Diagnostic cancer signatures

Abstract

The invention concerns a method for determining cancer in a subject, said method comprising performing Raman spectral analysis of a biological sample obtained from the subject and, optionally, comparing a portion of the spectrum from the sample with that of a control; use of the afore method to determine likelihood of cancer, in particular breast, pancreatic, lung and / or colorectal cancer; use of said method to further select a course of treatment for cancer, and a method of treatment comprising same; and a kit of parts for use in said method.

Description

[0001] Diagnostic Cancer Signatures

[0002] Field of the Invention

[0003] The invention concerns a method for determining cancer in a subject, said method comprising performing Raman spectral analysis of a biological sample obtained from the subject and, optionally, comparing a portion of the spectrum from the sample with that of a control; use of the afore method to determine likelihood of cancer, in particular breast, pancreatic, lung and / or colorectal cancer; use of said method to further select a course of treatment for cancer, and a method of treatment comprising same; and a kit of parts for use in said method.

[0004] Background of the Invention

[0005] Worldwide, cancer is the leading cause of deaths with 10 million deaths accounted for by cancer. The currently available methods for detection of cancer include flexible bronchoscopy, computed tomography (CT)-scan and X- ray. However, these technigues are not overly effective for early detection of the disease, as evidenced by the extremely poor rate of diagnosis of early stage disease.

[0006] Pan-cancer diagnostic tests have become a popular area of study in recent years, and could revolutionise current pathways to diagnosis. Raman spectroscopy offers a non-destructive, highly-sensitive label-free tool exploiting the biochemical changes in a sample under laser interrogation. Raman spectra contain a rich molecular fingerprint of a sample, whereby incident laser light induces the Raman effect; an inelastic scattering of light. This change in energy in the scattered light is related to the chemical constituents in a sample and its responsiveness to the laser energy. Due to its inherent ease of use, high reproducibility and non-invasiveness, Raman has previously been applied with success to a range of biofluids and tissue samples and has shown promise as a sensitive diagnostic tool, specifically in the form of surface enhanced Raman (SERS) technigues, to distinguish neoplastic from normal cells in cancers such as colorectal cancer, prostate, breast, cervical, gastric, oral and oesophageal cancer. However, many of these previous studies typically only demonstrate the potential for Raman spectroscopy utilising complex systems and methods to chemically I plasmonically enhance the Raman scattering signature from small sample sets, and as such a lack of reliability and diagnostic resolution has meant little of this has impacted the clinical setting.

[0007] There is therefore also clearly an unmet need for an improved cancer diagnostic tool and refinement of vibrational spectroscopic analysis in the clinical context that is able to finely resolve and identify those patients with cancer, and more so discern from those individuals with cancer the nature of the specific cancer that they are suffering, particularly when they present with vague symptoms. Further still, the ability to use a simple, highly powerful, non- invasive, technique that can achieve early diagnosis, and thus improve patient survival rates, is highly sought after.

[0008] Accordingly, we herein disclose a refined Raman spectral waveform signature and analysis that can be used to diagnose multiple cancer types, preferably by standard or non-surface enhanced Raman techniques, and, significantly and superiorly, even discern amongst those individuals the nature of the specific cancer they are suffering. In particular, through an exhaustive study of large datasets, we have been able to combine the complex nature of biological Raman spectra with interrogation via machine learning methods to provide a robust analysis into determining distinguishing characteristics between patient groups that represent reliable key diagnostic waveform signatures in multiple cancer types. Through subsequent validation, we have found such signatures provide for robust and superior diagnostic assays that pushes the diagnostic power of Raman analysis in cancer to a level of increased sensitivity and specificity across multiple cancer types that is a pre-requisite in any clinical diagnostic test. Statements of Invention

[0009] According to a first aspect of the invention there is provided a method for determining cancer in a subject, the method comprising: i) performing Raman spectral analysis of a biological sample obtained from the subject to produce a test sample spectrum; ii) comparing the test sample spectrum obtained in step i) with at least one control sample Raman spectrum from a control subject, iii) wherein a difference between the test sample spectrum and the control sample spectrum at one or more wavenumbers within one or more ranges selected from the group comprising: about 622 cm-1, about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 758 cm-1, about 803 cm-1, about 829 cm-1, about 853 cm-1, about 880 cm-1, about 943 cm-1, about 957 cm-1, about 1004 cm-1, about 1031 cm-1, about 1051 cm-1, about 1062 cm-1, about 1084 cm-1, about 1127 cm-1, about 1158 cm-1, about 1175 cm-1, about 1272 cm-1, about 1341 cm-1, about 1402 cm-1, about 1411 cm-1, about 1449 cm-1, about 1498 cm-1, about 1521 cm-1, about 1556 cm-1, about 1587 cm-1, about 1607 cm’ 1, about 1657 cm-1is indicative of a subject suffering from cancer.

[0010] In a preferred embodiment, said subject is a mammal, more preferably still human, equine, canine, feline, porcine, or any other domestic or agricultural species. In particularly preferred embodiments said subject is a human.

[0011] Remarkably, when analysing a biological sample, in particular blood, it has been found that these specific wavenumbers are able to accurately predict and diagnose multiple cancer types and thus provide a robust and highly accurate pan-cancer diagnostic method with heretofore undisclosed levels of sensitivity and specificity. In a preferred embodiment, the biological sample may be a tissue or biopsy sample, or a processed derivative thereof. In a preferred method, the biological sample is a blood sample. Reference to blood sample includes liquid or dried whole blood or a fraction thereof including serum and / or plasma, or a processed derivative thereof. Preferably, the blood sample is a liquid sample. Reference to a processed derivative includes reference to a blood sample, after it has been treated, typically for the purpose of preparing it for the method of the invention or preserving it prior to undertaking the said method and involves the use of conventional techniques well known to those skilled in the art of taking, preparing or preserving biological samples. Such techniques include, but are not limited to, freezing, thawing and / or dehydrating samples,

[0012] Reference herein to a control sample refers to a sample obtained from at least one control subject that has been shown not to have cancer using any one or more conventional techniques for identifying same such as, but not limited to bronchoscopy, sigmoidoscopy, colonoscopy, CT-scan, X-ray, ultrasound, MRI, biopsy or the like. As is disclosed herein, it has been found that, through rigorous analysis and machine learning, a refined Raman spectral waveform signature has been determined that can be used to diagnose and predict a variety of cancer types and, significantly and superiorly, not only predict the presence of cancer but actually diagnose the specific cancer a subject is likely to be suffering.

[0013] Preferably multiple control spectra are obtained for each control subject, ideally from multiple control subjects, and each spectrum is preferably subjected to one or more, preferably two or more, of wavenumber correction, baseline correction and vector normalisation.

[0014] As used herein, the term “cancer” refers to cells having the capacity for autonomous growth, i.e., an abnormal state or condition characterized by uncontrolled cell proliferation. The term is meant to include all types of cancerous growths or oncogenic processes, metastatic tissues or malignantly transformed cells, tissues, or organs, irrespective of histopathologic type or stage of invasiveness.

