actionable to Level 3 evidence and above were reported.

[0300] Annotation of mutated genes based on enrichment for genomic alterations in individual tumour types Genes were annotated according to tumour type enrichment using a previously described set of genes that show statistically significant enrichment for genomic alterations in individual tumour types when compared to all other cancer, based on AACR Project GENIE mutation data (Posner et al., 2022). For patients which were ‘clinically resolved’ or ‘highly suspected’, these were considered as supported if at least one gene was noted as statistically enriched in that tumour type. For patients with a ‘differential diagnosis’, mutations were considered supportive if only one of the potential differential diagnoses was present; those with mutations which are enriched for more than one potential diagnosis were not considered supportive.

[0301] EXAMPLE 1: Building a robust multi-class TOO classifier

[0302] To build a robust multi-class TOO classifier applicable to cfDNA samples, we had to address the significant challenge of high variability in ctDNA content within cfDNA (tumour fraction, TF) that results in predominant non-cancerous cfDNA signal dilution of tumour-specific signal, even in metastatic cancers. To overcome this without profiling thousands of cfDNA samples, we applied a bioinformatic approach to mimic cfDNA samples of varying TF (Chemi et al. 2022) (Figure 2). Firstly, we used publicly available DNA methylation data from tumour tissue using Infinium 450K methylation microarrays, mainly from The Cancer Genome Atlas (TCGA), representing 29 tumour classes (9,017 tumours (Hoadley et al. 2018), (Figure 3) and converted methylation beta values for each array probe into estimated T7-MBD-seg read counts. For controls, DNA methylation data was determined from the cfDNA of blood samples of 79 noncancer control (NCC) subjects. Then we created in silico mixtures by mixing the cancer class estimated counts with previously seguenced cfDNA samples from patients (Figure 2 and Methods). Each tumour class was present across a range of TFs (0.5-10%) and NCC clinical features (age, sex, ethnicity, smoking status, comorbidities). Preliminary classifier development demonstrated the normal liver tissue component present in NCC cfDNA could erroneously result in a liver cancer class prediction, this was remedied by the addition of non-cancerous liver tissue arrays (n=49) to the non-cancer class. In total 276,108 mixtures were created across 30 classes.

[0303] Tumour-specific genomic regions were calculated using differentially methylated regions (DMRs), comparing each of the 30 classes (100% tumour content) and selecting the 250 DMRs with greatest differences between each pairwise comparison (22,179 unigue regions, Figure 4). A Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction using these DMRs applied to 9,017 tumour samples, recapitulated class clustering seen using all regions (Figure 5). An ensemble classifier, termed CUPiD, was built comprising of 100 individual gradient-boosted tree sub-classifiers with each sub-classifier trained on in silico mixtures from 80% of the arrays and NCCs (Figure 2). The sub-classifiers were then tested on ‘held-out’ data sets comprised of mixtures of the 20% of the arrays and NCCs excluded from the training set. These sub-classifiers performed accurately across the 30 classes with a mean multi-class area under the receiver operator curve (AUROC) of 0.980 (standard deviation 0.00521 , Figure 6A). The resulting ensemble classifier, taking mean prediction score across each sub-classifier, had an AUROC of 0.984 (Figure 6B). A tumour prediction was called when the score for a single class was >0.5 ensuring the assigned class score was higher than all remaining classes combined. An unclassified prediction was reported where all scores were <0.5 or the non-cancer class was predicted.

[0304] EXAMPLE 2: Testing of CUPID on an independent test cohort with known primary cancer types

[0305] We tested CUPiD on an independent test cohort of 170 cfDNA samples, including 143 cancer patient samples from 13 different tumour types, and 27 NCC, profiled using T7-MBD-seq (Figure 7A-B). In the cancer cfDNA samples, CUPiD predicted a correct tumour type for 121 / 143 patients (overall sensitivity, 84.6%), an unclassified prediction for 18 / 143 patients (12.6%) and an incorrect prediction for 4 / 143 patients (2.8%) (Figure 8). None of the 27 NCC samples were predicted as a tumour class (Figure 8). In the 125 samples with a tumour prediction, the correct prediction was made in 121 / 125 of cases (TOO accuracy 96.8%). These results are comparable to other cfDNA TOO classifiers in development for early cancer detection, e.g., TOO accuracy using DNA methylation was recently reported between 87.0%- 90.2% in 1393 samples with a ‘cancer-like’ methylation signal across 21 different tumour classes (Klein et al., 2021).

[0306] The tumour fraction (TF) for each sample in our study was estimated using ichorCNA (Adalsteinsson et al. 2017) from shallow whole genome sequencing and varied by tumour type, ranging between 0-60.3% (Figure 9A). For 54 cancer cfDNA samples, the estimated TF was <3% limit of ichorCNA detection. A correct prediction was made in 37 of these cases (67.9%), 16 (30%) were unclassified and one had an incorrect prediction (Figure 9B). This demonstrates, despite low estimated TF, methylation profiling detects tumour signal and accurately predicts TOO; only 2 samples with estimated TF >3% were unclassified (Figure 9B).

