Machine learning-based diagnostic classifiers for predicting cancer tissue of origin

Abstract

A method capable of determining a probability of a subject having one or multiple cancer tissues of origin is disclosed. The method includes inputting the expression profile of a miRNA set obtained from a biopsy sample such as serum sample of the subject into a classifier which is based on a machine learning model such as a support vector machine model. In certain embodiments, the method can be used to simultaneously determine the probability of the subject having each of multiple different cancer types, which can enhance the prediction accuracy.

Description

[0001] MACHINE LEARNING-BASED DIAGNOSTIC CLASSIFIERS FOR

[0002] PREDICTING CANCER TISSUE OF ORIGIN

[0003] CROSS REFERENCE TO RELATED APPLICATION

[0004] This application claims the benefit of U.S. provisional patent application U.S. 63 / 669,686 filed July 10, 2024, the content of which is incorporated by reference in its entirety.

[0005] REFERENCE TO SEQUENCE LISTING SUBMITTED ELECTRONICALLY

[0006] The content of the electronically submitted sequence listing, file name miRNAset- classifiers.xml, size 366,487 bytes, and date of creation July 7, 2025, filed herewith, is incorporated herein by reference in its entirety.

[0007] FIELD OF THE INVENTION

[0008] The present invention relates generally to the technical field of disease screening, detection and diagnosis, and more specifically relates to a method, a system, and a non- transitory storage medium for the detection of one or multiple human cancers.

[0009] BACKGROUND

[0010] Cancer screening is crucial for early diagnosis when effective treatment can be applied thereto. However, currently only four cancer types have anatomic site-specific screening tests recommended by the United States Preventive Services Task Force for lung, breast, colorectal, and ovarian cancers respectively (Nelson HD, et al. Ann Intern Med. 2016; US Preventive Services Task Force, et al. JAMA. 2018; US Preventive Services Task Force, Davidson KW, et al. JAMA. 2021; US Preventive Services Task Force, Krist AH, et al. JAMA. 2021). Most other cancer types do not have a recommended screening method at all leaving them for incidental diagnosis or late-stage diagnosis when symptoms develop at which time effective treatment is close to impossible. There have been efforts that intend to find an effective solution that can comprehensively screen or diagnose multiple cancer types in one single assay, especially from a liquid biopsy sample (e.g. blood, urine, saliva, sweat, etc.) in a noninvasive or minimally invasive manner.

[0011] Noninvasive multi-cancer early detection (MCED) with or without tissue of origin (TOO) has the potential to reduce cancer-related mortality by analyzing circulating cell-free nucleic acids and / or proteins in blood. Accurate prediction of cancer tissue of origin following a positive MCED test would guide selection of confirmatory tests, thereby expediting the definitive diagnosis and prompt initiation of the most appropriate treatment tailored to the specific cancer type.

[0012] SUMMARY

[0013] Described herein is the development of machine learning-based diagnostic classifier(s) that are capable of predicting tissue of origin for one or multiple cancer types with high accuracy using microRNAs (miRNAs).

[0014] In a first aspect, a method of developing a classifier capable of distinguishing a population of individuals having a cancer tissue of origin (i.e. a classifier developing method) is disclosed.

[0015] The classifier developing method substantially comprises the following two steps: (a) providing a training set from a population of subjects; and (b) constructing the classifier based on the training set. Herein, the training set comprises an expression profile of a miRNA set obtained from a biopsy sample from each of the population of subjects. The miRNA set comprises x different miRNAs, where x is a positive integer (i.e. x >0). The population of subjects comprises two subsets of subjects identified as having and not having the cancer tissue of origin, respectively. In other words, the population of subjects may comprise a first subset of subjects and a second subset of subjects. The first subset of subjects are identified as having the cancer from a specific tissue of origin, and the second subset of subjects are identified as having the cancer not from that tissue of origin.

[0016] According to some embodiments of the method, the step (b) constructing the classifier based on the training set comprises the sub-steps of: (bl) performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, thereby selecting an optimal set of miRNAs; and (b2) building a second machine learning model based on the optimal set of miRNAs. Herein the optimal set of miRNAs is substantially a subset of the miRNA set, and comprises y different miRNAs, where y < x.

[0017] Herein, the term "cancer tissue of origin" may be exchangeable to "cancer type", and refers to the tissue(s) where the cancer originates or comes from.

[0018] Herein the term "optimal set of miRNAs" is defined as a subset of miRNAs that gives a classification accuracy (i.e. ROC) that meets a certain condition (e.g. the accuracy passes a certain threshold) in the analysis using the first machine learning model (e.g. RF model) with recursive feature elimination (RFE). According to some embodiments of the disclosure, the optimal set of miRNAs is selected or obtained when a highest level of ROC value is produced in the analysis using the first machine learning model with RFE, and further in certain embodiments, the first machine learning model is a RF model.

[0019] Herein in sub-step (bl), the first machine learning model is optionally selected from a method including but not limited to, a random forest (RF) model, a support vector machine (SVM), a neural network, classification and regression trees (CART), adaptive boosting, bagging, discriminant analysis, logistic regression, gradient boost, generalized linear models, generalized additive models, k-nearest neighbors, partial least squares, or ridge regression, and preferably the first machine learning model is a RF model.

[0020] Optionally in the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection thereby selecting an optimal set of miRNAs, the optimal set of miRNAs is selected using the first machine learning model along with / 77-fold cross validation, where m is a positive integer no less than 5, such as 5, 6, 7, 8, 9, 10, etc.

[0021] Further optionally in the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection thereby selecting an optimal set of miRNAs, the optimal set of miRNAs is selected based on a highest level of ROC value obtained when performing the RFE based on the first machine learning model.

[0022] Optionally in the sub-step (b2) of building a second machine learning model based on the optimal set of miRNAs, the second machine learning model is selected from a method including but not limited to, a support vector machine (SVM), a random forest (RF) model, neural network, classification and regression trees (CART), adaptive boosting, bagging, discriminant analysis, logistic regression, gradient boost, generalized linear models, generalized additive models, k-nearest neighbors, partial least squares, or ridge regression, and preferably the second machine learning model is SVM.

[0023] Further optionally the sub-step (b2) of building a second machine learning model based on the optimal set of miRNAs comprises / / -fold cross validation for parameter optimization, where n is a positive integer no less than 5, such as 5, 6, 7, 8, 9, 10, etc.

[0024] According to some embodiments of the method, the step (2) of constructing the classifier based on the training set further comprises, prior to the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, a sub-step of performing a limma analysis over the training set for a preliminary feature selection, thereby selecting top miRNAs. Herein the top miRNAs are substantially a subset of the miRNA set, and the optimal set of miRNAs is substantially a subset of the top miRNAs. In other words, the top miRNAs comprise z different miRNAs, where y < z < x. Accordingly in the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, thereby selecting an optimal set of miRNAs, the recursive feature elimination (RFE) is performed over the top miRNAs.

[0025] Herein, the limma analysis is substantially used to filter a set of miRNAs based on their statistical significances for distinguishing the cancer tissue of origin, such as in the differentiation of those having the cancer tissue of origin from those not having the cancer tissue of origin.

[0026] According to some embodiments of the method, the step (2) of constructing the classifier based on the training set further comprises, prior to the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, a sub-step of performing a quality control over the training set, comprising: filtering out low coefficient of variation (CV) miRNAs and / or highly correlated miRNAs from the miRNA set.

[0027] According to some embodiments of the method, the step (2) of constructing the classifier based on the training set further comprises, prior to the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, the sub-steps of: performing a quality control over the training set, comprising: filtering out low coefficient of variation (CV) miRNAs and / or highly correlated miRNAs from the miRNA set to thereby obtain a cleaned training set; and performing a limma analysis over the cleaned training set for a preliminary feature selection, thereby selecting top miRNAs. Herein the top miRNAs comprises z different miRNAs, where y < z < x. Accordingly in the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, thereby selecting an optimal set of miRNAs, the recursive feature elimination (RFE) is performed over the top miRNAs.

[0028] In any embodiments of the classifier developing method as described above, the biopsy sample can be selected from a body liquid sample, including but not limited to, a plasma sample, a serum sample, a urine sample, a saliva sample, or a sweat sample.

[0029] In a second aspect, the present disclosure further provides a method for determining the probability of an individual having a cancer tissue of origin. This method is substantially a diagnostic method for determining how likely an individual has a cancer tissue of origin.

[0030] The method substantially comprises the following two steps: (a) obtaining an expression profile of a predetermined miRNA set based on a liquid biopsy sample from the individual; and (b) inputting the expression profile of the predetermined miRNA set into a predetermined classifier to thereby obtain the probability of the individual having the cancer tissue of origin. Herein in the diagnostic method, the predetermined miRNA set and the predetermined classifier can be the optimal set of miRNAs and the classifier that are determined by the classifier developing method as described above in the first aspect. The individual is of the same species as the population of subjects, who can be a mammal and can preferably be a human being. The liquid biopsy sample can be selected from a body liquid sample, including but not limited to, a plasma sample, a serum sample, a urine sample, a saliva sample, or a sweat sample.

[0031] In the diagnostic method, there can be different embodiments for the cancer tissue of origin, which can be any cancer type, such as biliary tract cancer, bladder cancer, breast cancer, colorectal cancer, esophageal cancer, gastric cancer, glioma, liver cancer, lung cancer, ovarian cancer, pancreatic cancer, prostate cancer, and sarcoma, but is not limited by these above listed cancers.

[0032] According to some embodiments, the diagnostic method is for determining the probability of an individual having biliary tract cancer, and the predetermined miRNA set comprises at least one, and preferably all, of the miRNAs with SEQ ID NOS: 1-75.

[0033] According to some other embodiments, the diagnostic method is for determining the probability of an individual having bladder cancer, and the predetermined miRNA set comprises at least one, and preferably all, of the miRNAs with SEQ ID NO: SEQ ID NOS: 1, 3, 5, 14, 51, and 76-120.