[0015] Most preferably the cancer referred to herein includes any one or more of the following cancers: nasopharyngeal cancer, synovial cancer, hepatocellular cancer, renal cancer, cancer of connective tissues, melanoma, lung cancer, bowel cancer, colorectal cancer, brain cancer, throat cancer, oral cancer, liver cancer, bone cancer, pancreatic cancer, choriocarcinoma, gastrinoma, pheochromocytoma, prolactinoma, T-cell leukemia / lymphoma, tonsil, spleen, neuroma, von Hippel-Lindau disease, Zollinger-Ellison syndrome, adrenal cancer, anal cancer, bile duct cancer, bladder cancer, ureter cancer, glioma, oligodendroglioma, neuroblastoma, meningioma, spinal cord tumor, osteochondroma, chondrosarcoma, Ewing's sarcoma, carcinoid, carcinoid of gastrointestinal tract, fibrosarcoma, breast cancer, muscle cancer, Paget's disease, cervical cancer, rectal cancer, esophagus cancer, gall bladder cancer, cholangioma cancer, head cancer, eye cancer, nasopharynx cancer, neck cancer, kidney cancer, Wilms' tumor, liver cancer, Kaposi's sarcoma, prostate cancer, testicular cancer, Hodgkin's disease, non-Hodgkin's lymphoma, skin cancer, mesothelioma, myeloma, multiple myeloma, ovarian cancer, endocrine pancreatic cancer, glucagonoma, parathyroid cancer, penis cancer, pituitary cancer, soft tissue sarcoma, retinoblastoma, small intestine cancer, stomach cancer, thymus cancer, thyroid cancer, trophoblastic cancer, hydatidiform mole, uterine cancer, endometrial cancer, vagina cancer, vulva cancer, acoustic neuroma, mycosis fungoides, insulinoma, carcinoid syndrome, somatostatinoma, gum cancer, heart cancer, lip cancer, meninges cancer, mouth cancer, nerve cancer, palate cancer, parotid gland cancer, peritoneum cancer, pharynx cancer, pleural cancer, salivary gland cancer, tongue cancer and tonsil cancer. More preferably still, said cancer is selected from the group comprising the following cancers: colorectal, lung, pancreatic, and breast cancers.

[0016] Further, the methods described herein may also be able to identify precancerous conditions, and distinguish same from cancer. As used herein, reference to a precancerous condition refers to a sample provided from an individual confirmed as not having a cancer using any one or more conventional techniques for identifying same, such as those listed above and / or from a subject not yet with confirmed cancer but is observed have a pre- malignancy or pre-cancerous condition or disorder that, if left untreated, has an increased likelihood of developing into cancer ideally, using any one or more of the above conventional techniques for identifying same. For example, this includes, but is not limited to: ductal carcinoma in-situ, lobular carcinoma in- situ, Sclerosing adenosis, Small duct papilloma as breast cancer pre- malignancies; colorectal adenoma, sessile serrated lesion, Lynch syndrome, familial adenomatous polyposis as colorectal cancer pre-malignancies; or Mucinous cystic neoplasm (MCN), Intraductal papillary mucinous neoplasm (IPMN) or Neuroendocrine tumours as pancreatic cancer pre-malignancies.

[0017] As will be appreciated, in this manner, the method involves comparing spectra of test samples to spectra of control samples obtained from healthy individuals, or those having a precancerous condition as defined herein, and / or both. In this manner, it is possible to observe those differences in test spectra (where disease or precancerous condition is confirmed or suspected) compared to healthy / pre-malignant spectra i.e. to determine those differences that are indicative of cancer. Optionally by also comparing spectra of pre-malignant individuals with healthy controls, one can add a further level of analysis by determining specifically those spectral differences of a pre-cancerous disease state. In yet a further preferred embodiment said control, and / or said test sample are age / gender / weight-matched.

[0018] In a preferred method of the invention, Raman analysis is undertaken on a liquid blood or blood derivative sample obtained from the subject. This minimises additional drying processes. The blood or blood derivative sample is preferably fresh.

[0019] The blood or blood derivate sample may also be analysed once dried. The method may, therefore, comprise the step of drying the sample. The drying step may involve drying sample at room temperature or via assisted drying (e.g. vacuum drying). It is beneficial that the sample may be dried on a sample holder.

[0020] In the present invention, Raman analysis comprises irradiating the sample with a light source, preferably a laser light source, more preferably a laser light source configured to emit light in the wavelength band of infrared light, most preferably 785 nm. In particularly preferred embodiments the laser light source generates laser light having a power of from about 10 mW to about 1 W, and more preferably from about 50 mW to about 500 mW.

[0021] The output spectrum is preferably recorded and analysed in a spectral region between about 600 and about 1700 cm-1. This range has been determined to encompass the fullest spectral output that allows reproducible discrimination.

[0022] Reference herein to the term ‘about’ means plus or minus 5% and most preferably plus or minus 2%. For example, given the nature of the art it will be appreciated by those skilled in the field that there may be variation around recited wavenumbers owing to sample variability, for example, ±5 cm-1, ±4 cm-1, ±3 cm-1, ±2 cm-1, ±1 cm-1. As will be appreciated by those skilled in the art, in a preferred method of the invention the, or each, spectra preferably undergoes one or more conventional pre-processing steps prior to or following the comparison step to reduce the noise associated with the one or more spectra to provide the, or each, processed spectra. The pre-processing step(s) may comprise one or more of: data binning, smoothing, background subtraction (e.g. Extended multiplicative scatter correction), and / or normalisation such as vector normalisation and / or baseline correction, or other method known to those skilled in the art. Preferably multiple output spectra are obtained and each spectrum is preferably subjected to one or more, preferably two or more, of wavenumber correction, baseline correction and vector normalisation.

[0023] In preferred embodiments, the, or each, processed spectra may be further processed to provide one or more processed or derivative spectrum such as, but not limited to, dimensionally reduced spectra, feature reduced spectra or the like. The or each derivative spectra is then compared to similarly derived control spectra. As will be appreciated, also in this manner, such derived spectra can be prepared by comparing sample to control to readily identify additional or further features of interest. For example a Random Forest Gini importance derived spectra, or trace, can be prepared by comparing sample and control spectra to aid visual identification of significant spectral regions of difference or importance, and utilised to assess changes in spectra or areas of interest as defined herein.

[0024] As will be appreciated, the methods of the invention and, in particular, the identification of spectral wavenumber regions of diagnostic relevance (features of importance) for determining cancer status, were derived by post processing and comparing large spectral data sets obtained from healthy subjects with those known to have cancer or a precancerous condition. Such post- processing steps include the application of machine learning methods aimed to produce a mathematical model which allows the classification of one state or another. This is achieved using supervised learning, whereby examples of a given class (e.g. diseased or non-diseased) are fed into a machine learning algorithm and patterns within a given class are identified to allow for classification of unseen data, known as generalisation.

[0025] Feature selection (FS) techniques, sometimes referred to as feature reduction techniques, have been used to tackle a large data set with 1000+ variables and reduce this significantly. This is particularly useful with Raman spectra where the data is highly correlated and hence some wavenumbers may not provide any more additional information and can be removed. There are many methods of feature reduction known to those skilled in the art, such as principle component analysis, factor analysis, ElasticNet, random forest feature selection, etc. The preferred method utilised in this study was random forest feature selection (RFFS).

[0026] Random forests tend to correct for the pure decision trees tendency to overfit to the training set. In RF, the rule is to maximise the decrease of Gini impurity which occurs at each split. The Gini impurity gives a probability of a new data point being incorrectly classified given the current split performance. This indicates how often a variable is selected in each split and the magnitude of the variables discriminatory power for the classification problem at hand. We then selected the most important features coined random forest feature selection (RFFS).

[0027] The Leave-One-Out Cross-Validation, or LOOCV, procedure was used to evaluate and optimise the RFFS algorithm with the relatively small datasets here for each cancer type.

[0028] In a preferred method, step iii) comprises or consists of observing a difference wherein an increase or decrease in signal intensity at the same wavenumber or a shift in position of signal maxima or minima between wavenumbers between the test sample spectrum produced from the sample and the control sample spectrum is indicative of a subject suffering from cancer. Most preferably, an increase or decrease in signal intensity between wavenumbers between the spectrum produced from the sample and the control is observed.

[0029] The method is preferably a pan-cancer diagnostic method, wherein the method comprises or consists of comparing the signal intensity at one, or any combination of, wavenumbers selected from the group comprising or consisting of: about 666 cm-1, about 682 cm-1, about 829 cm-1, about 853 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 803 cm-1, about 1004 cm-1, about 1127 cm-1, about 1175 cm-1, about 1341 cm-1, about 1402 cm-1, about 1411 cm-1, about 1498 cm-1, and about 1657 cm-1, wherein a difference compared with control is indicative of a subject suffering from cancer, and more preferably still a combination of at least two wavenumbers, and more preferably still a combination of three, four, five, six, seven, eight, nine, ten, and / or eleven wavenumbers. Preferably, observing an increased signal intensity at a wavenumber selected from one or more, ideally all, of about 1004 cm-1, about 1127 cm-1, about 1341 cm-1, about 1402 cm-1, about 1411 cm-1and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 803 cm-1is indicative of a subject suffering from cancer. In particularly preferred embodiments, observing an increased signal intensity at a wavenumber selected from one or ideally both, of: about 1004 cm-1, about 1402 cm-1and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: about 666 cm-1, about 682 cm-1, 745 cm-1is indicative of a subject suffering from cancer.