[0307] For the 4 cases where CUPiD predicted an incorrect tumour type, 3 were patients with lung adenocarcinoma, of which 2 / 3 were predicted as lung cancer but squamous carcinoma and 1 / 3 breast cancer. The remaining incorrect prediction was cholangiocarcinoma predicted as pancreatic adenocarcinoma.

[0308] EXAMPLE 3: cfDNA mutation profiling of CUP patients

[0309] The profiling of CUP patients was then investigated in a 41 patient CUP pilot study. In particular, the study assessed the feasibility of cfDNA methylation and mutation profiling combined with TOO predictions.

[0310] Prior to profiling, most CUP cases had been categorised as adenocarcinomas (25 / 41 , 61 .0%) or poorly differentiated carcinomas (11 / 41 , 26.8%). At that stage, standard clinical investigations were unable to more specifically identify the primary cancer. Unsurprisingly, verifying TOO predictions is challenging given the intrinsic nature of CUP. Retrospectively, we reviewed clinical data including clinical history, pathology, radiology and discussions from CUP-dedicated multi-disciplinary team (MDT) meetings, including additional investigations after initial diagnosis. In 15 / 41 (36.6%) patients, a subsequent primary tumour diagnosis was made and patients received treatment for that tumour type (‘clinically resolved’).

[0311] For the patients who remained confirmed CUP (cCUP) throughout their cancer journey (n=26), 18 / 26 had tumour diagnoses suspected based on their clinical data; either one highly-suspected primary tumour diagnosis was made (n=11 , ‘highly suspected’) or multiple primary tumour diagnoses were suspected (n=7, ‘differential diagnosis’). A further 8 / 41 patients had no clinical or radiological clues for potential primary tumour diagnosis (‘no clinical suspicion’).

[0312] Initially, cfDNA mutational profiles in 41 patients with CUP were evaluated using a comprehensive 641 gene panel (see Methods). cfDNA mutation profiling was not successful in one patient. This patient appeared to have many mutations which are commonly found in genomic databases of population variation, suggesting the potential for contamination. However, cfDNA mutation profiling was successful in the remaining 40 patients. In these patients, we found 345 mutations across 33 patients (82.5% with at least one mutation). Per patient, the median number of mutations was 5 (range 0-77) and median variant allele frequency (VAF) was 10.4% (range 0-61.3), (Figure 10). OncoKB (Chakravarty et al. 2017) analysis predicted 60 alterations across 26 patients as oncogenic or likely oncogenic (Figure 10) with 7 / 26 patients harbouring alterations that are potentially actionable with level 3 or above evidence. This included 3 patients with PIK3CA mutations, 2 with non-V600E BRAF mutations and 1 IDH1 mutation, all targetable with FDA approved drugs in specified tumour types.

[0313] To determine if any identified alterations could inform tumour type in the CUP cohort, we considered gene alterations previously described as significantly enriched in particular tumour types (Posner et al. 2022) and asked whether these tumour type-enriched alterations (TTEAs) supported the clinical features. In no cases were TTEAs found that were supportive of only one tumour type, showing the limitations of mutation analysis. In 14 / 40 cases (35.0%) at least one TTEA was present and supportive (Figure 11). In 14 (35.0%) cases TTEAs were present but not supportive, or they were ambiguously supportive of several tumour types in cases of differential diagnosis. The remaining cases had no TTEAs (n=12) (Figure 11).

[0314] EXAMPLE 4: Application of CUPID to the CUP cohort to determine primary cancer cfDNA methylation profiling was successful across our entire cohort of 41 patients with CUP (41 / 41 passed QC). The estimated ichorCNA TF ranged between 0-53.4%, with 27 / 41 samples (65.8%) >3%, correlating well with the average VAF calculated through mutation analysis (r=0.74, p=<0.001), Figure 12). When applied to this cohort CUPiD yielded a tumour prediction in 32 / 41 of cases (78.0%) (Figure 13, Table 3). For 9 / 41 patients (22.0%), no prediction was made with 7 / 9 unclassified cases demonstrating no copy number changes (estimated TF <3%) and 5 / 7 of these patients also had no detectable mutations suggesting low ctDNA content (Figure 10, Figure 14). Of the 32 tumour predictions, 26 patients (81%) had predicted classes within 5 broad cancer categories (Figure 15). The commonest predictions were hepato-pancreatobiliary (7 / 32, 20% with 6 predicted cholangiocarcinomas) and female genital tract (6 / 32, 18.75%) (Figure 15). Although cholangiocarcinoma is rare, it is increasingly recognised within CUP cohorts and diagnostic biomarkers are limited (Saha et al. 2016, Conway et al. 2022)). Upper and lower gastrointestinal (4 / 32, 12.5%), lung (5 / 32, 15.2%) and urological cancers (4 / 32, 12.5%) were also frequently predicted by CUPiD. These tumour type predictions are consistent with historical data of primary tumour types found at autopsy in patients with CUP (Le Chevalier 1988, Mayordomo et al. 1993) and are commonly predicted in other large TOO studies based on tumour tissue profiling (Fizaki et al. 2019, Hainsworth et al. 2013, Lu et al. 2021 , Hayashi et al. 2019 and Moran et al. 2016). Interestingly, all of the CUPiD predicted tumour types have radically different treatment strategies compared to SOC chemotherapy for CUP, with all predictions except a single case for tumour types that warrant consideration for immunotherapy or targeted therapies as SOC treatment options in first- or second-line setting.