[0034] According to yet some other embodiments, the diagnostic method is for determining the probability of an individual having breast cancer, and the predetermined miRNA set comprises at least one, and preferably all, of the miRNAs with SEQ ID NOS: 14, 77-78, 80, 82, 91-92, and 121-188.

[0035] According to yet some other embodiments, the diagnostic method is for determining the probability of an individual having colorectal cancer, and the predetermined miRNA set comprises at least one, and preferably all, of the miRNAs with SEQ ID NOS: 1, 17, 37, 77, 89, 112, 116, 121, 128, 130-132, 135, 151, and 189-224.

[0036] According to yet some other embodiments, the diagnostic method is for determining the probability of an individual having esophageal cancer, and the predetermined miRNA set comprises at least one, and preferably all, of the miRNAs with SEQ ID NOS: 1, 6, 14, 54, 82, 93, 95, 97, 100, 104, 111, 120, 130, 141, 143, 147, 154, 156, 158, 186, 194, 211, 217-218, and 225-250.

[0037] According to yet some other embodiments, the diagnostic method is for determining the probability of an individual having gastric cancer, and the predetermined miRNA set comprises at least one, and preferably all, of the miRNAs with SEQ ID NOS: 6, 37, 54, 76, 78- 80, 82, 84-86, 90-91, 93, 95, 97, 100-101, 111-112, 120, 122, 127-128, 130, 132, 138, 141, 143-144, 147, 154, 156, 158, 174, 178, 186, 189, 191, 194-196, 205, 211, 225-232, 235-236, 238, 242, 244, and 251-293.

[0038] According to yet some other embodiments, the diagnostic method is for determining the probability of an individual having glioma, and the predetermined miRNA set comprises at least one, and preferably all, of the miRNAs with SEQ ID NOS: 6, 17, 120, 128, 154, 156, 158, 195, 225-226, 228, 247, and 294-306.

[0039] According to yet some other embodiments, the diagnostic method is for determining the probability of an individual having liver cancer, and the predetermined miRNA set comprises at least one, and preferably all, of the miRNAs with SEQ ID NOS: 1, 6, 9, 12, 14, 26, 28, 35, 39, 57, 76-78, 83, 89, 112, 120, 128, 132, 137, 151, 174, 189, 194, 214, 236, 263, 272, 280, 293, and 307-326.

[0040] According to yet some other embodiments, the diagnostic method is for determining the probability of an individual having lung cancer, and the predetermined miRNA set comprises miRNAs at least one, and preferably all, of the miRNAs with SEQ ID NOS: 3-4, 6, 37, 76, 79, 82, 86, 91, 93, 101, 112, 120, 122, 128, 130, 132, 141, 154, 156, 160, 165, 172, 177-178, 186, 189, 191, 194-196, 209, 211, 217-218, 225-226, 230, 232, 257, 263, 285-286, 299, and 326-331.

[0041] According to yet some other embodiments, the diagnostic method is for determining the probability of an individual having ovarian cancer, and the predetermined miRNA set comprises miRNAs at least one, and preferably all, of the miRNAs with SEQ ID NOS: 12, 76, 79, 101, 128, 130, 143, 154, 186, 195, 211, 225-226, 232, 236, 247, 254, 268, 280, 302, 329, and 332-335.

[0042] According to yet some other embodiments, the diagnostic method is for determining the probability of an individual having pancreatic cancer, and the predetermined miRNA set comprises at least one, and preferably all, of the miRNAs with SEQ ID NO: 1, 14, 25, 28, 48, 77, 89, 95, 114, 116, 120, 130, 140-141, 143, 147, 186, 194-195, 211, 218, 227-229, 231-232, 240, 244, 276, 293, and 336-355.

[0043] According to yet some other embodiments, the diagnostic method is for determining the probability of an individual having prostate cancer, and the predetermined miRNA set comprises at least one, and preferably all, of the miRNAs with SEQ ID NOS: 1, 10, 14, 23, 25, 77, 79, 83, 85-86, 89, 104, 120, 122, 143, 165-166, 172, 217, 239, 248, 254, 263, 283, 285, 292, 299, 300, 315, 328, 330, and 356-374. According to yet some other embodiments, the diagnostic method is for determining the probability of an individual having sarcoma, wherein the predetermined miRNA set comprises at least one, and preferably all, of the miRNAs with SEQ ID NO: 4, 6, 12, 18, 25, 54, 58, 69-70, 78, 84-86, 89, 91, 95, 100, 108, 118, 120-121, 130, 132-133, 141, 143, 146-147, 149, 151, 154, 161, 166, 172, 181, 186-187, 194, 205, 217-218, 225, 227-232, 234, 236, 238- 241, 243-245, 249, 253, 263, 272, 286, 291, 293, 299-300, 307, 313-314, 319, 322, 327, 369, and 375-426.

[0044] In a third aspect, the present disclosure further provides a method for determining a cancer tissue of origin for an individual. This method is substantially a diagnostic method for comprehensively estimating how likely an individual has a certain cancer tissue of origin among several different cancer tissues of origin.

[0045] The method substantially comprises the following three steps (a) obtaining an expression profile of a selected miRNA set based on a liquid biopsy sample from the individual; (b) inputting the expression profile of the selected miRNA set into a classifying algorithm comprising at least two predetermined classifiers, each for a unique cancer tissue of origin, thereby obtaining a probability of the individual having the unique cancer tissue of origin corresponding to each predetermined classifier; and (c) estimating the predicted cancer tissue of origin for the individual based on the probability of the individual having the unique cancer tissue of origin corresponding to the each predetermined classifier. Herein, each predetermined classifier can be obtained based on the classifier developing method as described above in the first aspect, and the selected miRNA set comprises a combination of an optimal set of miRNAs corresponding to each predetermined classifier as selected by the classifier developing method. Herein, the individual is of the same species as the population of subjects, who can be a mammal and can preferably be a human being. The liquid biopsy sample can be selected from a body liquid sample, including but not limited to, a plasma sample, a serum sample, a urine sample, a saliva sample, or a sweat sample.

[0046] According to certain embodiments of the method, step (c) of estimating the predicted cancer tissue of origin for the individual based on the probability of the individual having the unique cancer tissue of origin corresponding to the each predetermined classifier comprises the following two sub-steps: (cl) ranking the unique cancer tissue of origin corresponding to each predetermined classifier based on a corresponding probability Pi, thereby obtaining ranked cancer tissues of origin; and (c2) estimating the predicted cancer tissue of origin for the individual based on the ranked cancer tissues of origin. In other words, the above can be described as follows: n is the total number of predetermined classifiers in the classifying algorithm, with Pi corresponding to any cancer tissue of origin z (0 < z < n). In sub-step (cl), the unique cancer tissue of origin corresponding to each predetermined classifier is ranked based on a corresponding probability Pi, and after ranking, the ranked cancer tissues of origin can be described as Pi < P2 < ... < Pnin a highest-to-lowest rank.

[0047] According to some embodiments, step (c2) of estimating the predicted cancer tissue of origin for the individual based on the ranked cancer tissues of origin comprises the following sub-steps: (c2-l) picking up top m cancer tissues of origin (m < ri) and (c2-2) determining that the individual has the cancer tissue of origin based on the top m cancer tissues of origin. In other words, the estimation of how likely an individual has a certain cancer tissue of origin is substantially based on a comprehensive estimation of multiple cancer types.

[0048] According to some embodiments, m = 1 (i.e., the estimation is only on the top 1 cancer tissue of origin, or the highest ranked cancer tissue of origin). Herein optionally, the cancer tissue of origin is bladder cancer, breast cancer, esophageal cancer, gastric cancer, glioma, lung cancer, or prostate cancer, and the method is capable of reaching an accuracy equal to or more than 70%. Further optionally, the cancer tissue of origin is breast cancer, gastric cancer, glioma, or prostate cancer, and the method is capable of reaching an accuracy equal to or more than 80%. Further optionally, the cancer tissue of origin is breast cancer or glioma, and the method is capable of reaching an accuracy equal to or more than 90%.

[0049] According to some other embodiments, m = 2 (i.e., the estimation is only on the top 2 cancer tissues of origin, or the highest and the second highest ranked cancer tissues of origin), and the method is capable of reaching an accuracy equal to or more than 90% for certain cancer tissues of origin.

[0050] According to yet some other embodiments, m =3. Herein optionally, the cancer tissue of origin is bladder cancer, breast cancer, esophageal cancer, gastric cancer, glioma, liver cancer, lung cancer, or prostate cancer, and the method is capable of reaching an accuracy equal to or more than 90%. Further optionally, the cancer tissue of origin is bladder cancer, breast cancer, gastric cancer, glioma, lung cancer, or prostate cancer, and the method is capable of reaching an accuracy equal to or more than 95%. Further optionally, the cancer tissue of origin is gastric cancer, glioma, or lung cancer, and the method is capable of reaching an accuracy equal to or more than 98%.

[0051] According to some embodiments, the selected miRNA set comprises at least one, and possibly all, of the miRNAs whose sequences are set forth in SEQ ID NO: 1-426.

[0052] In a fourth aspect, the present disclosure further provides a system that can be used for implementing the method as described above in the first, second, and third aspects. Herein the system is substantially a computer system that comprises: a processor; and a non-transitory storage medium containing program instructions for execution by the processor. Herein the program instructions are configured to cause the processor to execute various steps of the classifier developing method as described above in the first aspect, and to execute various steps of the diagnostic method as described above in the second and third aspects of the disclosure.

[0053] In a fifth aspect, a non-transitory storage medium is further provided, which is configured to store computer-executable program instructions which, when executed by a processor, cause the processor to execute various steps of the cancer diagnostic model development method according to any embodiment as described above in the first aspect.

[0054] The terms as used throughout the disclosure are defined as follows.