[0030] In an alternative embodiment, the method comprises or consists of comparing the signal intensity at one, or any combination of, wavenumbers selected from the group comprising or consisting of: about 644 cm-1, about 666 cm-1, about 682 cm-1, about 745 cm-1, about 853 cm-1, about 1004 cm-1, about 1031 cm-1, about 1062 cm-1, about 1084 cm-1, about 1127 cm-1, about 1158 cm-1, about 1175 cm-1, about 1341 cm-1, about 1402 cm-1, about 1411 cm-1, about 1498 cm-1, wherein a difference compared with control is indicative of a subject suffering from Pancreatic cancer, and more preferably still a combination of at least two wavenumbers, and more preferably still a combination of three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, and / or sixteen wavenumbers. Preferably, observing an increased signal intensity at a wavenumber selected from one or more, ideally all, of about 644 cm-1, about 853 cm-1, about 1004 cm-1, about 1031 cm-1, about 1062 cm-1, about 1084 cm-1, about 1127 cm-1, about 1175 cm-1, about 1341 cm-1, about 1402 cm-1, about 1411 cm-1, about 1498 cm-1and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: about 666 cm-1, about 682 cm-1, about 745 cm-1, about 1158 cm-1is indicative of a subject suffering from pancreatic cancer.

[0031] In a further alternative embodiment, the method comprises or consists of comparing the signal intensity at one, or any combination of, wavenumbers selected from the group comprising or consisting of: about 622 cm-1, about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 803 cm-1, about 829 cm-1, about 853 cm-1, about 880 cm-1, about 1051 cm-1, about 1175 cm-1, about 1272 cm-1, about 1402 cm-1, about 1411 cm-1, about 1498 cm-1, about 1657 cm-1wherein a difference compared with control is indicative of a subject suffering from lung cancer, and more preferably a combination of at least two wavenumbers, and more preferably still a combination of at least two wavenumbers, and more preferably still a combination of three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, sixteen, seventeen and / or eighteen wavenumbers. Preferably, observing an increased signal intensity at a wavenumber selected from one or more, ideally all, of about 829 cm-1, about 853 cm-1, about 880 cm-1, about 1051 cm-1, about 1402 cm-1, about 1411 cm-1, and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: 622 cm-1, about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 803 cm-1, about 1175 cm-1, about 1272 cm-1, about 1498 cm-1, about 1657 cm-1is indicative of a subject suffering from lung cancer.

[0032] In yet a further alternative embodiment, the method comprises or consists of comparing the signal intensity at one, or any combination of, wavenumbers selected from the group comprising or consisting of: the group comprising: about 622 cm-1, about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 758 cm-1, about 803 cm-1, about 829 cm-1, about 853 cm-1, about 1004 cm-1, about 1127 cm-1, about 1175 cm-1, about 1449 cm-1, about 1498 cm-1, about 1657 cm-1wherein a difference compared with control is indicative of a subject suffering from breast cancer, and more preferably a combination of at least two wavenumbers, and more preferably still a combination of three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, sixteen, and / or seventeen wavenumbers. Preferably, observing an increased signal intensity at a wavenumber selected from one or more, ideally all, of : about 622 cm-1, about 758 cm-1, about 1004 cm-1, about 1127 cm-1, about 1175 cm-1, and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 803 cm-1, about 829 cm-1, about 853 cm-1, about 1449 cm-1, about 1498 cm-1is indicative of a subject suffering from breast cancer.

[0033] In yet a further alternative embodiment, the method comprises or consists of comparing the signal intensity at one, or any combination of, wavenumbers selected from the group comprising or consisting of: the group comprising about 622 cm-1about 829 cm-1, about 943 cm-1, about 957 cm-1, about 1004 cm-1, about 1341 cm-1, about 1402 cm-1, about 1498 cm-1, about 1521 cm-1, about 1556 cm-1, about 1587 cm-1, about 1607 cm-1, about 1657 cm-1wherein a difference compared with control is indicative of a subject suffering from colorectal cancer, and more preferably a combination of at least two wavenumbers, and more preferably still a combination of three, four, five, six, seven, eight, nine, ten, eleven, twelve, and / or thirteen wavenumbers. Preferably, observing an increased signal intensity at a wavenumber selected from one or more, ideally all, of : about 829 cm-1, about 1004 cm-1, about 1341 cm-1, about 1402 cm-1, about 1498 cm-1, about 1556 cm-1, about 1587 cm-1, about 1607 cm-1, about 1657 cm-1, and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: about 622 cm-1, about 943 cm-1, about 957 cm-1, about 1521 cm-1, is indicative of a subject suffering from colorectal cancer.

[0034] According to a second aspect of the invention there is provided a method for determining pancreatic cancer in a subject, the method comprising: i) performing Raman spectral analysis of a biological sample obtained from the subject to produce a test sample spectrum; ii) comparing the test sample spectrum obtained in step i) with a control sample Raman spectrum from a control subject, iii) wherein a difference between the test sample spectrum and the control sample spectrum at one or more wavenumbers within one or more ranges, ideally all ranges, selected from the group comprising: about 644 cm-1, about 666 cm-1, about 682 cm-1, about 745 cm-1, about 853 cm-1, about 1004 cm-1, about 1031 cm-1, about 1062 cm-1, about 1084 cm-1, about 1127 cm-1, about 1158 cm-1, about 1175 cm-1, about 1341 cm-1, about 1402 cm-1, about 1411 cm-1, about 1498 cm’ is indicative of a subject suffering from pancreatic cancer.

[0035] In a preferred method of the second aspect of the invention, observing an increased signal intensity at a wavenumber selected from one or more, ideally all, of about 644 cm-1, about 853 cm-1, about 1004 cm-1, about 1031 cm-1, about 1062 cm-1, about 1084 cm-1, about 1127 cm-1, about 1175 cm-1, about 1341 cm-1, about 1402 cm-1, about 1411 cm-1, about 1498 cm-1and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: about 666 cm-1, about 682 cm-1, about 745 cm-1, about 1158 cm-1is indicative of a subject suffering from pancreatic cancer.

[0036] According to a third aspect of the invention there is provided a method for determining lung cancer in a subject, the method comprising: i) performing Raman spectral analysis of a biological sample obtained from the subject to produce a test sample spectrum; ii) comparing the test sample spectrum obtained in step i) with a control sample Raman spectrum from a control subject, iii) wherein a difference between the test sample spectrum and the control sample spectrum at one or more wavenumbers within one or more ranges, ideally all ranges, selected from the group comprising: about 622 cm-1, about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 803 cm-1, about 829 cm-1, about 853 cm-1, about 880 cm-1, about 1051 cm-1, about 1175 cm-1, about 1272 cm-1, about 1402 cm-1, about 1411 cm-1, about 1498 cm-1, about 1657 cm-1is indicative of a subject suffering from lung cancer.

[0037] In a preferred method of the third aspect of the invention, observing an increased signal intensity at a wavenumber selected from one or more, ideally all, of about 829 cm-1, about 853 cm-1, about 880 cm-1, about 1051 cm-1, about 1402 cm-1, about 1411 cm-1, and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: 622 cm-1, about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 803 cm-1, about 1175 cm-1, about 1272 cm-1, about 1498 cm-1, about 1657 cm-1is indicative of a subject suffering from lung cancer. According to a fourth aspect of the invention there is provided a method for determining breast cancer in a subject, the method comprising: i) performing Raman spectral analysis of a biological sample obtained from the subject to produce a test sample spectrum; ii) comparing the test sample spectrum obtained in step i) with a control sample Raman spectrum from a control subject; iii) wherein a difference between the test sample spectrum and the control sample spectrum at one or more wavenumbers within one or more ranges, ideally all ranges, selected from the group comprising: about 622 cm-1, about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 758 cm-1, about 803 cm-1, about 829 cm-1, about 853 cm-1, about 1004 cm-1, about 1127 cm-1, about 1175 cm-1, about 1449 cm-1, about 1498 cm’ 1, about 1657 cm-1is indicative of a subject suffering from breast cancer.