[0315] For the 33 patients that were ‘clinically resolved’ (n=15) or suspected diagnoses (n=18), 26 had a CUPiD tumour type prediction and 23 (88.5%) of these predictions aligned with confirmed primary tumour type or one of the suspected differential diagnoses (Figures 11 and 13). Three CUPiD predictions were inconsistent with clinical data of which 2 were ‘clinically resolved’ cases: the first, predicted as bladder cancer, was a rare yolk sac tumour not existing within the training data of the CUPiD classifier; the second, predicted as cholangiocarcioma, was a patient eventually determined to have gastric cancer; the final patient, predicted as an upper Gl malignancy, remained as cCUP with clinical features suspicious of a breast primary (Figure 13). CUPiD predictions were also made in 6 / 8 patients where no primary was suspected or confirmed, exemplifying the potential of TOO molecular profiling in the most uncertain cases.

[0316] The 15 patients with ‘clinically resolved’ primary tumours suffered a protracted period of diagnostic uncertainty and most had treatment with suboptimal empiric chemotherapy prior to primary tumour determination (Figure 16). The median time to diagnosis was 7.1 months (range 0.4-47.2 months) and an invasive repeat biopsy was performed in 6 patients to make a final diagnosis. With adequate TF, a liquid biopsy-based TOO classifier measured at suspected cancer diagnosis could have dramatically accelerated a confirmed diagnosis in a large proportion of patients and potentially negated need for repeat, invasive biopsies. The experimental turnaround time for CUPiD is currently less than three weeks with potential to reduce via assay automation.

[0317] EXAMPLE 5: Application of CUPID using fewer genomic regions

[0318] It was then decided to look at the effect of taking the CUPiD classifier and altering the set of DMRs used.

[0319] First, the variable importance of the genomic regions used in the original CUPiD classifier was considered. The top N of these genomic regions was then selected, for varying numbers of N, and we fit the same ensemble classifier (with the same in silico mixtures and parameters) with each reduced set of DMRs. Variable importance was generated using the vip R package, vO.4.1 .

[0320] To determine performance, two measures were considered. First, the area under the curve (AUC) on the held-out measures, both for the individual 100 sub classifiers and the performance of the averaged ensemble (although on average this is only across 4 sub classifiers rather than the 100 used on unseen data) was analysed. This includes held-out samples from all 30 classes, and so represents performance across the whole set of tumour types. The results are shown in Figure 17A.

[0321] The 170 cfDNA samples utilised in Example 2 were then considered. These comprised 143 cancer samples from 13 classes and 27 non-cancer controls, so do not comprise a full validation of all the classes, but are actual patient cfDNA samples. The models of Figure 17A were then applied to the 170 cfDNA samples in the test cohort. Figure 17B shows the results, which demonstrates how many genomic regions may be required for accurate predictors in this model.

[0322] These classifiers, on these two metrics, maximise performance at 5000 windows, compared to the larger 22179 total windows used in Example 4, suggesting that the removal of uninformative windows allows the classifiers to focus into the most important ones. It should be noted that these classifiers used knowledge of window importance as given by the larger CUPiD classifier, and therefore required the development of that first larger classifier.

[0323] EXAMPLE 6: Varying numbers of pairwise DMRs

[0324] Another way to vary the number of windows going into the classifier is to take the top N windows that are differentially methylated between each pairwise comparison. For CUPiD the top 250 were used in each comparison, which led to 22179 distinct windows in total, as many windows are in multiple comparisons.

[0325] The AUC was analysed for varying numbers of windows, with the results shown in Figure 18A. Figure 18A shows that the AUC steadily increases as the number of DMRs is increased. The models of Figure 18A were then applied to the 170 cfDNA samples in the test set. Figure 18B shows the results.

[0326] This analysis shows that the ROC AUC is still increasing as additional DMRs are added, but the performance on the cfDNA samples is marginally better with the top 25 pairwise DMRs than the top 250 pairwise DMRs. This may be due to the distribution of the cfDNA sample cohort that was considered here.

[0327] EXAMPLE 7 Building a classifier using different genomic regions to those used in CUPiD

[0328] A classifier was then built without using the 22,179 windows used in CUPiD. Pairwise DMRs were calculated in the same way as before, then the top N in each pairwise comparison were selected after filtering out the windows used in CUPiD. These are therefore at least the 251st DMR in each pairwise comparison. The AUC for classifiers with 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) are shown in Figure 19A, which shows that the AUC steadily increases as the number of DMRs is increased. These models were then applied to the 170 cfDNA samples in the test cohort, of which Figure 19B shows the results.

[0329] The performance of these classifiers was then compared to the performance of CUPiD. Performance of both of the AUC metrics and the cfDNA calls was lower for these classifiers than for the equivalent top N classifier using the CUPiD regions. The ensemble classifier AUC is shown in Figure 20A, while the performance on the actual cfDNA cohort is shown in Figure 20B.