[0055] In general terms, a “subject” or "individual" as used herein means a mammal such as a primate including a human and a chimpanzee, a pet animal including a dog and a cat, a livestock animal including cattle, a horse, sheep, and a goat, and a rodent including a mouse and a rat. It is to be noted that the whole disclosure concerns more specifically human subjects, but can optionally be applied to other non-human mammals as well. Herein, the subject or individual may be of any condition, who may have already been determined to be a cancer patient but the cancer tissue of origin has not yet been determined or is unknown, or may have not yet been estimated or determined for the cancer disease condition.

[0056] Unless indicated or defined otherwise, the terms or abbreviations such as "gene", “nucleic acid”, “nucleotide”, “polynucleotide”, “DNA”, “RNA”, and “miRNA” abide by common use in the art.

[0057] As used herein, the term “polynucleotide” is interchangeable with “nucleic acid”, and refers to as a nucleic acid including all of RNA, DNA, and RNA / DNA (chimera). The DNA includes all of cDNA, genomic DNA, and synthetic DNA. The RNA includes all of total RNA, mRNA, rRNA, miRNA, siRNA, snoRNA, snRNA, non-coding RNA and synthetic RNA. As used herein, the term “fragment” is a polynucleotide having a nucleotide sequence having a consecutive portion of a polynucleotide and desirably has a length of 15 or more nucleotides, e.g. 15, 16, 17, 18, 19, etc. nucleotides.

[0058] Within the scope of the whole disclosure, the term “microRNA" or "miRNA” is intended to mean a 15- to 25-nucleotide non-coding RNA that is transcribed as an RNA precursor having a hairpin-like structure, cleaved by a dsRNA-cleaving enzyme which has RNase III cleavage activity, integrated into a protein complex called RISC, and involved in the suppression of translation of mRNA, unless otherwise specified. The term “miRNA” as used herein includes not only a “miRNA” represented by a particular nucleotide sequence (or SEQ ID NO) but a precursor of the “miRNA” (pre-miRNA or pri-miRNA), and miRNAs having biological functions equivalent thereto, for example, a congener (i.e., a homolog or an ortholog), a variant (e.g., a genetic polymorph), and a derivative. Such a precursor, a congener, a variant, or a derivative can be specifically identified using miRBase Release 20 (Kozomara A & Griffiths-Jones S. Nucleic Acids Res. 2011 Jan;39(Database issue):D152-7.), and examples thereof can include an “miRNA” having a nucleotide sequence hybridizing under stringent conditions described later to a complementary sequence of any particular nucleotide sequence represented by any of SEQ ID NOs: 1 to 426. The term “miRNA” as used herein may be a gene product of a miR gene. Such a gene product includes a mature miRNA (e.g., a 15- to 25- nucleotide or 19- to 25-nucleotide non-coding RNA involved in the suppression of translation of mRNA as described above) or a miRNA precursor (e.g., pre-miRNA as described above).

[0059] As used herein, the term “probe” includes a polynucleotide that is used for specifically detecting an RNA resulting from the expression of a gene or a polynucleotide derived from the RNA, and / or a polynucleotide complementary thereto.

[0060] As used herein, the term “primer”, or “amplification primers” includes a polynucleotide that specifically recognizes and amplifies an RNA resulting from the expression of a gene or a polynucleotide derived from the RNA, and / or a polynucleotide complementary thereto.

[0061] In this context, the complementary polynucleotide (complementary strand or reverse strand) means a polynucleotide in a complementary base relationship based on A:T (U) and G:C base pairs with the full-length sequence of a polynucleotide consisting of a nucleotide sequence defined by any of SEQ ID NOs: 1 to 426 or a nucleotide sequence derived from the nucleotide sequence by the replacement of the nucleotide "U" (or "u") with the nucleotide "T" (or "t"), or a partial sequence thereof (here, this full-length or partial sequence refers to as a plus strand for the sake of convenience). However, such a complementary strand is not limited to a sequence completely complementary to the nucleotide sequence of the target plus strand and may have a complementary relationship to an extent that permits hybridization under stringent conditions to the target plus strand.

[0062] As used herein, the term “stringent conditions” refers to conditions under which a nucleic acid probe hybridizes to its target sequence to a larger extent (e.g., a measurement value equal to or larger than a mean of background measurement values + a standard deviation of the background measurement values x 2) than that for other sequences. The stringent conditions are dependent on a sequence and differ depending on an environment where hybridization is performed. A target sequence complementary 100% to the nucleic acid probe can be identified by controlling the stringency of hybridization and / or washing conditions. Specific examples of the “stringent conditions” will be mentioned later.

[0063] As used herein, the term “Tm value” means a temperature at which the double-stranded moiety of a polynucleotide is denatured into single strands so that the double strands and the single strands exist at a ratio of 1 : 1.

[0064] As used herein, the term “variant” means, in the case of a nucleic acid, a natural variant attributed to polymorphism, mutation, or the like; a variant containing the deletion, substitution, addition, or insertion of 1, 2, or 3 or more nucleotides in a nucleotide sequence represented by any of SEQ ID NOs: 1 to 426 or a nucleotide sequence derived from the nucleotide sequence by the replacement of the nucleotide "U" (or "u") with the nucleotide "T" (or "t"), or a partial sequence thereof; a variant containing the deletion, substitution, addition, or insertion of 1 or 2 or more nucleotides in a nucleotide sequence of a premature miRNA of a sequence represented by any of SEQ ID NOs: 1 to 426 or a nucleotide sequence derived from the nucleotide sequence by the replacement of the nucleotide "U" (or "u") with the nucleotide "T" (or "t"), or a partial sequence thereof; a variant that exhibits % identity of approximately 90% or higher, approximately 95% or higher to each of these nucleotide sequences or the partial sequences thereof; or a nucleic acid hybridizing under the stringent conditions defined above to a polynucleotide or an oligonucleotide comprising each of these nucleotide sequences or the partial sequences thereof. A variant can be prepared by use of a well-known technique such as site-directed mutagenesis or PCR-based mutagenesis.

[0065] The term “percent(%) identity” can be determined with or without an introduced gap, using a protein or gene search system based on BLAST or FASTA described above (Zhang Z, et al. J Comput Biol. 2000; Altschul SF, et al. J Mol Biol. 1990; Pearson WR, et al. Proc Natl AcadSci USA. 1988).

[0066] The term “derivative” is meant to include a modified nucleic acid, for example, a derivative labeled with a fluorophore or the like, a derivative containing a modified nucleotide (e.g., a nucleotide containing a group such as halogen, alkyl such as methyl, alkoxy such as methoxy, thio, or carboxymethyl, and a nucleotide that has undergone base rearrangement, double bond saturation, deamination, replacement of an oxygen molecule with a sulfur atom, etc.), PNA (peptide nucleic acid; Nielsen PE, et al. Science. 1991), and LNA (locked nucleic acid; Obika S, et al. Tetrahedron Letters. 1998) without any limitation.

[0067] The “nucleic acid” capable of specifically binding to a polynucleotide selected from the miRNAs described above is a synthesized or prepared nucleic acid and specifically includes a “nucleic acid probe” or a “primer”. The “nucleic acid” is utilized directly or indirectly for detecting the presence or absence of cancer in a subject, for diagnosing the severity, the degree of amelioration, or the therapeutic sensitivity of cancer, or for screening for a candidate substance useful in the prevention, amelioration, or treatment of cancer. The “nucleic acid” includes a nucleotide, an oligonucleotide, and a polynucleotide capable of specifically recognizing and binding to a transcript represented by any of SEQ ID NOs: 1 to 426, or a synthetic cDNA nucleic acid thereof in vivo, particularly, in a sample such as a body fluid (e.g., blood or urine), in relation to the development of cancer. The nucleotide, the oligonucleotide, and the polynucleotide can be effectively used as probes for detecting the aforementioned gene expressed in vivo, in tissues, in cells, or the like on the basis of the properties described above, or as primers for amplifying the aforementioned gene expressed in vivo.

[0068] The term “detection” as used herein is interchangeable with the term “examination”, “measurement”, or “detection or decision support”. As used herein, the term “evaluation” is meant to include diagnosis or evaluation support on the basis of examination results or measurement results.

[0069] As used within the scope of the disclosure, each of the terms “ / ?-value”, “accuracy”, “AUC”, “sensitivity”, and “specificity” is generally to be understood to have the common definition that is well appreciated by people skilled in the art, and is specifically defined as follows:

[0070] As used herein, the term “ / ?-value”, "p value", "p-value", or alike, refers to a probability at which a more extreme statistic than that actually calculated from data under a null hypothesis is observed in a statistical test. Thus, smaller “ / ? value” means more significant difference between subjects to be compared.

[0071] The term “AUC” means area under the curve of a Receiver Operating Characteristic curve. The term “accuracy” means a value of (the number of true positives + the number of true negatives) / (the total number of cases). The accuracy indicates the ratio of samples that were correctly identified to all samples and serves as a primary index to evaluate detection performance. Within the scope of this present disclosure, the term refers to the percentage of the correct predictions out of the total predictions made.

[0072] As used herein, the term “sensitivity” means a value of (the number of true positives) / (the number of true positives + the number of false negatives). High sensitivity allows cancer to be detected, leading to clinical treatment interventions.

[0073] As used herein, the term “specificity” means a value of (the number of true negatives) / (the number of true negatives + the number of false positives). High specificity prevents needless extra examination for healthy subjects misjudged as being cancer patients, leading to reduction in burden on patients and reduction in medical expense.

[0074] As used herein, the term "machine learning" refers to a process that enables computers to learn autonomously by identifying patterns and making data-based decisions, which typically uses mathematical models to help computers learn from data without direct instruction. By using algorithms to identify patterns, machine learning creates data models that make predictions, and predictions improve as more data is collected and analyzed, much like how humans learn from experience.

[0075] As used herein, the term "classifier" refers to an algorithm or system used in machine learning that automatically sorts or categorizes data into one or more classes based on learned characteristics from input data, which can be used to predict the class of new data by analyzing the features of the input. Within the scope of this disclosure, it refers to a trained machine learning model that can, based on the expression profile of a predetermined miRNA set (i.e. feature vector) obtained from a biopsy sample of a subject (i.e. input), predict the category of such input data (e.g. the cancer tissue of origin).