[0038] In a preferred method of the fourth aspect of the invention observing an increased signal intensity at a wavenumber selected from one or more, ideally all, of : about 622 cm-1, about 758 cm-1, about 1004 cm-1, about 1127 cm-1, about 1175 cm-1, and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 803 cm-1, about 829 cm-1, about 853 cm-1, about 1449 cm-1, about 1498 cm-1is indicative of a subject suffering from breast cancer.

[0039] According to a fifth aspect of the invention there is provided a method for determining colorectal cancer in a subject, the method comprising: i) performing Raman spectral analysis of a biological sample obtained from the subject to produce a test sample spectrum; ii) comparing the test sample spectrum obtained in step i) with a control sample Raman spectrum from a control subject; iii) wherein a difference between the test sample spectrum and the control sample spectrum at one or more wavenumbers within one or more ranges, ideally all ranges, selected from the group comprising: about 622 cm-1about 829 cm-1, about 943 cm-1, about 957 cm-1, about 1004 cm-1, about 1341 cm-1, about 1402 cm-1, about 1498 cm-1, about 1521 cm-1, about 1556 cm-1, about 1587 cm-1, about 1607 cm-1, about 1657 cm-1is indicative of a subject suffering from colorectal cancer.

[0040] In a preferred method of the fifth aspect of the invention, observing an increased signal intensity at a wavenumber selected from one or more, ideally all, of : about 829 cm-1, about 1004 cm-1, about 1341 cm-1, about 1402 cm-1, about 1498 cm-1, about 1556 cm-1, about 1587 cm-1, about 1607 cm-1, about 1657 cm-1, and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: about 622 cm-1, about 943 cm-1, about 957 cm-1, about 1521 cm-1, is indicative of a subject suffering from colorectal cancer.

[0041] According to a further aspect of the invention, there is provided a method for monitoring the progression of cancer in a subject comprising repeating one or more of the afore, and ideally the same, method(s) periodically.

[0042] Ideally, Raman analysis is correlated with known cancer staging techniques such that of a simple in vitro assay or biopsy used to reliably inform a clinician about, not only the existence of a given cancer, but also its stage or progression.

[0043] As will be appreciated by those skilled in the art, in the above method of the invention comparing signal intensity with respect to wavenumber from the test sample spectrum and / or analyzing relative spectral difference(s) between the test sample spectrum and the control sample spectrum at one or more wavenumbers within one or more ranges disclosed herein can be used to assess how effective a treatment regimen is working, for example, by assaying the signal intensity levels during the course of a given therapy to determine if there is a change in signal intensity in response to said treatment.

[0044] Additionally, or alternatively, there is provided a method for treating cancer comprising performing any one of the afore methods and then, depending upon the outcome of the method, undertaking a suitable or selected course of treatment.

[0045] According to a further aspect of the invention there is provided a kit for use in determining cancer in a biological sample from subject, said kit comprising: a) a Raman spectrometer for performing spectral analysis of a biological sample obtained from the subject; b) a processing unit, wherein said processing unit processes the test sample spectrum according to a method as defined herein; and c) an output unit arranged to provide an output indicative of a subject suffering from cancer in the subject according to a determination made by the processing unit in step b).

[0046] In a preferred kit of the invention, said Raman spectrometer comprises a laser light source, more preferably a laser light source arranged to emit light in the wavelength band of infrared light, most preferably 785 nm. Alternatively or additionally, the light source preferably generates laser light having a power of from about 10 mW to about 1 W, and more preferably from about 50 mW to about 500 mW.

[0047] Particularly suitable Raman spectrometers comprise a long pass or Raman edge filter, dispersive device and charge coupled device, configured to produce a spectral response of intensity vs energy (relative Raman shift in wavenumbers). Throughout the description and claims of this specification, the words “comprise” and “contain” and variations of the words, for example “comprising” and “comprises”, mean “including but not limited to” and do not exclude other moieties, additives, components, integers or steps. Throughout the description and claims of this specification, the singular encompasses the plural unless the context otherwise requires. In particular, where the indefinite article is used, the specification is to be understood as contemplating plurality as well as singularity, unless the context requires otherwise.

[0048] All references, including any patent or patent application, cited in this specification are hereby incorporated by reference. No admission is made that any reference constitutes prior art. Further, no admission is made that any of the prior art constitutes part of the common general knowledge in the art.

[0049] Preferred features of each aspect of the invention may be as described in connection with any of the other aspects.

[0050] Other features of the present invention will become apparent from the following examples. Generally speaking, the invention extends to any novel one, or any novel combination, of the features disclosed in this specification (including the accompanying claims and drawings). Thus, features, integers, characteristics, compounds or chemical moieties described in conjunction with a particular aspect, embodiment or example of the invention are to be understood to be applicable to any other aspect, embodiment or example described herein, unless incompatible therewith.

[0051] Moreover, unless stated otherwise, any feature disclosed herein may be replaced by an alternative feature serving the same or a similar purpose.

[0052] The Invention will now be described by way of example only with reference to the Examples below and to the following Figures wherein: Figure 1. a) ROC curve from LOOCV where all cancers are included as only one class, against propensity score matched controls. Marked on the plots with a red dot are the model performance in the case where the probability threshold for cancer is 0.5. In addition, model performance where the probability threshold has been adjusted to achieve a minimum sensitivity of 90% are marked with a green cross, b) Shows the feature importance (Gini importance) for the general cancer control model normalised between 0 and 1.

[0053] Figure 2. Random forest Gini importance (i.e. wavenumber importance) for each cancer type overlaid on a typical serum spectrum.

[0054] Figure-3. ROC curves for each cancer type from LOOCV. Marked on the plots with a red dot are the model performance in the case where the probability threshold for cancer is 0.5. In addition, model performance where the probability threshold has been adjusted to achieve a minimum sensitivity of 90% are marked with a green cross.

[0055] Figure-4. Feature importance for a model trained to distinguish pancreatic cancer from colorectal cancer overlaid on a typical human blood serum sample.

[0056] Figure-5. Feature importance for a model trained to distinguish pancreatic cancer from lung cancer overlaid on a typical human blood serum sample.

[0057] Figure 6. Feature importance for a model trained to distinguish pancreatic cancer from breast cancer overlaid on a typical human blood serum sample.

[0058] Figure 7. Feature importance for a model trained to distinguish lung cancer from breast cancer overlaid on a typical human blood serum sample.

[0059] Figure 8. Feature importance for a model trained to distinguish lung cancer from colorectal cancer overlaid on a typical human blood serum sample. Figure 9. Feature importance for a model trained to distinguish breast cancer from colorectal cancer overlaid on a typical human blood serum sample.

[0060] Table 1. Breakdown of the number of samples from each cancer type and propensity score matched controls based on age and sex. This gives a total number of cancers of 263 with matched control samples;

[0061] Table 2. Model performances for each cancer type using LOOCV with a probability threshold of 0.5.

[0062] Table 3. Model performance for each cancer type using LOOCV where the model probability has been thresholded using the AUC to achieve a minimum of 90% sensitivity.

[0063] Table 4. Tentative peak assignments for pancreatic cancer.

[0064] Table 5. Metabolite pathways from most important model features in pancreatic cancer from the online KEGG pathway mapper.

[0065] Table 6. Matched Raman peaks with mass-spectrometry for the top 100 features seen in lung cancer.

[0066] Table 7. Metabolite pathways from most important model features in lung cancer from the online KEGG pathway mapper.

[0067] Table 8. Tentative peak assignments for Breast cancer.

[0068] Table 9. Metabolite pathways from most important model features in breast cancer from the online KEGG pathway mapper. Table 10. Tentative peak assignments from Raman spectroscopy and mass- spectrometry for the top 100 features seen in colorectal cancer.

[0069] Table 11. Metabolite pathways from most important model features in colorectal cancer from the online KEGG pathway mapper.