[0330] EXAMPLE 8: CUPID v1.1. performance on an independent test cohort with increased number and variety of known primary cancer types

[0331] A robust multi-class TOO classifier was built as per Example 1 , except that the pairwise comparisons did not contain a non-cancer control class, other than a Normal Liver class. This resulted in 30 classes in total, these being 29 primary malignant tumour classes and one normal liver class. The training data still contained NCC data (in addition to primary malignant tumour data), with genomic regions identified using pairwise comparisons of tumour classes. However, the set of NCC cfDNA samples used were split in a different way between the training and test set to v1 . Some additional NCC samples were also used. Following pairwise comparison, genomic regions in the non-cancer control data having a median beta value > 0.25 were excluded from the training data, before selecting the top 250 regions up and down from each pairwise comparison. This classifier may otherwise be referred to as CUPiD v1 .1 .

[0332] Table 5 lists all genomic regions used in CUPiDvl .1 . Table 6 lists the 250 genomic regions of CUPiD v1 .1 . which were identified by the inventors as the most informative genomic regions for discriminating between primary malignant tumour types, as determined by variable importance score for pairwise comparison between different primary malignant tumour class types. 174 of the 250 genomic regions of Table 6 are also found in the top 250 genomic regions of the earlier CUPiD model (Table 2). These 174 specific genomic regions are presented in Table 7. Table 8 lists the 500 genomic regions of CUPiD v1 .1 . which were identified by the inventors as the most informative genomic regions for discriminating between primary malignant tumour types, as determined by variable importance score for pairwise comparison between different primary malignant tumour class types.

[0333] The performance of Cupido v1.1 was assessed. Figure 21 shows the proportion of samples that were predicted by CUPiD v1 .1 as correct, incorrect or no prediction when applied to the same validation cohorts of 143 cfDNA known cancer samples (“known” cohort) and 41 cfDNA CUP samples as used in Examples 2-4 (“CUP” cohort). An additional six skin cutaneous melanoma samples and 18 breast cancer samples (hormone positive or HER2 positive subtypes) were added to the original 143 “known” samples totalling 167 cfDNA known cancer samples (“Combo” cohort).

[0334] For the CUP cohort, clinical correlation informed the likelihood of a prediction being correct: A prediction was deemed correct if a subsequent primary tumour diagnosis was made or highly suspected clinically or that the predicted tumour type was one of the potential differential diagnoses made clinically. No suspicion represented the samples from the CUP cohort where there was no suspected primary or potential differential diagnosis made clinically. Predictions were made forthose samples but it was not possible to confirm if these predictions were correct or incorrect.

[0335] The known cohort had a sensitivity of 84% and a false negative tumour rate (FNTR) of 0.03. The combo 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).

[0336] EXAMPLE 9: Use of broader output classifications to determine likely primary cancer

[0337] It was next decided to use broader cancer categories as the primary malignant tumour classification in the output of CUPiD v1 .1 . Results are shown in Figure 22. This figure shows how using eight broad cancer categories as the possible output would be applied to the validation cohorts of 143 cfDNA known cancer samples .

[0338] To achieve this, training of the model (using the original number of classes) remained the same, with the model still outputting a prediction score for each class. To obtain the broader prediction score, the broad cancer categories were pre-defined. The prediction scores for each subclass were then added up to form an accumulated score within each broad category. If the accumulated score was >0.5 threshold, then this was considered to be a prediction forthat particular broad category.

[0339] For example, the broad class of lung was the result of summing the prediction scores for LUAD and LUSC. This can mean that for a sample with a LUAD score of 0.35 and a LUSC score of 0.25, it would have originally resulted in an unclassified prediction (all below 0.5), but summing them gives 0.6, so a prediction of the broad class of lung. Table 10, below, details how each specific cancer class was defined broadly.

[0340] Table 10: Broad cancer grouping of 29 tumour classes

[0341] The broad cancer groups are labelled at the top of each graph in Figure 22 A to H. The prediction scores for each of the cancer types per broad categories were accumulated. The dashed line represents the 0.5 prediction threshold. Only eight of the broad reporting categories are present in the graph as those were the categories for which we had test samples. Samples were classed as correct predictions if the accumulated cancer type prediction of 0.5 or above was within the correct broad category.

[0342] Figure 22I shows that in the known cancer cohort, two samples with an incorrect prediction when reporting separately were predicted correctly on the broad reporting approach. One LUAD sample (TAR00338) had an incorrect LUSC prediction score of 0.57 but was correctly in the broader lung class, and a cholangiocarcinoma (CHOL) sample (TAR00248) had an incorrect pancreatic adenocarcinoma (PAAD) prediction score of 0.99 but was correctly in the broader upper abdominal class. Additionally, two samples with no prediction when using the separate reporting method were correctly predicted at the broad cancer category level, one sample was predicted as a lung cancer (TAR00279) and the other predicted as an upper abdominal cancer (TAR00203).