[0076] As used herein, the term "receiver operating characteristic" or "ROC" refers to a graphical plot used to show the diagnostic ability of binary classifiers by plotting the true positive rate against the false positive rate. The area under the ROC curve (i.e. "AUC") measures the performance of machine learning algorithms and helps in analyzing classification algorithms.

[0077] As used herein, the term "recursive feature elimination" or "RFE", refers to a feature selection algorithm used to select a subset of the most relevant features from a dataset, which works by recursively removing the least important features and refitting the model until the desired number of features is reached.

[0078] Unless specified elsewhere, the following summarizes the available technologies that can be used for the determination of the expression profile of the miRNA biomarker set.

[0079] It is to be noted that determination of the expression profile of the miRNA biomarker set substantially includes the determination of the expression level of each and every miRNA contained in the miRNA biomarker set. Preferably, expression levels for all of the miRNA contained in the miRNA biomarker set can be determined simultaneously in one single experiment that is well-controlled. Yet optionally, it is possible that expression levels of these miRNAs are determined in more than one experiment and by different experiment procedure. As used herein, measuring or detecting the expression of any of the miRNAs contained in the miRNA biomarker set comprises measuring or detecting any nucleic acid transcript corresponding to the miRNA.

[0080] Typically, expression can be detected or measured on the basis of miRNA or corresponding reverse transcribed cDNA levels. Any quantitative or qualitative method for measuring RNA levels, or cDNA levels can be used. Suitable methods of detecting or measuring miRNA or cDNA levels include, for example, Northern Blotting, microarray analysis, RNA-sequencing, RNA in-situ hybridization, or a nucleic acid amplification procedure, such as reverse-transcription PCR (RT-PCR) or real-time RT-PCR, also known as quantitative RT-PCR (qRT-PCR), or digital RT-PCR. Such methods are well known in the art. Other techniques include digital, multiplexed analysis of gene expression, such as the nCounter® (NanoString Technologies, Seattle, WA) gene expression assays, which are further described in US20100112710 and US20100047924.

[0081] Detecting a nucleic acid of interest generally involves hybridization between a target (e.g. miRNA or cDNA) and a probe. Sequences of the miRNAs used in various cancer gene expression profiles are known. Therefore, one of skills in the art can readily design hybridization probes for detecting those miRNAs (see e.g., Sambrook J, et al. 2001. Molecular Cloning: A Laboratory Manual. 3rd ed). For example, polynucleotide probes that specifically bind to the miRNA transcripts described herein (or cDNA synthesized therefrom) can be created using the nucleic acid sequences of the miRNA or cDNA targets themselves by routine techniques (e.g., PCR or synthesis). As used herein, the term “probe” means a part or portion of a polynucleotide sequence comprising about 10 or more contiguous nucleotides, about 15 or more contiguous nucleotides, about 20 or more contiguous nucleotides. In certain embodiments, the polynucleotide probes will comprise 10 or more nucleic acids, 15 or more nucleic acids, or 20 or more nucleic acids. In order to confer sufficient specificity, the probe may have a sequence identity to a complement of the target sequence of about 90% or more, such as about 95% or more (e.g., about 98% or more or about 99% or more) as determined, for example, using the well-known Basic Local Alignment Search Tool (BLAST) algorithm (available through the National Center for Biotechnology Information (NCBI), Bethesda, MD).

[0082] Each probe may be substantially specific for its target, to avoid any cross hybridization and false positives. An alternative to using specific probes is to use specific reagents when deriving materials from transcripts (e.g., during cDNA production, or using target-specific primers during amplification). In both cases specificity can be achieved by hybridization to portions of the targets that are substantially unique within the group of miRNAs being analyzed, for example hybridization to the polyA tail would not provide specificity. If a target has multiple splice variants, it is possible to design a hybridization reagent that recognizes a region common to each variant and / or to use more than one reagent, each of which may recognize one or more variants.

[0083] Stringency of hybridization reactions is readily determinable by one of ordinary skill in the art, and generally is an empirical calculation dependent upon probe length, washing temperature, and salt concentration. In general, longer probes may require higher temperatures for proper annealing, while shorter probes may require lower temperatures. Hybridization generally depends on the ability of denatured nucleic acid sequences to reanneal when complementary strands are present in an environment below their melting temperature. The higher the degree of desired homology between the probe and hybridizable sequence, the higher the relative temperature that can be used. As a result, it follows that higher relative temperatures would tend to make the reaction conditions more stringent, while lower temperatures less so.

[0084] “Stringent conditions” or “high stringency conditions,” as defined herein, are identified by, but not limited to, those that: (1) use low ionic strength and high temperature for washing, for example 0.015 M sodium chloride / 0.0015 M sodium citrate / 0.1% sodium dodecyl sulfate at 50°C; (2) use during hybridization a denaturing agent, such as formamide, for example, 50% (v / v) formamide with 0.1% bovine serum albumin / 0.1% Ficoll / 0.1% polyvinylpyrrolidone / 50 mM sodium phosphate buffer at pH 6.5 with 750 mM sodium chloride, 75 mM sodium citrate at 42°C; or (3) use 50% formamide, 5*SSC (0.75 M NaCl, 0.075 M sodium citrate), 50 mM sodium phosphate (pH 6.8), 0.1% sodium pyrophosphate, 5 Denhardt's solution, sonicated salmon sperm DNA (50pg / ml), 0.1% SDS, and 10% dextran sulfate at 42°C, with washes at 42°C in 0.2* SSC (sodium chloride / sodium citrate) and 50% formamide at 55°C, followed by a high-stringency wash of 0. l x SSC containing EDTA at 55°C. “Moderately stringent conditions” are described by, but not limited to, those in Sambrook J, et al. 2001. Molecular Cloning: A Laboratory Manual. 3rd ed., and include the use of washing solution and hybridization conditions (e.g., temperature, ionic strength and % SDS) less stringent than those described above. An example of moderately stringent conditions is overnight incubation at 37°C in a solution comprising: 20% formamide, 5xSSC (150 mM NaCl, 15 mM trisodium citrate), 50 mM sodium phosphate (pH 7.6), 5x Denhardt's solution, 10% dextran sulfate, and 20 mg / mL denatured sheared salmon sperm DNA, followed by washing the filters in l x SSC at about 37-50°C. The skilled artisan will recognize how to adjust the temperature, ionic strength, etc. as necessary to accommodate factors such as probe length and the like. In certain embodiments, microarray analysis, Northern blot, RNA in-situ hybridization, or a PCR-based method is used. In this respect, measuring the expression of the foregoing miRNAs in a biological sample can comprise, for instance, contacting a sample containing or suspected of containing cancer cells with polynucleotide probes specific to the miRNAs of interest, or with primers designed to amplify a portion of the miRNAs of interest, and detecting binding of the probes to the nucleic acid targets or amplification of the nucleic acids, respectively. Detailed protocols for designing PCR primers are known in the art. In certain embodiments, miRNAs obtained from a sample may be subjected to qRT-PCR. Reverse transcription may occur by any methods known in the art, such as through the use of an Omniscript RT Kit (Qiagen). The resultant cDNA may then be amplified by any amplification technique known in the art. miRNA expression may then be analyzed through the use of, for example, control samples as described below. As described herein, the over- or underexpression of miRNAs relative to controls may be measured to determine a miRNA expression profile for an individual biological sample. Similarly, detailed protocols for preparing and using microarrays to analyze miRNA expression are known in the art and described herein.

[0085] As used herein, RNA-sequencing (RNA-seq), also called Whole Transcriptome Shotgun Sequencing, refers to any of a variety of high-throughput sequencing techniques used to detect the presence and quantity of RNA transcripts in real time (Wang Z, et al. Nat Rev Genet. 2009). RNA-seq can be used to reveal a snapshot of a sample’ s miRNAs from a genome at a given moment in time. In certain embodiments, miRNA is converted to cDNA fragments via reverse transcription prior to sequencing, and, in certain embodiments, miRNA can be directly sequenced without conversion to cDNA. Adaptors may be attached to the 5’ and / or 3’ ends of the miRNAs, and the miRNA or cDNA may optionally be amplified, for example by PCR. The fragments are then sequenced using high-throughput sequencing technology, such as, for example, those available from Roche (e.g., the 454 platform), Illumina, Inc., and Applied Biosystem (e.g., the SOLiD system).

[0086] It is noted that in interpreting both the specification and the claims, all terms should be interpreted in the broadest possible manner consistent with the context. In particular, the terms “comprise” and “comprising” should be interpreted as referring to elements, components, or steps in a non-exclusive manner, indicating that the referenced elements, components, or steps may be present, or utilized, or combined with other elements, components, or steps that are not expressly referenced. The meaning of “a,” “an,” and “the” includes plural reference unless the context clearly dictates otherwise. Also, as used in the description herein, the meaning of “in” includes “in” and “on” unless the context clearly dictates otherwise. The numbers expressing quantities of ingredients, properties such as concentration, reaction conditions, and so forth, used to describe and claim certain embodiments of the invention are to be understood as being modified in some instances by the term “about”. As used herein, the terms "about", "around", "approximately" or alike, when referring to a specified, measurable value (such as a parameter, an amount, a temporal duration, and the like), is meant to encompass the specified value and variations of and from the specified value, such as variations of + / -20% or less, alternatively variations of + / -10% or less, ... and from the specified value, insofar as such variations are appropriate to perform in the disclosed embodiments. Thus, the value to which the modifier "about" or "approximately" refers is itself also specifically disclosed. The recitation of ranges of values herein is merely intended to serve as a shorthand method of referring individually to each separate value falling within the range. The use of any and all examples, or exemplary language (e.g., “such as”) provided with respect to certain embodiments herein is intended merely to better illuminate the invention and does not pose a limitation on the scope of the invention otherwise claimed.

[0087] BRIEF DESCRIPTION OF THE DRAWINGS

[0088] FIG. 1 illustrates a computerized system according to some embodiments of the disclosure.