[0070] Table 12. Pancreatic cancer vs. colorectal cancer performance with and without probability thresholding to achieve a minimum of 90% sensitivity.

[0071] Table 13. Pancreatic cancer vs. lung cancer performance metrics in cross- validation.

[0072] Table 14. Pancreatic cancer vs. breast cancer performance metrics in cross- validation.

[0073] Table 15. Lung cancer vs. breast cancer performance metrics in cross- validation.

[0074] Table 16. Lung cancer vs. colorectal cancer performance metrics in cross- validation.

[0075] Table 17. Breast cancer vs. colorectal cancer performance metrics in cross- validation.

[0076] Table 18. Confusion matrix with metrics calculated based on the clinical “gold standard” actual diagnosis, and the predicted diagnosis based on the Raman measurement. Methods

[0077] Serum Collection

[0078] Patient characteristics at time of sampling may define the accuracy of the resultant spectrum. Patients are preferentially fasted for 4 hours pre-sampling and not having diseases of the liver. Details of patient medication are also recorded. Blood samples are taken by a skilled phlebotomist via normal standard operating procedures. Vacutainer™ Serum Separator blood collection tubes were used to collect the blood. The collection tubes were then handled according to the manufacturer's best practice protocols in order to produce liquid serum. Samples were typically frozen and stored to preserve before analysis. They were analysed by Raman spectroscopy as a liquid sample.

[0079] Raman Spectroscopy of Liquid Samples (785 nm Laser)

[0080] Liquid samples were pipetted into a receptacle in the form of a stainless-steel sample holder which had multiple wells. This was then placed into the spectrometer onto a stainless-steel cooling plate. Using an objective the 785 nm laser light was focused to 1.2 mm above the base of the well into the liquid sample. Data points were then taken using 165-175 mW laser power for 5 s exposure time in the spectral region between 610 cm-1and 1718 cm-1. This was then averaged over 30 acquisitions to produce one spectrum. This process was then repeated to produce 5 replicates per sample and is used in the diagnostic model to check on degree of spectral variances associated with ‘sampling’ reproducibility.

[0081] Spectral Data Pre-Processing

[0082] Pre-processing of Raman spectral data was undertaken using a computational package developed to help optimise diagnostic model performance by testing the best combination of pre-processing steps and the order in which they are applied. Whilst good diagnostic performance can be achieved using various other pre-processing methods and / or using an alternative configuration or order of pre-processing steps, excellent diagnostic performance was achieved undertaking a preferred order of such steps: binning, data smoothing, baseline correction and normalisation. Each pre-processing step can be described in the following way:

[0083] Data binning: This method minimises the effect of observational errors within a data set, and in spectral data, increases the signal to noise ratio. Binning has the added benefit of reducing the data dimensionality which can decrease the computational burden on data processing with highly multivariate data. In the analysis of data here, binning was set to 1.

[0084] Data smoothing: This process minimises noise in a data set by fitting a curve to replace the raw data thus increasing the signal-to-noise ratio. In this study, where the pre-processing sequence contains a smoothing step, the Savitsky- Golay (SG) filter has been applied. SG functions by operating a sliding window along the data set and fitting a polynomial of order n to that window thus replacing spectra with a fitted polynomial of that window. In the preferred analysis here, the filter window length was set to 9 and the order of the polynomial was 4.

[0085] Background removal: This process attempts to remove any background contributions to the spectra from fluorescence without removing features from the spectra (peaks). It can be achieved by a number of methods:

[0086] Rolling Circle Filter (RCF): This is a high pass filter for the removal of background fluorescence which works by rolling under the spectra storing the minimum distance the between the circle and the data. This is repeated and evaluated for shifted windows along the spectra, resulting in a list of the minimum distance between the circumference of the circle and the spectra which is then removed from the spectra as the background contribution. The RCF removes broad features from the spectra, thus maintaining the sharp characteristic Raman peaks. The tunable parameter with the RCF is the radius of the circle, which is typically chosen such that broad features are effectively removed and such that the circle cannot “roll” into Raman peaks.

[0087] Polynomial Baseline Removal: This method fits an n order polynomial to the spectra using least squares, aiming to remove broader features uncharacteristic of Raman while maintaining peak structures. The method works by smoothing the spectrum in such a way that maintains the Raman peaks while fitting the background.

[0088] Extended multiplicative scatter correction: This method is a powerful pre-processing technique that isolates and removes complicated multiplicative effects caused by physical phenomena so that spectral features are more readily determined.

[0089] Derivative Baseline Removal: By taking the derivative of the measured response with respect to the wavenumber (or index), one can evaluate the derivative of spectra, accentuating the maxima and minima indicative of the Raman spectral feature. One can either take the first or second derivative to achieve this, both of which remove low frequency data from the spectra as a high pass filter. The preferred method used here was chosen to be polynomial background removal of order 5.

[0090] Normalisation: Generally considered an essential step when analysing spectroscopic data due to the innate fluctuation of intensity counts in the spectra from day to day. Normalising the spectra aims to strip out this fluctuation in intensity and allow the analyst to compare spectra more easily. Whilst various methods can be employed such as min-max, vector normalisation, normalisation to a major peak and standard normal variate, the preferred method employed here was standard normal variate.

[0091] Samples

[0092] Table 1 shows the number of patients in each cancer type. Patients are matched using propensity score matching based on sex and gender. Controls are determined by clear CT scan, or colonoscopy for CRC, and cancers confirmed by histopathology from tumour biospy.

[0093] Machine Learning Classification

[0094] Classification of Raman spectra were performed using a Random Forest machine learning algorithm. Briefly, random forest works by building n decision trees, where each tree begins at a random variable, or wavenumber, and the resultant model combines all n trees for an ensemble result. A benefit of random forest is the ability to extract important variables in decision making, and in the case of Raman spectroscopy one can map these wavenumbers back to vibrational modes of molecules for identification. In addition, RF performs well out-of-the-box, and through bagging tends not to overfit as other classical ML methods fall into. This information of important features can also be used to rebuild models with fewer variables to remove any which do not contribute to classification.

[0095] Performance assessment of each model is described using the metrics sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy defined in Table 18.

[0096] In addition, we observe the area-under-the-receiver-operator-curve (AUC, or ALIROC, or ROC curve) which is, by definition, the area under the plot which shows how well a binary classifier performs when varying the discrimination threshold. Figure 1 a) shows an example ROC curve, where optimal performance is in the top left of the plot, and an AUC = 1 is a perfect classifier.

[0097] Uncertainty in Model Performance

[0098] Uncertainties in the performance of machine learning algorithms in this case where LOOCV is utilised is calculated using the equation where 1.96 originates from the 95% confidence interval, metric refers to the metric in question for example sensitivity of the model, prevalence the proportion of cases in the data set, and N the total number of samples in the data set. In this study, the prevalence is 50%, or 0.5, for each cancer type trialled.

[0099] Since the data sets are from between N = 61 at a minimum for lung cancer (31 lung patients), to N = 300 maximum for CRC cancer (150 CRC patients), it is expected that uncertainties will be large, ranging from 5% to 17% for some metrics. This is clearly large and will be addressed in the discussion section.

[0100] As Raman spectroscopy provides the fingerprint of vibrational modes of a sample, this can lead to multiple potential assignments for the same peak. To provide assurances of the matched Raman peaks, 61 serum samples underwent both Raman analysis and mass-spectrometry. Correlations with the Raman wavenumbers and mass-spectrometry measurements combined with existing Raman peak libraries have been used to build a data base of metabolites with Raman peaks.

[0101] Results

[0102] In the first instance, a machine learning model has been developed for general cancer detection incorporating the full data set containing 263 cancers with 263 matched controls. Figure 1 shows the receiver-operator-characteristic (ROC) curve for the model. With a probability set to 0.5, the model achieves a sensitivity of 78.3% and a specificity of 89.4%, and when established with threshold to achieve a minimum sensitivity of 90% (probability cut-off=0.374) it achieves a sensitivity of 90.1% and specificity of 83.3%. The feature importance for the all cancers vs. controls can be seen in figure 1b over a typical serum spectrum. Feature importance from the RF algorithm shows which wavenumbers prove least "impure" on splitting decisions. This means peaks in the feature importance corresponds to wavenumbers where the clearest differences between cancer and control are seen. From the feature importance we are then able to remove poorly differentiating wavenumbers and hence simplify models. With more date one would expect the noise around the feature importance to reduces as the distinguishing characteristics are more well established.