[0343] This broader output classification increased the sensitivity of CUPiD compared to when the output could comprise any one or more of 29 primary malignant tumour classifications, each primary malignant tumour classification classifying a distinct anatomical and / or histological tumour subtype. In particular, the broader output classification increased the sensitivity of CUPiD from 84% to 87% in the known cohort (Figure 23). The broader output classification increased the sensitivity of CUPiD from 86% to 89% in the combo cohort, while it increased the sensitivity of CUPiD from 70% (23 / 33) to 76% (25 / 33) in the CUP cohort (Figure 23). The two incorrect broad predictions in the CUP cohort were both for samples where the primary tumour type was not confirmed clinically but narrowed to several differential clinical diagnoses and therefore true diagnosis remains uncertain. EXAMPLE 10: Benchmarking of the CuPiD v1.1 model

[0344] To benchmark the CUPiD v1 .1 model, the AUC (which may otherwise be referred to as AUROC) was analysed for varying numbers of windows / DMRs per pairwise comparison. The total number of DMRs in the model, depending on the number of DMRs per pairwise comparison, is shown in Table 11 below. Table 11 also shows the number of DMRs excluding repeated regions. This means the total number of unique DMRs in the model, as some of the DMRs were the top DMR (or in the top 2, 5, and so on) for multiple pairwise comparisons.

[0345] Table 11 : Total number of DMRs in the model, depending on number of DMRs per cancer type comparison.

[0346] Figure 24 shows the ensemble AUC for the held-out in silico mixture sets for each number of DMRs per pairwise comparison. This figure shows that the AUC steadily increases as the number of DMRs per pairwise comparison increases.

[0347] It was next decided to assess performance of the CUPiD v1 .1 model when considering varying numbers of windows / DMRs per pairwise comparison, as separated by tumour content (TC). Tumour content was separated into the groups of below 3%, between 3 and 5% and 5% and over. Figure 25 shows the ensemble AUC for the held-out in silico mixture sets for each number of DMRs per pairwise comparison. This demonstrates that increasing the number of DMRs per pairwise comparison is most beneficial for the AUC of the mixture sets with low tumour content. It also demonstrates that smaller numbers of DMRs per pairwise comparison are still effective, especially when tumour content is at least 3%.

[0348] We next looked at prediction performance when varying the number of DMRs per pairwise comparison. Figure 26 shows the proportions of correct, incorrect and no predictions for the held-out in silico mixture sets for each number of DMRs per pairwise comparison. The graph does not include NCC only mixture sets. These graphs show that AUC and CUPiD performance is still relatively high at only 1 DMR per pairwise comparison, and begins to plateau after 50 DMRs per pairwise comparison. Varying numbers of windows / DMRs in CUPiD v1 .1 were then tested on the known cfDNA samples used in earlier examples (Figure 27). Briefly, Known (n=143) were the predictions in the cohort of 143 cfDNA samples from known primary cancer samples and CUP (n=41) were the predictions in the cohort of 41 cfDNA CUP samples. The no suspicion proportions represent the samples that had no clinically suspected TOO. Predictions were made by CUPiD for these samples, but they cannot be classified as correct or incorrect. CUPiD still performed well when making predictions in the cohort of known samples, when using only 1 DMR per comparison. A plateau was formed after 50 DMRs per comparison. Increasing the number of DMRs caused incorrect predictions to become no predictions in the cohort of CUP samples, thereby demonstrating improved accuracy.

[0349] EXAMPLE 11: Comparison of top 250 ranked DMRs in CUPiD v1 versus CUPiD v1.1

[0350] The top 250 DMRs for each model (Table 2 for CUPiD v1 and Table 6 for CUPiD v1 .1), based on ranking, were compared. Figure 28 shows a plot of variable importance rankings of all DMRs in v1 against all DMRs in v1 .1 . This Figure shows that 174 of the top 250 DMRs in CUPiD v1 .1 were also found in the top 250 DMRs of CUPiD v1 . This Figure also shows that while the highest ranked DMR in CUPiD v1 .1 was not found in CUPiD v1 , the next four highest ranked DMRs (indeed the next nine highest) in CUPiD v1 .1 were also found in CUPiD v1 .

[0351] DISCUSSION

[0352] In summary, we have developed CUPiD, an accurate TOO liquid biopsy with encouraging sensitivity and accuracy in known tumour types and clinically consistent predictions for patients with CUP whose cfDNA contains adequate TF. As cfDNA mutation and methylation profiling can be performed from the same blood draw, this approach enables identification of potentially actionable alterations alongside TOO predictions to aid treatment stratification. Applying CUPiD to a pilot cohort of 41 CUP patients resulted in predicted tumour types in 32 / 41 , with all 32 patients having the potential to benefit from radically different, tumour specific treatment strategies compared to CUP SOC chemotherapy.

[0353] References

[0354] A number of publications are cited above in order to more fully describe and disclose the invention and the state of the art to which the invention pertains. Full citations for these references are provided below. The entirety of each of these references is incorporated herein.

[0355] Amemiya HM, Kundaje A, Boyle AP. The ENCODE Blacklist: Identification of Problematic Regions of the Genome. Sci Rep. 2019 Jun 27;9(1):9354. doi: 10.1038 / s41598-019-45839-z. PMID: 31249361 ; PMCID: PMC6597582. Adalsteinsson VA, Ha G, Freeman SS, Choudhury AD, Stover DG, Parsons HA, et al. Scalable whole- exome sequencing of cell-free DNA reveals high concordance with metastatic tumors. Nature Communications. 2017;8(1):1324.

[0356] 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.