[0089] FIGS. 2A-2L show the performance of 12 cancer classification SVM models on the testing set;

[0090] FIGS. 3 A-3L show the plots of AUC vs. number of miRNAs for the 12 cancer types.

[0091] DETAILED DESCRIPTION

[0092] This present disclosure provides machine learning-based diagnostic classifier(s) that is / are capable of predicting tissue of origin for multiple cancer types with high accuracy using miRNAs, and further provides a diagnostic method using these diagnostic classified s) for the determination of how likely a subject has one or multiple of these cancer tissues of origin.

[0093] In one aspect, a method of developing a classifier capable of distinguishing a population of individuals having a cancer tissue of origin (i.e. a classifier developing method) is provided.

[0094] The classifier developing method substantially comprises the following two steps: SI 00: providing a training set from a population of subjects; and S200: constructing the classifier based on the training set.

[0095] Herein, the training set is a dataset that comprises an expression profile of a miRNA set (containing x miRNAs, x >0) obtained from a biopsy sample, such as a liquid biopsy sample (e.g. a plasma sample, a serum sample, a urine sample, a saliva sample, or a sweat sample, etc.) from each of the population of subjects. The population of subjects comprises two subsets of subjects identified as having and not having the cancer tissue of origin, respectively.

[0096] Herein, the data in the training set may have already been determined, and thus may have been stored in a storage medium in a local computer system, in a remote server, or in a cloud. The training set data may as well have not yet been determined, and may need to take efforts to be determined experimentally and processed (e.g. data cleaning and normalization, etc.). The experimental process determining an expression profile for a given miRNA set is well known to artisans skilled in the art, which may include the assaying of biopsy samples (e.g. liquid biopsy samples such as plasma sample, serum sample, saliva sample, etc.) obtained from each of the population of subjects (with or without the cancer tissue of origins). The expression profile of a miRNA set may include a normalized expression level for each miRNA in the set. Determining of an expression of one or more miRNAs can be realized by various different technical solutions as described above, and will be skipped for detailed description herein. The raw expression data were normalized by average expression value of pre-selected reference miRNAs, such as certain internal control miRNAs. Non-limiting examples for such internal control miRNAs include miR-149-3p, miR-2861, and miR-4463, but can also include other miRNAs.

[0097] According to some embodiments of the classifier developing method, the step S200 of constructing the classifier based on the training set comprises the sub-steps of:

[0098] S220: performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, thereby selecting an optimal set of miRNAs; and

[0099] S240: building a second machine learning model based on the optimal set of miRNAs. Herein the optimal set of miRNAs is substantially a subset of the miRNA set, and comprises y different miRNAs, where y < x. The term "optimal set of miRNAs" as used herein refers to a subset of miRNAs that gives a classification accuracy (i.e. ROC) that meets a certain condition (e.g. the accuracy passes a certain threshold) in the analysis using the first machine learning model (e.g. RF model) with recursive feature elimination (RFE). According to some embodiments of the disclosure, the optimal set of miRNAs is selected or obtained when a highest level of ROC value is produced in the analysis using the first machine learning model with RFE, and further in certain embodiments, the first machine learning model is a RF model.

[0100] Herein in sub-step S220, the first machine learning model can be a random forest (RF) model, but can optionally be other machine learning models such as a support vector machine (SVM), a neural network, classification and regression trees (CART), adaptive boosting, bagging, discriminant analysis, logistic regression, gradient boost, generalized linear models, generalized additive models, k-nearest neighbors, partial least squares, or ridge regression.

[0101] Optionally in the sub-step S220 of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection thereby selecting an optimal set of miRNAs, the optimal set of miRNAs can be selected using the first machine learning model along with m-fold cross validation (m > 5, i.e., m = 5, 6, 7, 8, 9, 10, etc.).

[0102] Further optionally in the sub-step S220 of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection thereby selecting an optimal set of miRNAs, the optimal set of miRNAs can be selected based on a highest level of ROC value obtained when performing the RFE based on the first machine learning model.

[0103] In the sub-step S240 of building a second machine learning model based on the optimal set of miRNAs, the second machine learning model can be a support vector machine (SVM), but can optionally be other machine learning models such as a random forest (RF) model, neural network, classification and regression trees (CART), adaptive boosting, bagging, discriminant analysis, logistic regression, gradient boost, generalized linear models, generalized additive models, k-nearest neighbors, partial least squares, or ridge regression.

[0104] Further optionally the sub-step S240 of building a second machine learning model based on the optimal set of miRNAs can comprise / / -fold cross validation for parameter optimization (n > 5, i.e., m = 5, 6, 7, 8, 9, 10, etc.).

[0105] According to some embodiments, the step S200 of constructing the classifier based on the training set may further comprise, prior to the sub -step S220 of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, a sub-step of:

[0106] S210: performing a limma analysis over the training set for a preliminary feature selection, thereby selecting top miRNAs.

[0107] Herein the top miRNAs (containing z different miRNAs) are substantially a subset of the miRNA set (containing x different miRNAs), and the optimal set of miRNAs (containing y different miRNAs) is substantially a subset of the top miRNAs, i.e. y < z < x. Accordingly in the sub-step S220 of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, thereby selecting an optimal set of miRNAs, the recursive feature elimination (RFE) is performed over the top miRNAs.

[0108] Herein, the "limma analysis" refers to “linear models for microarray data (limma) model” is substantially used to filter a set of miRNAs based on their statistical significances for distinguishing the cancer tissue of origin, such as in the differentiation of those having the cancer tissue of origin from those not having the cancer tissue of origin. According to some embodiments, the step S200 of constructing the classifier based on the training set may further comprise, prior to the sub -step S220 of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, a sub-step of:

[0109] S21 O' : performing a quality control over the training set, comprising: filtering out low coefficient of variation (CV) miRNAs and / or highly correlated miRNAs from the miRNA set.

[0110] Herein, the "low coefficient of variation (CV) miRNAs" refer to those miRNAs that have “near-zero variance” by satisfying the following two characteristics simultaneously: 1) the ratio of the frequency of the most prevalent value over the second most frequent value (called the “frequency ratio”) is greater than 19 (= 95 / 5); and 2) the “percent of unique values” defined as the number of unique values divided by the total number of samples (times 100) is less than 10; The "highly correlated miRNAs" refer to those miRNAs that have absolute correlations above 0.9.

[0111] According to some embodiments of the method, the step S200 of constructing the classifier based on the training set further comprises, prior to the sub-step S220 of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, the sub-steps of:

[0112] S205: performing a quality control over the training set, comprising: filtering out low coefficient of variation (CV) miRNAs and / or highly correlated miRNAs from the miRNA set to thereby obtain a cleaned training set; and

[0113] S215: performing a limma analysis over the cleaned training set for a preliminary feature selection, thereby selecting top miRNAs.

[0114] Accordingly in the sub-step S210 of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, thereby selecting an optimal set of miRNAs, the recursive feature elimination (RFE) is performed over the top miRNAs.

[0115] By utilizing the classifier developing method provided herein, different classifiers can be developed for determining the probability of an individual having different cancer tissues of origin respectively. Herein, the cancer tissues of origin can be of any cancer type, such as biliary tract cancer, bladder cancer, breast cancer, colorectal cancer, esophageal cancer, gastric cancer, glioma, liver cancer, lung cancer, ovarian cancer, pancreatic cancer, prostate cancer, sarcoma.

[0116] In another aspect, the present disclosure further provides a method utilizing the diagnostic classifier(s) developed by the classifier developing method as described above for the determination of the probability of an individual having a cancer tissue of origin.

[0117] The method substantially comprises the following two steps: (a) obtaining an expression profile of a predetermined miRNA set based on a liquid biopsy sample from the individual; and

[0118] (b) inputting the expression profile of the predetermined miRNA set into a predetermined classifier to thereby obtain the probability of the individual having the cancer tissue of origin.

[0119] Herein in the diagnostic method, the predetermined miRNA set and the predetermined classifier can be the optimal set of miRNAs and the classifier that are determined by the classifier developing method as described above. The individual is of the same species as the population of subjects, who can be a mammal and can preferably be a human being. The liquid biopsy sample can be selected from a body liquid sample, including but not limited to, a plasma sample, a serum sample, a urine sample, a saliva sample, or a sweat sample.

[0120] In the diagnostic method, there can be different embodiments for the cancer tissue of origin, which can be any cancer type, such as biliary tract cancer, bladder cancer, breast cancer, colorectal cancer, esophageal cancer, gastric cancer, glioma, liver cancer, lung cancer, ovarian cancer, pancreatic cancer, prostate cancer, and sarcoma, but is not limited by these above listed cancers.

[0121] According to different embodiments, the diagnostic method can be customized for determining the probability of an individual having each of the above 13 cancer tissues of origin, and their corresponding predetermined miRNA sets are summarized in Table 4 below.

[0122] In yet another aspect, the present disclosure further provides a method utilizing the diagnostic classifier(s) developed by the classifier developing method as described above for the estimation of how likely an individual has a certain cancer tissue of origin among several different cancer tissues of origin.

[0123] Differing from the diagnostic method described above, this method is capable of simultaneously determining the probabilities of an individual having each of multiple different cancer tissues of origin, thus offering a more comprehensive and more accurate diagnosis result.

[0124] The method substantially comprises the following three steps:

[0125] (A) obtaining an expression profile of a selected miRNA set based on a liquid biopsy sample from the individual;

[0126] (B) inputting the expression profile of the selected miRNA set into a classifying algorithm comprising at least two predetermined classifiers, each for a unique cancer tissue of origin, thereby obtaining a probability of the individual having the unique cancer tissue of origin corresponding to each predetermined classifier; and (C) estimating the predicted cancer tissue of origin for the individual based on the probability of the individual having the unique cancer tissue of origin corresponding to the each predetermined classifier.