[0103] These results are very promising with high sensitivity and specificity. If, for example, a pan-cancer test without cancer specificity was a feasible solution, RS could be utilised. However, being able to ascertain the specific location of the cancer via a blood-based Raman test could streamline current NHS pathways.

[0104] Tables 2 and 3 show the model performances for paired cancer vs. control samples given in table 1 in leave-one-out-cross-validation (LOOCV). Figure 2 shows the feature importance for these paired cancer vs. control models. This shows distinct differences between the cancer types in many regions, along with similarities. Understanding the differences between cancer types metabolomically can help with detection, treatment, and outcomes if cancers are caught at earlier stages.

[0105] Figure 3 shows the ROC curve for each of the cancer models along with performance of the model with the probability set to 0.5 (red dot) which links the performances in table 2, and the performance with a threshold probability to achieve a minimum sensitivity of 90% (green cross) which links to the performances in table 3.

[0106] Pancreatic Cancer As shown in Figure 2, the most important regions for the Random Forest classification of pancreatic cancers are primarily in the wavenumber regions 620-800 cm-1, 1000-1200 cm-1, and 1350-1550 cm-1. Approximate peak assignments are given in Table 4 where wavenumbers are matched to the nearest peak. This is to account for some important features falling on shoulders of peaks and have been combined for brevity. From the random forest feature importance, the top 10 metabolites for pancreatic cancer diagnosis via Raman are; L-Phenylalanine, L-Proline, Pantothenate, D- Fructose, Sarcosine, Alanine, L-Tryptophan, Citric acid, Phenol, and L- Glutamine.

[0107] Table 5 shows the output from feeding the top metabolites for pancreatic cancer as determined from RS into the Kyoto Encyclopaedia of Genes and Genomes (KEGG) pathway mapper. The pathways with the top hits include the alanine, aspartate, and glutamate metabolism, and the valine, leucine, and isoleucine biosynthesis / degradation. Statistical significance has been seen previously between pancreatic ductal adenocarcinoma and other pancreatic cancers plus chronic pancreatitis for various amino acids. Of these amino acids highlighted, our results for L-Phenylalanine, L-Proline, Alanine, L-Tryptophan and L-Glutamine are in agreement which also align with the KEGG pathways.

[0108] Lung Cancer

[0109] In lung cancer the top 100 features mapped to metabolites in Raman and mass- spectrometry correspond to: L-Phenylalanine, Pyruvate, Citric acid, Alanine, D- Fructose, L-Proline, Sarcosine, L-Tryptophan, Phenol, and Pantothenate.

[0110] The most important metabolite in the Raman lung cancer model is L- Phenylalanine which in our study is generally attributed to a decrease for lung cancer patients. Pyruvate is identified as the second most important metabolite in the Raman-based lung cancer model which is known as a main anaplerotic source in the TCA cycle. This also links the third most important metabolite in the lung cancer model Citric Acid. Table 7 shows the top metabolites from lung cancer matched with the KEGG pathways. The pathway Alanine, aspartate, and glutamate metabolism contains the highest number of hits with the top Raman metabolites which links with the affected TCA cycle.

[0111] Breast Cancer

[0112] The top 100 features mapped to Raman and mass-spectrometry correspond to the metabolites: L-Lysine, L-Methionine, L-Proline, L-Valine, Sucrose, L- Phenylalanine, Alanine, L-Tryptophan, Phenol, and Pantothenate.

[0113] We find L-Methionine is the second most important metabolite for breast cancer detection which is increased in breast cancer patients. Table 9 shows the top metabolite pathways from KEGG for the Raman model for breast cancer. Like with lung and pancreatic cancer, alanine, aspartate, and glutamate metabolism have the highest number of matches to Raman metabolites. This commonality could indicate a pathway affected across each cancer type.

[0114] Colorectal Cancer

[0115] In colorectal cancer, the top 10 metabolites matched with important Raman features and mass-spectrometry correspond to: L-Proline, Sucrose, Methylmalonate, D-Fructose, Sarcosine, L-Phenlyalanine, Alanine, L- Tryptophan, Citric acid, and Panthothenate.

[0116] The most important metabolite in the Raman-based model is L-Proline, which we find an increase in colorectal cancer patients. While the second most important metabolite found for colorectal cancer in the Raman model was Sucrose, aside from links to high dietary Sucrose being linked to colorectal cancer no other links to cancer have been found in the literature. Table 11 shows the KEGG pathways which match to the top Raman metabolites from the model. Unlike pancreatic, lung, and breast cancer which saw the most statistical significance on the alanine, aspartate and glutamate pathway, CRC sees the most significance with the valine, leucine, and isoleucine biosynthesis. However, alanine, aspartate and glutamate metabolism is the second most significant pathway identified for CRC. This could indicate this pathway is affected in most, or all cancers.

[0117] Paired Cancer Analyses

[0118] We now move to analyse the cancers against one another, excluding control samples from analysis. By building ML models between 2 cancer types, for example colorectal and pancreatic cancer, we can establish if Raman is able to elucidate differences between serum from cancer patients with different cancer types. We can then match any differences in the Raman peaks with the spectral library to establish which metabolites vary between cancer types. With knowledge of specific metabolites and pathways affected by different cancers comparatively, it may be possible to unlock more information about the fundamental alterations occurring in cancer metabolism and inform treatment decisions.

[0119] Pancreatic Cancer vs. Colorectal Cancer

[0120] Using the same methodology as the different cancer types vs. controls such that feature importance is obtained via 100 RF models and an average Gini importance plot taken. The subsequent top 100 features are then fed into a RF algorithm with LOOCV to evaluate performance classifying one cancer type from another.

[0121] Model performance in the case of pancreatic cancers vs. colorectal cancers is 90.9 % sensitivity (where colorectal cancer is considered the positive case), 70.5% specificity, 75.5% PPV, 88.6% NPV, 80.7% accuracy, and finally 0.815 AUC. Note this corresponds to a case where the probability has a threshold to achieve a sensitivity of at least 90%. Table 12 shows the performance with and without probability thresholding for a pancreatic vs. colorectal cancer model. Figure 4 shows the feature importance for the pancreatic vs. colorectal cancer model overlaid on a typical serum spectrum. The most important features selected correspond to the wavenumbers 644 cm-1 , 850 cm-1, 800 cm -1, 1050 cm-1 tentatively attributed to L-Lysine, L-Phenylalanine, L- Methionine, L-Proline, Phenol, Alanine, L-Tryptophan, Undine, Pantothenate, L-Alanine, GABA, L-Glutamate, L-Glutamine, L-Leucine, D-Fructose. Note arrows indicate how pancreatic cancer peaks vary from CRC, i.e. , each peak is increased for pancreatic cancer.

[0122] Pancreatic cancer vs. Lung cancer

[0123] Table 13 shows the model performance in cross-validation for a pancreatic cancer vs. lung cancer model with 31 age and sex matched patients. A sensitivity of 90.3%, specificity of 64.5%, PPV of 71.8% and NPV of 87.0% is achieved with a probability threshold adjusted to ensure a minimum of 90% sensitivity.

[0124] Distinct differences from pancreatic cancer and lung cancer can be seen in figure 5 with a feature importance plot over a typical serum spectrum. Random forest selected the most important features between pancreatic cancer and lung cancer as 1004 cm-1, 1020 cm-1, 625 cm-1, 1704 cm-1tentatively attributed to L-Phenylalanine, L-Tryptophan, Pyruvate, Methionine, L-Proline, L- Valine, Citric acid, L-Leucine, D-Fructose, Phenol, and Creatinine.