[0357] Chan-On W, Nairismagi ML, Ong CK, Lim WK, Dima S, Pairojkul C, Lim KH, 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 distinct mutational patterns in liver fluke-related and non- infection-related bile duct cancers. Nat Genet. 2013 Dec;45(12):1474-8. doi: 10.1038 / ng.2806. Epub 2013 Nov 3. PMID: 24185513.

[0358] 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 and subtyping of small cell lung cancers. Nature Cancer, 3, 1260-1270.

[0359] Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '16). Association for Computing Machinery, New York, NY, USA, 785-794. https: / / doi.org / 10.1145 / 2939672.2939785

[0360] Conway A-M, Morris GC, Smith S, Manoharan P, Mitchell C, Backen A, et al. Intrahepatic cholangiocarcinoma hidden within cancer of unknown primary British Journal of Cancer. 2022;in press.

[0361] Davis S, Meltzer PS. GEOquery: a bridge between the Gene Expression Omnibus (GEO) and BioConductor. Bioinformatics. 2007 Jul 15;23(14):1846-7. doi: 10.1093 / bioinformatics / btm254. Epub 2007 May 12. PMID: 17496320.

[0362] Edgar, R., Tan, P. P. C., Portales-Casamar, E. & Pavlidis, P. 2014. Meta-analysis of human methylomes reveals stably methylated sequences surrounding CpG islands associated with high gene expression. Epigenetics & Chromatin, 7, 28.

[0363] Ewels, P.A., Peltzer, A., Fillinger, S. et al. The nf-core framework for community-curated bioinformatics pipelines. Nat Biotechnol 38, 276-278 (2020). https: / / doi.org / 10.1038 / s41587-020-0439-x

[0364] Fizazi K, Maillard A, Penel N, Baciarello G, Allouache D, Daugaard G, et al. A phase III trial of empiric chemotherapy with cisplatin and gemcitabine or systemic treatment tailored by molecular gene expression analysis in patients with carcinomas of an unknown primary (CUP) site (GEFCAPI 04). Annals of Oncology. 2019;30:v851-v. Hainsworth JD, Rubin MS, Spigel DR, Boccia RV, Raby S, Quinn R, et al. Molecular gene expression profiling to predict the tissue of origin and direct site-specific therapy in patients with carcinoma of unknown primary site: A prospective trial of the Sarah cannon research institute. 2013. p. 217-23.

[0365] Hand, D.J., Till, R.J. A Simple Generalisation of the Area Underthe ROC Curve for Multiple Class Classification Problems. Machine Learning 45, 171-186 (2001). https: / / doi.Org / 10.1023 / A:1010920819831

[0366] Hayashi H, Kurata T, Takiguchi Y, Arai M, Takeda K, Akiyoshi K, et al. Randomized phase II trial comparing site-specific treatment based on gene expression profiling with carboplatin and paclitaxel for patients with cancer of unknown primary site. Journal of Clinical Oncology. 2019;37(7):570-9.

[0367] 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 Tumors from 33 Types of Cancer. Cell. 2018;173(2):291-304.e6.

[0368] Jusakul, A., et al. Whole-Genome and Epigenomic Landscapes of Etiologically Distinct Subtypes of Cholangiocarcinoma. Cancer Discov 7, 1116-1135 (2017).

[0369] Karczewski, K.J., et al. The mutational constraint spectrum quantified from variation in 141 ,456 humans. Nature 581 , 434-443 (2020).

[0370] Klein EA, Hubbell E, Maddala T, Aravanis A, Beausang JF, Filippova D, et al. Development of a comprehensive cell-free DNA (cfDNA) assay for early detection of multiple tumor types: The Circulating Cell-free Genome Atlas (CCGA) study. Journal of Clinical Oncology. 2018;36(15_suppl):12021 -.

[0371] 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).

[0372] Kuhn, M. & Vaughan, D. yardstick: Tidy characterizations of model performance. R package version 1.0.0. (2022).

[0373] Le Chevalier T, Cvitkovic E, Caille P, Harvey J, Contesso G, Spielmann M, et al. Early metastatic cancer of unknown primary origin at presentation. A clinical study of 302 consecutive autopsied patients.

[0374] Archives of internal medicine. 1988;148(9):2035-9.

[0375] 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 Acids Res, 45, e103.

[0376] Li, H. Aligning sequence reads, clone sequences and assembly contigs with BWA-MEM. arXiv preprint arXiv:1303.3997 (2013).

[0377] Li, H., et al. The Sequence Alignment / Map format and SAMtools. Bioinformatics (Oxford, England) 25, 2078-2079 (2009).

[0378] Lienhard, M., Grasse, S., Rolff, J., Frese, S., Schirmer, U., Becker, M., Borno, S., Timmermann, B.,

[0379] Chavez, L., Sultmann, H., Leschber, G., Fichtner, I., Schweiger, M. R. & Herwig, R. 2017. QSEA- modelling of genome-wide DNA methylation from sequencing enrichment experiments. Nucleic Acids Research, 45.

[0380] Lienhard, M., Grimm, C., Morkel, M., Herwig, R. & Chavez, L. 2014. MEDIPS: Genome-wide differential coverage analysis of sequencing data derived from DNA enrichment experiments. Bioinformatics, 30, 284-286.