[0127] Herein, each predetermined classifier can be obtained substantially based on the classifier developing method as described above, and the selected miRNA set substantially comprises a combination of an optimal set of miRNAs corresponding to each predetermined classifier as selected by the classifier developing method. Herein, the individual is of the same species as the population of subjects, and the liquid biopsy sample can be selected from a body liquid sample, including but not limited to, a plasma sample, a serum sample, a urine sample, a saliva sample, or a sweat sample.

[0128] According to certain embodiments of the method, step (C) of estimating the predicted cancer tissue of origin for the individual based on the probability of the individual having the unique cancer tissue of origin corresponding to the each predetermined classifier comprises the following two sub-steps:

[0129] (Cl) ranking the unique cancer tissue of origin corresponding to each predetermined classifier based on a corresponding probability Pi, thereby obtaining ranked cancer tissues of origin; and

[0130] (C2) estimating the predicted cancer tissue of origin for the individual based on the ranked cancer tissues of origin.

[0131] In other words, the above can be described as follows: n is the total number of predetermined classifiers in the classifying algorithm, with Pt corresponding to any cancer tissue of origin z (0 < z < ri). In sub-step (Cl), the unique cancer tissue of origin corresponding to each predetermined classifier is ranked based on a corresponding probability Pi, and after ranking, the ranked cancer tissues of origin can be described as Pi < P2 < ... < Pnin a highest- to-lowest rank.

[0132] According to some embodiments, step (C2) of estimating the predicted cancer tissue of origin for the individual based on the ranked cancer tissues of origin comprises the following sub -steps:

[0133] C2(a): picking up top m cancer tissues of origin (zzz < zz); and

[0134] C2(b): determining that the individual has the cancer tissue of origin based on the top zzz cancer tissues of origin.

[0135] In other words, the estimation of how likely an individual has a certain cancer tissue of origin is substantially based on a comprehensive estimation of multiple cancer types. According to some embodiments, m = 1 (i.e., the estimation is only on the top 1 cancer tissue of origin, or the highest ranked cancer tissue of origin). Herein optionally, the cancer tissue of origin is bladder cancer, breast cancer, esophageal cancer, gastric cancer, glioma, lung cancer, or prostate cancer, and the method is capable of reaching an accuracy equal to or more than 70%. Further optionally, the cancer tissue of origin is breast cancer, gastric cancer, glioma, or prostate cancer, and the method is capable of reaching an accuracy equal to or more than 80%. Further optionally, the cancer tissue of origin is breast cancer or glioma, and the method is capable of reaching an accuracy equal to or more than 90%.

[0136] According to some other embodiments, m = 2 (i.e., the estimation is only on the top 2 cancer tissues of origin, or the highest and the second highest ranked cancer tissues of origin). The method is capable of reaching an accuracy equal to or more than 90% for certain cancer tissues of origin.

[0137] According to yet some other embodiments, m = 3 (i.e., the estimation is only on the top 3 cancer tissue of origin, or the highest, the second highest, and the third highest ranked cancer tissues of origin). Herein optionally, the cancer tissue of origin is bladder cancer, breast cancer, esophageal cancer, gastric cancer, glioma, liver cancer, lung cancer, or prostate cancer, and the method is capable of reaching an accuracy equal to or more than 90%. Further optionally, the cancer tissue of origin is bladder cancer, breast cancer, gastric cancer, glioma, lung cancer, or prostate cancer, and the method is capable of reaching an accuracy equal to or more than 95%. Further optionally, the cancer tissue of origin is gastric cancer, glioma, or lung cancer, and the method is capable of reaching an accuracy equal to or more than 98%.

[0138] According to some preferred embodiments of the method, the selected miRNA set comprises the 426 miRNAs whose sequences are set forth in SEQ ID NO: 1-426 (see Table 2). The method as such substantially provides a comprehensive and simultaneous estimation of a total of 13 different cancer types (biliary tract cancer, bladder cancer, breast cancer, colorectal cancer, esophageal cancer, gastric cancer, glioma, liver cancer, lung cancer, ovarian cancer, pancreatic cancer, prostate cancer, and sarcoma), based on the expression profile of the 426 miRNAs and by means of the 13 classifiers that are respectively developed by the aforementioned classifier developing method.

[0139] It is to be noted that there can be different combination of classifiers for the prediction of different cancer types according to different embodiments. For example, one method may contain the unique miRNAs shown in the top two rows (i.e. biliary tract cancer and bladder cancer), and can thus be used to estimate how likely an individual has biliary tract cancer or bladder cancer by utilizing the two classifiers specifically for the two cancer types. It is to be further noted that the method is not limited to the 13 cancer types, but can expand to other cancer types as sufficient data are available for the training and testing of machine learning models for the development of classifiers that specifically target those cancer types.

[0140] In yet another aspect, this present disclosure further provides a computerized solution, which substantially serves, in a computerized and automatic manner, to implement the various steps and / or sub-steps of the classifier developing method and the diagnostic / estimating method as described above. Such a computerized solution may be applied in a situation where the implementation of the various steps of the method described above is automated by running a software program comprising program instructions in a computer, which brings about advantages such as high efficiency and great convenience.

[0141] Specifically, such a computerized solution may include a computerized system or computer system (or short as a "system"). The system comprises a collection of hardware (e.g. processor, memory, I / O interface, storage medium, etc.) and software (i.e. computer programs, including operation system software, and specific program software, etc.), which are configured to collaboratively work so as to collectively implement all or some steps of the cancer detection diagnostic model development method as described above. According to some embodiments, the system comprises a processor (i.e. controller) and a computer-readable non-transitory storage medium that is communicatively coupled to the processor. The non- transitory storage medium is configured to contain a software (i.e. program instructions) for execution by the processor, and the program instructions are configured to cause the processor to execute the various different steps and sub-steps in the cancer diagnostic model development method as described above.

[0142] As used herein and throughout the disclosure, the “processor” is interpreted to be exchangeable with “central controller” or “central computing unit (CPU)”, and can be deemed to be a single core or multi core processor, or a plurality of processors for parallel processing. The term “non-transitory,” as used herein, is intended to describe a tangible computer-readable storage medium excluding propagating electromagnetic signals, but are not intended to otherwise limit the type of physical computer-readable storage device that is encompassed by the phrase. Examples may include any tangible or non-transitory storage media or memory media such as electronic, magnetic, or optical media (e.g., disk or CD / DVD-ROM), or nonvolatile memory storage (e.g., “flash” memory), etc.

[0143] As illustrated in FIG. 1, the system 100 can, in addition to the processor 10 and the computer-readable non-transitory storage medium 20, further comprise a bus 30, a memory 40, an I / O interface 50, and a communication interface 60. The processor 10, the storage medium 20, the memory 40, the I / O interface 50 and the communication interface 60 are all communicatively coupled with one another through the bus 30.

[0144] The storage medium 20 stores computer-executable program instructions which, when executed by the processor 10, cause the processor 10 to execute steps (l)-(3) of the method as described above. The memory 40 is configured to transiently store the program instructions obtained from the storage medium 20, and the processor 10 is configured to execute the program instructions transiently stored in the memory 40. The I / O interface 50 allows an input / output between the system 100 and a user, realizing the control of the system 100. The communication interface 60 can allow the system 100 to be communicatively connected to another computing device to exchange data. It is to be noted that these computer hardware components can be locally arranged, or can be remotely arranged through a network, such as an intranet, an internet, or a cloud.

[0145] In the following, in order to illustrate some aspects of the present disclosure, two examples are provided below.

[0146] Example 1

[0147] Briefly in this example, eight serum miRNA microarray datasets from GEO totaling 6,283 patients across 13 cancer types were used in this study. The patients were split, with an approximate 3:2 ratio, into a training (n=3,844) and a testing set (n=2,439). An ensemble of classifiers was constructed in the training set via the “one vs. rest” approach, thus one classifier for each cancer type. Random forest models with recursive feature elimination (RFE) selected the optimal set of miRNAs that was fed into support vector machine models to generate a prediction probability for each cancer type. The type with the highest probability was considered the predicted cancer type. The performance of these classifiers was evaluated in the testing set in two steps with the 1st using all cancer types and the 2nd using the top 2 or 3 cancer types from the 1st step to achieve a refined prediction. RFE selected 426 miRNAs for building the 13 classification models. In the testing set comprising 2,439 patients across 12 cancer types, the classifiers correctly predicted cancer types for 1,922 (79%) samples based on the highest prediction probability. The accuracy increased to 92% and 95% based on top 2 and 3 predictions. In particular, based on top 3 predictions, the accuracy was >95% for bladder, breast, prostate, gastric, glioma and lung cancers, >85% for ovarian, liver and esophageal cancers, 78% for pancreatic cancer and sarcoma, and 67% for colorectal cancer. With 95% accuracy in narrowing TOO down to 3 organ sites, the miRNA-based TOO classifiers could be used clinically as a reflex test for the simple and highly accurate MCED screening models previously developed (Zhang A, Hu H. Cancers (Basel). 2022). Together, they support the development of an inexpensive, accurate and noninvasive blood test for MCED with TOO.

[0148] More details for the example are provided below.

[0149] I) Datasets:

[0150] Eight serum miRNA microarray datasets from GEO had been previously used to develop three datasets that were used as the source datasets (Zhang A, Hu H. Cancers (Basel). 2022; Zhang A, Hu H. Journal of Clinical Oncology. 2022) for the current work. Description of these 11 datasets has been reported in the following references (Zhang A, Hu H. Cancers (Basel). 2022; Zhang A, Hu H. Journal of Clinical Oncology. 2022; Asakura K, et al. Commun Biol. 2020; Yokoi A, et al. Nat Commun. 2018; Yamamoto Y, et al. Hepatol Commun. 2020; Usuba W, et al. Cancer Sci. 2019; Sudo K, et al. JAMA Netw Open. 2019; Abe S, et al. Gastric Cancer. 2021; Ohno M, i aJ JAMA Netw Open. 2019; UrabeF, et al. Clinical Cancer Research. 2019), which are incorporated by reference in their entirety. Combination of these above datasets gives the combined datasets that contain a total of 6,283 subjects having cancer (i.e. cancer patients) (see Table 1).