[0125] Pancreatic Cancer vs. Breast Cancer

[0126] Table 14 shows the performance in cross-validation for a pancreatic vs. breast cancer model with 38 matched patients. With probability thresholding, the model achieves a sensitivity of 91.9%, specificity of 67.6%, PPV of 73.9%, NPV of 89.3%, accuracy of 79.7%, and an AUC of 0.892. Note due to breast cancer patients only having female participants, age and gender propensity score matching has been unable to be implemented where breast is paired with other cancers with the exception of CRC. Matching is, however, still based on age where gender matching has not been possible.

[0127] Figure 6 shows the feature importance for a pancreatic vs. breast cancer machine learning model. The most important wavenumbers highlighted from random forest are 1600 cm-1, 644 cm-11200 cm-1tentatively attributed to L-Phenylalanine, L-Methionine, L-Tryptophan, L-Lysine, L-Proline, Phenol, Alanine, Uridine, Pantothenate, L-Alanine, L-Proline, L-Leucine, D-Fructose, Pyruvate, and Hypoxanthin.

[0128] Lung Cancer vs. Breast Cancer

[0129] Model performance for lung cancer vs. breast cancer is 94.6% sensitivity, 62.2% specificity, 71.4% PPV, 92.0% NPV, 78.4% accuracy, and an AUC of 0.893 with an 90% sensitivity threshold. Table 15 shows the full performance from cross- validation with and without probability thresholding.

[0130] Figure 7 shows the feature importance for a lung vs. breast cancer model. The most important wavenumbers include 1600 cm-1 , 644 cm-1, 1195 cm -1 which are tentatively attributed to L-Phenylalanine, L-Methionine, L-Tryptophan, L-Lysine, L-Proline, Phenol, Alanine, Uridine, Pantothenate, L-Alanine, L- Proline, L-Leucine, D-Fructose, Pyruvate, and Hypoxanthin. Notably these wavenumbers are the same as the most important wavenumbers in the paired model pancreatic vs. breast cancer. The common factor being breast cancer which indicates a similarity between pancreatic cancer and lung cancer.

[0131] Lung Cancer vs. Colorectal Cancer

[0132] Model performance for lung cancer vs. colorectal cancer is 90.3% sensitivity, 83.9% specificity, 84.9% PPV, 89.7% NPV, 87.1% accuracy, and an AUC of 0.898 with an 90% sensitivity threshold as shown in table 16. Figure 8 shows the feature importance for the lung vs colorectal model. The most important wavenumbers include 644 cm-1, 850 cm-1, 800 cm-1, 1050 cm-1tentatively attributed to L-Lysine, L-Phenylalanine, L-Methionine, L-Proline, Phenol, Alanine, L-Tryptophan, Undine, Pantothenate, L-Alanine, GABA, L-Glutamate, L-Glutamine, L-Leucine, D-Fructose.

[0133] Breast Cancer vs. Colorectal Cancer

[0134] Table 17 shows the model performance for colorectal cancer vs. breast cancer is 97.3% sensitivity, 83.8% specificity, 85.7% PPV, 96.9% NPV, 90.5% accuracy, and an AUC of 0.954 with an 90% sensitivity threshold.

[0135] Figure 9 shows the feature importance for the breast cancer vs colorectal cancer model. The most important wavenumbers are 644 cm-1, 1280 cm-1, 1004 cm-1, 1600 cm-1tentatively attributed to L-Lysine, L-Phenylalanine, L- Methionine, L-Proline, Phenol, Alanine, L-Tryptophan, Uridine, Pantothenate, L- Alanine, Sarcosine, and Citric acid.

[0136] Discussion

[0137] We have presented an application of Raman spectroscopy combined with machine learning for the detection of multiple cancer types from human blood serum. For cancer more generally, our model performed with a sensitivity of 90.1% and a specificity of 83.3%. These results alone are positive and suggest the potential for Raman to be used as either a screening procedure to flag patients requiring further investigation, or as a step prior to secondary care to prevent unnecessary and expensive exploratory investigations.

[0138] As a tool for specific cancer detection, we found our models performed highly for pancreatic cancer (where thresholds have been applied to ensure a minimum of 90% sensitivity) with sensitivity of 90.9% and specificity of 77.3%. For lung cancer the model achieves a sensitivity of 90.3% and a specificity of 64.5%. For breast cancer achieves a sensitivity of 92.1% and specificity of 65.8%. Finally, for colorectal cancer achieves a sensitivity of 91.3% and specificity of 44.0%. These results prove positive as a proof of principle.

[0139] Distinct differences in feature importance for the different cancer types. Notably for pancreatic cancer these are attributed to L-Phenylalanine, L-Proline, Alanine,

[0140] L-Tryptophan and L-Glutamine which is in line with another study which showed statistical significance with pancreatic cancers. For lung cancer the most important metabolites from the Raman model are L-Phenylalanine, Pyruvate, and Citric acid. For breast cancer the most important features from Raman are attributed to L-Lysine, L-Methionine, and L-Proline. With colorectal cancer, the most important feature from the Raman model include L-Proline, and Sucrose.

[0141] Raman spectroscopy combined with machine learning has great potential for detecting multiple cancer types.

[0142] Table 1

[0143] Table 2

[0144] Table 3

[0145] Table 4

[0146] Table 5

[0147] 36 Table 6

[0148] Table 7 Table 8

[0149] Table 9 Table 10

[0150] Table 11 Table 12

[0151] Table 13

[0152] Table 14

[0153] Table 15

[0154] Table 16

[0155] Table 17 Table 18

Claims

CLAIMS1. A method for determining cancer in a subject, the method comprising: i) performing Raman spectral analysis of a biological sample obtained from the subject to produce a test sample spectrum; ii) comparing the test sample spectrum obtained in step i) with at least one control sample Raman spectrum from a control subject, iii) wherein a difference between the test sample spectrum and the control sample spectrum at one or more wavenumbers within one or more ranges selected from the group comprising: about 622 cm-1, about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 758 cm-1, about 803 cm-1, about 829 cm-1, about 853 cm-1, about 880 cm-1, about 943 cm-1, about 957 cm-1, about 1004 cm-1, about 1031 cm-1, about 1051 cm-1, about 1062 cm-1, about 1084 cm-1, about 1127 cm-1, about 1158 cm-1, about 1175 cm-1, about 1272 cm-1, about 1341 cm-1, about 1402 cm-1, about 1411 cm-1, about 1449 cm-1, about 1498 cm-1, about 1521 cm-1, about 1556 cm-1, about 1587 cm-1, about 1607 cm-1, about 1657 cm-1is indicative of a subject suffering from cancer.

2. The method according to claim 1, wherein said subject is a mammal, and wherein said mammal is optionally a human, equine, canine, feline, porcine or any other domestic or agricultural species.

3. The method according claim 1 or claim 2, wherein said biological sample is a blood sample, optionally a liquid blood or blood derivative sample.

4. The method according to any of the preceding claims, wherein said cancer is selected from: nasopharyngeal cancer, synovial cancer, hepatocellular cancer, renal cancer, cancer of connective tissues, melanoma, lung cancer, bowel cancer, colorectal cancer, brain cancer,throat cancer, oral cancer, liver cancer, bone cancer, pancreatic cancer, choriocarcinoma, gastrinoma, pheochromocytoma, prolactinoma, T-cell leukemia / lymphoma, tonsil, spleen, neuroma, von Hippel-Lindau disease, Zollinger-Ellison syndrome, adrenal cancer, anal cancer, bile duct cancer, bladder cancer, ureter cancer, glioma, oligodendroglioma, neuroblastoma, meningioma, spinal cord tumor, osteochondroma, chondrosarcoma, Ewing's sarcoma, carcinoid, carcinoid of gastrointestinal tract, fibrosarcoma, breast cancer, muscle cancer, Paget's disease, cervical cancer, rectal cancer, esophagus cancer, gall bladder cancer, cholangioma cancer, head cancer, eye cancer, nasopharynx cancer, neck cancer, kidney cancer, Wilms' tumor, liver cancer, Kaposi's sarcoma, prostate cancer, testicular cancer, Hodgkin's disease, non-Hodgkin's lymphoma, skin cancer, mesothelioma, myeloma, multiple myeloma, ovarian cancer, endocrine pancreatic cancer, glucagonoma, parathyroid cancer, penis cancer, pituitary cancer, soft tissue sarcoma, retinoblastoma, small intestine cancer, stomach cancer, thymus cancer, thyroid cancer, trophoblastic cancer, hydatidiform mole, uterine cancer, endometrial cancer, vagina cancer, vulva cancer, acoustic neuroma, mycosis fungoides, insulinoma, carcinoid syndrome, somatostatinoma, gum cancer, heart cancer, lip cancer, meninges cancer, mouth cancer, nerve cancer, palate cancer, parotid gland cancer, peritoneum cancer, pharynx cancer, pleural cancer, salivary gland cancer, tongue cancer and tonsil cancer.