[0381] Liu MC, Oxnard GR, Klein EA, Swanton C, Seiden MV, Cummings SR, et al. Sensitive and specific multicancer detection and localization using methylation signatures in cell-free DNA. Annals of Oncology. 2020.

[0382] Lu MY, Chen TY, Williamson DFK, Zhao M, Shady M, Lipkova J, et al. Al-based pathology predicts origins for cancers of unknown primary. Nature. 2021 ;594(7861):106-10.

[0383] Martin, M. Cutadapt removes adapter sequences from high-throughput sequencing reads. EMBnet. journal 17, 10-12 (2011).

[0384] 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).

[0385] Mayordomo JI, Guerra JM, Guijarro C, Garcia-Prats MD, Gomez A, Lopez-Brea M, et al. Neoplasms of unknown primary site: A clinicopathological study of autopsied patients. Tumori. 1993;79(5):321-4.

[0386] Mclnnes, L, Healy, J, UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction, ArXiv e-prints 1802.03426, 2018

[0387] McLaren, W., et al. The ensembl variant effect predictor. Genome biology 17, 1-14 (2016).

[0388] Melville, J. uwot: the Uniform Manifold Approximation and Projection (UMAP) method for dimensionality reduction. R package version 0.1.14. (CRAN, 2022).

[0389] Moran S, Martinez-Cardus A, Sayols S, Musulen E, Balana C, Estival-Gonzalez A, et al. Epigenetic profiling to classify cancer of unknown primary: a multicentre, retrospective analysis. 2016. p. 1386-95.

[0390] Moss J, Magenheim J, Neiman D, Zemmour H, Loyfer N, Korach A, et al. Comprehensive human celltype methylation atlas reveals origins of circulating cell-free DNA in health and disease. Nature Communications. 2018;9(1).

[0391] Nice 2010. Metastatic malignant disease of unknown primary origin in adults: diagnosis and management. In: EXCELLENCE, N. I. F. H. A. C. (ed.). https: / / www.nice.org.uk / guidance / cg104.

[0392] Peran I, Madhavan S, Byers SW, McCoy MD. Curation of the Pancreatic Ductal Adenocarcinoma Subset of the Cancer Genome Atlas Is Essential for Accurate Conclusions about Survival-Related Molecular Mechanisms. Clin Cancer Res. 2018 Aug 15;24(16):3813-3819. doi: 10.1158 / 1078-0432. CCR-18-0290. Epub 2018 May 8. PMID: 29739787.

[0393] Posner, A., Prall, O. W., Sivakumaran, T., Etemadamoghadam, D., Thio, N., Pattison, A., Balachander,

[0394] 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 expression profiling to assist tissue of origin diagnosis in cancer of unknown primary. J Pathol.

[0395] 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 clinical trials: the TARGET study.

[0396] Saha SK, Zhu AX, Fuchs CS, Brooks GA. Forty-Year Trends in Cholangiocarcinoma Incidence in the U.S.: Intrahepatic Disease on the Rise. The Oncologist. 2016;21 (5):594-9.

[0397] Shah, S.P., et al. Integrating copy number polymorphisms into array CGH analysis using a robust HMM. Bioinformatics 22, e431-439 (2006).

[0398] Smith, T., Heger, A. & Sudbery, I. UMI-tools: modeling sequencing errors in Unique Molecular Identifiers to improve quantification accuracy. Genome Res 27, 491-499 (2017).

[0399] Van der Auwera, G.A. & O'Connor, B.D. Genomics in the cloud: using Docker, GATK, and WDL in Terra, (O'Reilly Media, 2020).

[0400] Wang P, Dong Q, Zhang C, Kuan PF, Liu Y, Jeck WR, Andersen JB, Jiang W, Savich GL, Tan TX, Auman JT, Hoskins JM, Misher AD, Moser CD, Yourstone SM, 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 1 and 2 occur frequently in intrahepatic cholangiocarcinomas and share hypermethylation 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.

[0401] For standard molecular biology techniques, see Sambrook, J., Russel, D.W. Molecular Cloning, A

[0402] Laboratory Manual. 3 ed. 2001 , Cold Spring Harbor, New York: Cold Spring Harbor Laboratory Press

[0403] 53

[0404] Table 1 : The top 250 genomic regions identified by the inventors from each of 30 pairwise comparisons. The DMRs may be those disclosed at https: / / zenodo.org / doi / 10.5281 / zenodo.4071360, as of 10 November 2023. The entire contents of https: / / zenodo.org / doi / 10.5281 / zenodo.4071360 are expressly incorporated herein by reference as “Table 1”.

[0405]

[0406]

[0407] Table 2: Top 250 genomic regions ranked by average variable importance in model.

[0408]

[0409] Table 3: List of classes used within CUPiD, with the number of methylation arrays for each class. All array samples come from The Cancer Genome Atlas (TCGA) or Gene Expression Omnibus (GEO), with regrouping of some TCGA classes for CUPiD v1

[0410]

[0411]

[0412] Table 4: CUPiD Predictions for the 41 patients in the CUP cohort

[0413] Table 5: The top 250 genomic regions identified by the inventors from each of 30 pairwise comparisons in CUPiD v1.1. The genomic regions may be those disclosed at https: / / doi.org / 10.5281 / zenodo.14054859, as of 8 November 2024. The entire contents of https: / / doi.org / 10.5281 / zenodo.14054859 are expressly incorporated herein by reference as “Table 5”.