[0151] Table 1. Final combined dataset for developing TOO classifiers (N=6 283),

[0152] II) Training and Testing Sets:

[0153] The patients were split, with an approximately 3:2 ratio, into a training (n=3,844, see Table 2A) and a testing set (n=2,439, see Table 2B). Specifically, the training set comprises 60% or at least 100 cancer cases of the 13 cancer types, with the exception for the biliary tract cancer for which all the 40 patients were used. All of the remaining patients are used for the testing set, and for the 12 cancer types.

[0154] Table 2A. Training set used for developing the TOO classifiers (n=3,844). Table 2B. Testing set used for developing the TOO classifiers (n=2,439).

[0155] III) Machine Learning Analysis

[0156] The classification was done via the “one vs rest” approach, i.e., each of the 13 cancer types is compared to all the remaining 12 cancer types.

[0157] In the training set, the following approach was used to build 13 classification models to discriminate each cancer types versus the remaining 12 cancer types (“one vs rest”). Specifically, the approach includes the following three steps: (1) limma analysis was done to select the top 500 miRNAs; (2) Recursive Feature Elimination (RFE) based on random forest (RF) models was done to select the optimal set of miRNAs from the top 500 miRNAs; and (3) a Support Vector Machine (SVM) model was built based on the optimal set of miRNAs. It is noted that in the above step (2), machine learning models other than RF may alternatively be used (see section IV below), and that in the above step (3), machine learning models other than SVM may alternatively be used (see section IV below).

[0158] For the testing set, a two-stage prediction was used to refine the final cancer type prediction. Specifically, in Stage 1, a primary prediction using 13 SVM classification models was carried out, whereby each sample was predicted to give 13 prediction probabilities, one for each cancer type, and the cancer type with the highest prediction probability is the predicted cancer type for that sample. In Stage 2, a secondary prediction based on top 2 or 3 cancer types from the primary prediction was further carried out. If the highest prediction probability is <= 0.5, the secondary prediction was done among the top 3 cancer types. If the highest prediction probability is > 0.5, the secondary prediction was done among the top 2 cancer types.

[0159] IV) Feature Selection

[0160] In this section, more details are provided for the feature selection in the analysis.

[0161] Firstly, a filtering process was carried out for the data, in which low coefficient of variation (CV) miRNAs and highly correlated miRNAs were first filtered out, and then a limma analysis was carried out to further filter miRNAs based on statistical significance (with "adjusted -value < 0.05" determined as statistically significant) for differentiating each cancer from the rest (see Table 3), and finally, 500 miRNAs were selected for the subsequent Recursive Feature Elimination (RFE) for each cancer type.

[0162] Table 3. Summary of significantly upregulated or downregulated miRNAs when comparing each cancer with the rest in the limma analysis.

[0163] Note: "Down" for significantly upregulated miRNAs; "Up" for significantly upregulatec miRNAs; and "NotSig" for miRNAs that did not reach statistical significance (adjusted - value < 0.05).

[0164] Secondly, final feature selection was performed over the top 500 miRNAs filtered out in the above filtering process for each cancer type, so as to select the optimal set of miRNAs for building diagnostic models in the subsequent steps. The feature selection was carried out based on Recursive Feature Elimination (RFE), and random forest (RF) models along with 10- fold cross validation were used to determine the optimal set of miRNAs for each cancer type. It is noted that there is no limitation to the use of the RF model in here. In addition to the RF model, other machine learning models (e.g. SVM, neural network, classification and regression trees (CART), adaptive boosting, bagging, discriminant analysis, logistic regression, gradient boost, generalized linear models, generalized additive models, k-nearest neighbors, partial least squares, ridge regression, etc.) can alternatively be used. Table 4 lists the RFE-selected optimal set of miRNAs for each of the 13 cancer types that gave the highest ROC. Please note that this RFE based analysis can give different sets of miRNAs that produce different ROC values, and only the set of miRNAs that produced the highest ROC value was regarded as the optimal set of miRNAs for the subsequent development of the classification models.

[0165] Table 4. The optimal miRNA set selected by RFE for each of 13 cancer types.

[0166]

[0167]

[0168]

[0169]

[0170]

[0171]

[0172] Note: "Sens" means sensitivity; "Spec" means specificity; "Accu" means accuracy.

[0173] The combination of all the optimal sets of miRNAs for the 13 cancer types as listed above in Table 4 gives a miRNA set that consists of 426 unique miRNAs, whose sequence information and miRBase IDs are provided in Table 5. Table 5. Unique miRNAs used for building the diagnostics model for 13 cancer types.

[0174]

[0175]

[0176]

[0177]

[0178]

[0179] V) Building Classification Models

[0180] Based on the optimal sets of miRNAs respectively for the 13 cancer types listed in Table 4 above, a support vector machine (SVM, see Hsu CW, et al. IEEE Trans Neural Netw. 2002) model along with 10-fold cross validation was used to train the training set data to optimize tuning parameters, thereby building the classification models for each of the 13 cancer types (see Table 6 for the optimal SVM models developed for each of the 13 cancer types). When developing the classification models, the tuning parameter "Sigma" was held constant at a value listed in Table 6, and ROC was then used to select the optimal SVM model using the largest value (see Table 6).

[0181] It is noted that there is no limitation to the use of the SVM model in here. In addition to the SVM model, other machine learning models (e.g. RF, neural network, classification and regression trees (CART), adaptive boosting, bagging, discriminant analysis, logistic regression, gradient boost, generalized linear models, generalized additive models, k-nearest neighbors, partial least squares, ridge regression, etc.) can also be used.

[0182] Table 6. The optimal SVM models developed for each of the 13 cancer types.

[0183] VI) Prediction on the Testing Set

[0184] In order to test the performance of the above 13 developed classifiers, the testing data from the Testing set was used for the two-stage prediction. Because all the 40 biliary cancer cases had been used in the Training set, the Testing set composed of data from 12 cancer types.

[0185] Specifically, in Stage 1, a primary prediction using each of the above 13 SVM classification models was carried out. Each sample was predicted to give 13 prediction probabilities, one for each cancer type, and the cancer type with the highest prediction probability is regarded as the predicted cancer type for that sample. The performances of the 12 cancer classification SVM models on the Testing set is shown in FIGS. 2A-2L and summarized in Table 7. Out of 2439 patients in the Testing set, 1816 (74%) had correct predicted cancer type; 400 cases with 2nd highest probability matched to the true cancer type; and 102 cases with 3rd highest probability matched to the true cancer type. Thus, the accuracy based on top 2 prediction is 91%, and the accuracy based on top 3 prediction is 95%.

[0186] In Stage 2, a secondary prediction based on top 2 or 3 cancer types from the primary prediction was further carried out, and the following results were obtained. The performance of the 12 cancer classification SVM models on the Testing set is shown in Table 8. Out of 2439 patients in the Testing set, 1922 (79%) had correct predicted cancer type which was improved from 74%, the accuracy based on top 2 prediction is 92%, and the accuracy based on top 3 prediction is 95%.

[0187] Table 8. Accuracies of the SVM model for individual cancer type based on top 3 predictions in the secondary prediction.

[0188] VII. Additional analysis

[0189] Some additional analysis was performed to evaluate the number of miRNAs needed to make accurate predictions. For each classification model for each cancer type, differing number of miRNAs (i.e. 5, 10, 15, . . . . ) were selected from the optimal set of miRNAs for each cancer type and a SVM model was built, and the SVM model was then applied to the testing set to calculate the AUC of ROC curve for the cancer type. The process was repeated 10 times for each set of miRNAs selected and the AUC was averaged. The average AUC was then plotted against the number of miRNAs to examine the effect of the number of miRNAs on the performance of SVM model as measured by AUC. FIGS. 3A-3L show the plots of AUC vs. number of miRNAs for each cancer type.

[0190] As shown in the figures, subsets of the identified optimal list of the miRNAs for cancer classification can also be used to develop effective cancer classifiers. However, for most of the cancer types, it appears that at least 10 or 15 miRNAs were required to achieve good performance (AUC of at least 0.8); for a few cancer types, 5 miRNAs are sufficient to achieve such performance.

[0191] Example 2

[0192] This example intends to provide specific demonstration of the 12 classification models (i.e. classifiers) developed above using real cancer patients. It is noted that due to the lack of testing data, biliary tract cancer is not included in the analysis.

[0193] Specifically, a total of 12 cancer patients from the GEO database who have been definitively diagnosed were tested. Firstly, the expression profile of the 426-miRNA panel (Table 5) for each of the 12 cancer patients (from a serum sample) was obtained. Secondly, the expression profile of the 426-miRNA panel for each cancer patient was inputted in a classifying algorithm that consists of the 12 developed classifiers, each for a different cancer type, thereby obtaining the probabilities of this cancer patients having each of the 12 cancer types. The probabilities for each of these 12 cancer patients are shown in Tables 9A and 9B.

[0194] Table 9A. SVM classifier-based prediction results for 12 cancer patients (part I).

[0195] Table 9B. SVM classifier-based prediction results for 12 cancer patients (part II).

[0196] As shown in Tables 9A-9B, for each cancer patient, the SVM classification models that have been developed for predicting the 12 cancer types respectively by the classifier developing method described above output a probability of this cancer patient having each of the 12 cancer types (see each column for the 12 cancer types). Importantly, among the 12 predicted cancer types, the one with the highest probability ("Top Predicted Cancer Type") precisely matched with the diagnosed cancer type for each cancer patient. In one example, for cancer patient (GSM3106951), the top predicted cancer type is bladder cancer (P = 0.5833), which is exactly the diagnosed cancer type for the patient. In another example, for cancer patient (GSM2850714), although three cancer types are predicted with close probabilities (i.e. P = 0.2799, 0.2295, and 0.1875 for breast cancer, colorectal cancer, and ovarian cancer, respectively), breast cancer gave the highest probability and still notably matched well with the diagnosed cancer type for the patient. Thus, the aforementioned SVM classification models that have been developed using the classifier developing method described above are capable of accurately predicting the cancer type for each of these 12 cancer patients.