5. The method according to claim 4, wherein said cancer is selected from: colorectal, lung, pancreatic, and breast cancer.

6. The method according to any of the preceding claims, wherein said control and / or test samples are age, gender and weight-matched.

7. The method according to any of the preceding claims, wherein said Raman spectral analysis comprises irradiating said sample with a laser light source, wherein said laser light source is optionally configured toemit light in the wavelength band of infrared light, most preferably 785 nm.

8. The method according to claim 7, wherein the laser light source generates laser light having a power of from about 10 mW to about 1 W.

9. The method according to any of the preceding claims, wherein the, or each, spectra preferably undergoes one or more conventional preprocessing steps prior to or following the comparison step to reduce the noise associated with the one or more spectra to provide the, or each, processed spectra.

10. The method according to claim 9, wherein the one or more preprocessing step(s) are selected from the group comprising: data binning, smoothing, background subtraction such as Extended multiplicative scatter correction, and / or normalisation such as vector normalisation and / or baseline correction.

11. The method according to any of the preceding claims, wherein step iii) comprises or consists of observing a difference wherein an increase or decrease in signal intensity at the same wavenumber or a shift in position of signal maxima or minima between wavenumbers between the test sample spectrum produced from the sample and the control sample spectrum is indicative of a subject suffering from cancer.

12. The method according to any of the preceding claims, wherein said method comprises or consists of comparing the signal intensity at one, or any combination of, wavenumbers selected from the group comprising or consisting of: about 666 cm-1, about 682 cm-1, about 829 cm-1, about 853 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 803 cm-1, about 1004 cm-1, about 1127 cm-1, about 1175 cm-1, about 1341 cm-1, about 1402 cm-1, about 1411 cm-1, about 1498 cm-1,and about 1657 cm-1, wherein a difference compared with control is indicative of a subject suffering from cancer.

13. The method according to claim 12, wherein an increased signal intensity at a wavenumber selected from one or ideally both, of: about 1004 cm-1, about 1402 cm-1and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: about 666 cm-1, about 682 cm-1, 745 cm-1is indicative of a subject suffering from cancer.

14. The method according to any of claims 1 to 11 , wherein said method comprises or consists of comparing the signal intensity at one, or any combination of, wavenumbers selected from the group comprising or consisting of: about 644 cm-1, about 666 cm-1, about 682 cm-1, about 745 cm-1, about 853 cm-1, about 1004 cm-1, about 1031 cm-1, about 1062 cm-1, about 1084 cm-1, about 1127 cm-1, about 1158 cm-1, about 1175 cm-1, about 1341 cm-1, about 1402 cm-1, about 1411 cm-1, about 1498 cm-1, wherein a difference compared with control is indicative of a subject suffering from pancreatic cancer.

15. The method according to claim 14, wherein an increased signal intensity at a wavenumber selected from one or more, ideally all, of about 644 cm-1, about 853 cm-1, about 1004 cm-1, about 1031 cm-1, about 1062 cm-1, about 1084 cm-1, about 1127 cm-1, about 1175 cm-1, about 1341 cm-1, about 1402 cm-1, about 1411 cm-1, about 1498 cm-1and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: about 666 cm-1, about 682 crrr 1, about 745 cm-1, about 1158 cm-1is indicative of a subject suffering from pancreatic cancer.

16. The method according to any of claims 1 to 11, wherein said method comprises or consists of comparing the signal intensity at one, or any combination of, wavenumbers selected from the group comprising or consisting of: about 622 cm-1, about 644 cm-1, about 666 cm-1, about682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 803 cm-1, about 829 cm-1, about 853 cm-1, about 880 cm-1, about 1051 cmr 1, about 1175 cm-1, about 1272 cm-1, about 1402 cm-1, about 1411 cm’ 1, about 1498 cm-1, about 1657 cm-1wherein a difference compared with control is indicative of a subject suffering from lung cancer.

17. The method according to claim 16, wherein an increased signal intensity at a wavenumber selected from one or more, ideally all, of about 829 cm-1, about 853 cm-1, about 880 cm-1, about 1051 cm-1, about 1402 cm-1, about 1411 cm-1, and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: 622 cm-1, about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 803 cm-1, about 1175 cm-1, about 1272 cm-1, about 1498 cm-1, about 1657 cm-1is indicative of a subject suffering from lung cancer.

18. The method according to any of claims 1 to 11, wherein said method comprises or consists of comparing the signal intensity at one, or any combination of, wavenumbers selected from the group comprising or consisting of: the group comprising: about 622 cm-1, about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 758 cm-1, about 803 cm-1, about 829 cm-1, about 853 cm-1, about 1004 cm-1, about 1127 cm-1, about 1175 cm-1, about 1449 cm-1, about 1498 cm-1, about 1657 cm-1wherein a difference compared with control is indicative of a subject suffering from breast cancer.

19. The method according to claim 18, wherein an increased signal intensity at a wavenumber selected from one or more, ideally all, of : about 622 cm-1, about 758 cm-1, about 1004 cm-1, about 1127 cm-1, about 1175 cm-1, and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: about 644 cm-1, about 666 cm-1, about 682 cm-1, about 700 cm-1, about 719 cm-1, about 745 cm-1, about 803 cm-1, about 829 cm-1, about 853 cm-1, about 1449cm-1, about 1498 cm-1is indicative of a subject suffering from breast cancer.

20. The method according to any of claims 1 to 11 , wherein said method comprises or consists of comparing the signal intensity at one, or any combination of, wavenumbers selected from the group comprising or consisting of: the group comprising about 622 cm-1about 829 cm-1, about 943 cm-1, about 957 cm-1, about 1004 cm-1, about 1341 cm-1, about 1402 cm-1, about 1498 cm-1, about 1521 cm-1, about 1556 cm-1, about 1587 cm-1, about 1607 cm-1, about 1657 cm-1wherein a difference compared with control is indicative of a subject suffering from colorectal cancer.

21. The method according to claim 20, wherein an increased signal intensity at a wavenumber selected from one or more, ideally all, of : about 829 cm-1, about 1004 cm-1, about 1341 cm-1, about 1402 cm-1, about 1498 cm-1, about 1556 cm-1, about 1587 cm-1, about 1607 cm-1, about 1657 cm-1, and / or observing a decreased signal intensity at a wavenumber selected from one or more, ideally all, of: about 622 cm-1, about 943 cm-1, about 957 cm-1, about 1521 cm-1, is indicative of a subject suffering from colorectal cancer.

22. A method for monitoring the progression of cancer in a subject, said method comprising repeating one or more of the methods according to any one of claims 1 to 21 , and ideally the same, method(s) periodically.

23. A method for treating cancer, said method comprising performing the method according to any one or more of claims 1 to 21 and then, depending upon the outcome of the method, undertaking a suitable or selected course of treatment.

24. A kit for use in determining cancer in a biological sample from subject, said kit comprising:a. a Raman spetrometer for performing spectral analysis of a biological sample obtained from the subject; b. a processing unit, wherein said processing unit processes the test sample spectrum according to a method according to any one or more of claims 1 to 21 ; and c. an output unit arranged to provide an output indicative of a subject suffering from cancer in the subject according to a determination made by the processing unit in step b).

25. The kit according to claim 24, wherein said Raman spectrometer comprises a laser light source, wherein said laser light source: a. is arranged to emit light in the wavelength band of infrared light, most preferably 785 nm; and / or b. generates laser light having a power of from about 10 mW to about 1 W.

Images