[0414]

[0415]

[0416]

[0417]

[0418] Table 6: Top 250 genomic regions ranked by average variable importance in CuPiD v1.1 model.

[0419] Table 7: Genomic regions common to Tables 2 and 6.

[0420] Table 8: The top 500 genomic regions ranked by average variable importance in CUPiD v1.1 model. The genomic regions may be those disclosed at https: / / doi.org / 10.5281 / zenodo.14055157, as of 8 November 2024. The entire contents of https: / / doi.org / 10.5281 / zenodo.14055157 are expressly incorporated herein by reference as “Table 8”.

[0421] Table 9: The top 500 genomic regions ranked by average variable importance in CUPiD v1 model. The genomic regions may be those disclosed at https: / / doi.org / 10.5281 / zenodo.14055306, as of 8 November 2024. The entire contents of https: / / doi.org / 10.5281 / zenodo.14055306 are expressly incorporated herein by reference as “Table 9”.

Claims

Claims:1 . A computer-implemented method for determining the primary cancer of a malignancy of unknown origin (MUO) of a subject, the method comprising:(i) providing sequence data of cell-free DNA (cfDNA) obtained from a liquid sample from the subject, the sequence data comprising a region methylation signature (“sample data”);(ii) providing the sample data as input to a machine learning model comprising a multi-class classifier trained on a training data set comprising region methylation signatures from at least two primary malignant tumour types (“training data”), each primary malignant tumour type having a distinct location and / or histological subtype;(iii) receiving an output from the machine learning model, the output indicative of a primary malignant tumour type classification or an unknown type classification for the sample data; and(iv) determining the primary cancer of the subject 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), naive bayes or k-nearest neighbour algorithm, 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 tumour types comprise at least three primary malignant tumour types, optionally at least four primary malignant tumour types, wherein each primary malignant tumour type has a distinct location and / or histological subtype.

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

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

8. The method of any one of the preceding claims, wherein the at least two primary malignant tumour types are selected from: adrenocortical carcinoma (ACC), bladder urothelial carcinoma (BLCA), breast invasive carcinoma (BRCA), cervical squamous cell carcinoma (CervSq), cholangiocarcinoma (CHOL), diffuse large B-cell lymphoma (DLBC), glioblastoma multiforme (GBM), non-squamous gynaelogical carcinoma (Gynae), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), acute myeloid leukaemia (LAML), brain lower grade glioma (LGG), liver hepatocellular carcinoma (LIHC), colon and rectum adenocarcinoma (Lower Gl), 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 tumours (TGCT), thyroid carcinoma (THCA), thymoma (THYM), stomach and oesophageal adenocarcinoma (Upper Gl), head and neck and oesophageal squamous cell carcinoma (UpperSq) and uveal melanoma (UVM).

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

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

11. The method of any one of the preceding claims, wherein each region methylation signature comprises a DNA methylation status of at least one genomic region, optionally a plurality of genomic regions.

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

13. The method of claim 11 or claim 12, wherein the region methylation signature from each primary malignant tumour type comprises 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 regions are selected from the genomic regions listed in Table 1 .

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

16. The method of any one of claims 11 to 14, wherein the genomic regions are 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 regions comprise at least the top five genomic regions listed in Table 2.

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

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

20. The method of any one of claims 11 to 16 and 18, wherein the genomic regions are selected from the genomic regions listed in Table 7.21 . The method of any one of claims 11 to 20, wherein the genomic regions comprises genomic regions identified from a pairwise comparison of DNA methylation status of genomic regions between the at least two primary malignant tumour types, optionally wherein the genomic regions comprise at least 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 which have the greatest difference in methylation status as identified from a pairwise comparison.

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

23. The method of any one of the preceding claims, wherein the training data comprises region methylation signatures derived from primary malignant tumour tissue samples obtained from a plurality of individuals each having a known primary malignant cancer.

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

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

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

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

28. A method of determining the primary cancer of a subject having a malignancy of unknown origin (MUO) , the method comprising: i) analysing cell-free DNA (cfDNA) obtained from a liquid sample from the subject in order to obtain sequence data comprising a region methylation signature (“sample data”); and ii) performing the method of any one of claims 1 to 27 on the sample data to determine the primary cancer of the subject.

29. A method of selecting a cancer treatment for a subject having a malignancy of unknown origin (MUO), the method comprising: i) performing the method of any one of claims 1 to 27 to determine the likely primary cancer of the subject; and ii) selecting an anti-cancer treatment that targets the likely primary cancer.

30. A method of treating a subject having malignancy of unknown origin (MUO), the method comprising: i) performing the method of any one of claims 1 to 27 to determine the likely primary cancer of the subject; and ii) administering an anti-cancer treatment that targets the likely primary cancer.31 . A system comprising: a processor; and a computer readable medium comprising instructions that, when executed by the processor, cause the processor to perform the steps of the method of any one of claims 1 to 27.

32. One or more computer readable media comprising instructions that, when executed by one or more processors, cause the one or more processors to performs the steps of the method of any one of claims 1 to 27.

Images