[0197] References

[0198] Abe S, et al. Gastric Cancer. 2021;24(4):835-843.

[0199] Altschul SF, et al. J Mol Biol. 1990 Oct 5;215(3):403-10.

[0200] Asakura K, et al. Commun Biol. 2020;3(l).

[0201] GEO (Gene Expression Omnibus): https: / / www.ncbi.nlm.nih.gov / geo /

[0202] Hsu CW, et al. IEEE Trans Neural Netw. 2002;13(2):415-25.

[0203] Kozomara A & Griffiths-Jones S. Nucleic Acids Res. 2011 Jan;39(Database issue):D152-7.

[0204] Nelson HD, et al. Ann Intern Med. 2016 Feb 16;164(4):244-55.

[0205] Nielsen PE, et al. Science. 1991 Dec 6;254(5037): 1497-500.

[0206] Obika S, et al. Tetrahedron Letters. 1998 Jul 23;39(30):5401-5404.

[0207] Ohno M, et al. JAMA Netw Open. 2019;2(12):el916953.

[0208] Pearson WR, et al. Proc Natl Acad Sci USA. 1988 Apr;85(8):2444-8.

[0209] Ritchie ME, et al. Nucleic Acids Res. 2015 Apr 20;43(7):e47.

[0210] Sambrook J, et al. 2001. Molecular Cloning: A Laboratory Manual. 3rd ed. Plainview, N.Y.: Cold Spring Harbor Laboratory Press.

[0211] Sudo K, et al. JAMA Netw Open. 2019;2(5):el94573.

[0212] Urabe F, et al. Clinical Cancer Research. 2019;25(10):3016-3025.

[0213] US Preventive Services Task Force, et al. JAMA. 2018 Aug 21;320(7):674-686.

[0214] US Preventive Services Task Force, et al. JAMA. 2021 May 18;325(19): 1965-1977.

[0215] US Preventive Services Task Force, Krist AH, et al. JAMA. 2021 Mar 9;325(10):962-970.

[0216] Usuba W, et al. Cancer Sci. 2019; 110(1).

[0217] Wang Z, et al. Nat Rev Genet. 2009 Jan;10(l):57-63.

[0218] Yamamoto Y, et al. Hepatol Commun. 2020;4(2):284-297.

[0219] Yokoi A, et al. Nat Commun. 2018;9(l).

[0220] Zhang Z, et al. J Comput Biol. 2000 Feb-Apr;7(l-2):203-14.

[0221] Zhang A, Hu H. Cancers (Basel). 2022;14(6): 1450.

[0222] Zhang A, Hu H. Journal of Clinical Oncology. 2022;40(16_suppl):3065-3065.

Claims

CLAIMS1. A method of developing a classifier capable of distinguishing a population of individuals having a cancer tissue of origin, comprising the steps of:(a) providing a training set from a population of subjects, wherein the training set comprises an expression profile of a miRNA set obtained from a biopsy sample from each of the population of subjects, wherein the miRNA set comprises x different miRNAs, where x is a positive integer, and wherein the population of subjects comprises two subsets of subjects identified as having and not having the cancer tissue of origin, respectively; and(b) constructing the classifier based on the training set, comprising the sub-steps of:(bl) performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, thereby selecting an optimal set of miRNAs, wherein the optimal set of miRNAs comprises y different miRNAs, where y < x and(b2) building a second machine learning model based on the optimal set of miRNAs.

2. The method of claim 1, wherein the first machine learning model is optionally selected from a random forest (RF) model, a support vector machine (SVM), a neural network, classification and regression trees (CART), adaptive boosting, bagging, discriminant analysis, logistic regression, gradient boost, generalized linear models, generalized additive models, k- nearest neighbors, partial least squares, or ridge regression, and preferably the first machine learning model is a RF model.

3. The method of claim 1, wherein in the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection thereby selecting an optimal set of miRNAs, the optimal set of miRNAs is selected using the first machine learning model along with m-fold cross validation, where m is a positive integer no less than 5.

4. The method of claim 1, wherein in the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection thereby selecting an optimal set of miRNAs, the optimal set of miRNAs is selected based on a highest level of ROC value obtained when performing the RFE based on the first machine learning model.

5. The method of claim 1, wherein the second machine learning model is optionally selected from a support vector machine (SVM), a random forest (RF) model, neural network, classification and regression trees (CART), adaptive boosting, bagging, discriminant analysis, logistic regression, gradient boost, generalized linear models, generalized additive models, k- nearest neighbors, partial least squares, or ridge regression, and preferably the second machine learning model is SVM.

6. The method of claim 1, wherein the sub-step (b2) of building a second machine learning model based on the optimal set of miRNAs comprises / / -fold cross validation for parameter optimization, where n is a positive integer no less than 5.

7. The method of claim 1, wherein the step (2) of constructing the classifier based on the training set further comprises, prior to the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, a sub-step of performing a limma analysis over the training set for a preliminary feature selection, thereby selecting top miRNAs, wherein the top miRNAs comprises z different miRNAs, where y < z < x, wherein: in the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, thereby selecting an optimal set of miRNAs, the recursive feature elimination (RFE) is performed over the top miRNAs.

8. The method of claim 3, wherein the step (2) of constructing the classifier based on the training set further comprises, prior to the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, a sub-step of performing a quality control over the training set, comprising: filtering out low coefficient of variation (CV) miRNAs and / or highly correlated miRNAs from the miRNA set.

9. The method of claim 3, wherein the step (2) of constructing the classifier based on the training set further comprises, prior to the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, the sub-steps of: performing a quality control over the training set, comprising: filtering out low coefficient of variation (CV) miRNAs and / or highly correlated miRNAs from the miRNA set to thereby obtain a cleaned training set; andperforming a limma analysis over the cleaned training set for a preliminary feature selection, thereby selecting top miRNAs, wherein the top miRNAs comprises z different miRNAs, where < z < x; wherein: in the sub-step (bl) of performing recursive feature elimination (RFE) based on a first machine learning model for feature selection, thereby selecting an optimal set of miRNAs, the recursive feature elimination (RFE) is performed over the top miRNAs.

10. The method of claim 1, wherein the biopsy sample is a liquid biopsy sample, selected from a plasma sample, a serum sample, a urine sample, a saliva sample, or a sweat sample.

11. A method for determining the probability of an individual having a cancer tissue of origin, comprising:(a) obtaining an expression profile of a predetermined miRNA set based on a liquid biopsy sample from the individual;(b) inputting the expression profile of the predetermined miRNA set into a predetermined classifier to thereby obtain the probability of the individual having the cancer tissue of origin; wherein: the predetermined miRNA set is the optimal set of miRNAs selected from, and the predetermined classifier is the classifier developed by, the method according to any one of claims 1-10; and the individual is of the same species as the population of subjects.

12. The method of claim 11, wherein the liquid biopsy sample is selected from a plasma sample, a serum sample, a urine sample, a saliva sample, or a sweat sample.

13. The method of claim 11, wherein the cancer tissue of origin is one of the following: biliary tract cancer, wherein the predetermined miRNA set comprises at least one of the miRNAs with SEQ ID NOS: 1-75; bladder cancer, wherein the predetermined miRNA set comprises at least one of the miRNAs with SEQ ID NO: SEQ ID NOS: 1, 3, 5, 14, 51, and 76-120; breast cancer, wherein the predetermined miRNA set comprises at least one of the miRNAs with SEQ ID NOS: 14, 77-78, 80, 82, 91-92, and 121-188;colorectal cancer, wherein the predetermined miRNA set comprises at least one of the miRNAs with SEQ ID NOS: 1, 17, 37, 77, 89, 112, 116, 121, 128, 130-132, 135, 151, and 189- 224; esophageal cancer, wherein the predetermined miRNA set comprises at least one of the miRNAs with SEQ ID NOS: 1, 6, 14, 54, 82, 93, 95, 97, 100, 104, 111, 120, 130, 141, 143, 147, 154, 156, 158, 186, 194, 211, 217-218, and 225-250; gastric cancer, wherein the predetermined miRNA set comprises at least one of the miRNAs with SEQ ID NOS: 6, 37, 54, 76, 78-80, 82, 84-86, 90-91, 93, 95, 97, 100-101, 111- 112, 120, 122, 127-128, 130, 132, 138, 141, 143-144, 147, 154, 156, 158, 174, 178, 186, 189, 191, 194-196, 205, 211, 225-232, 235-236, 238, 242, 244, and 251-293; glioma, wherein the predetermined miRNA set comprises at least one of the miRNAs with SEQ ID NOS: 6, 17, 120, 128, 154, 156, 158, 195, 225-226, 228, 247, and 294-306; liver cancer, wherein the predetermined miRNA set comprises at least one of the miRNAs with SEQ ID NOS: 1, 6, 9, 12, 14, 26, 28, 35, 39, 57, 76-78, 83, 89, 112, 120, 128, 132, 137, 151, 174, 189, 194, 214, 236, 263, 272, 280, 293, and 307-326; lung cancer, wherein the predetermined miRNA set comprises miRNAs at least one of the miRNAs with SEQ ID NOS: 3-4, 6, 37, 76, 79, 82, 86, 91, 93, 101, 112, 120, 122, 128, 130, 132, 141, 154, 156, 160, 165, 172, 177-178, 186, 189, 191, 194-196, 209, 211, 217-218, 225- 226, 230, 232, 257, 263, 285-286, 299, and 326-331; ovarian cancer, wherein the predetermined miRNA set comprises miRNAs at least one of the miRNAs with SEQ ID NOS: 12, 76, 79, 101, 128, 130, 143, 154, 186, 195, 211, 225-226, 232, 236, 247, 254, 268, 280, 302, 329, and 332-335; pancreatic cancer, wherein the predetermined miRNA set comprises at least one of the miRNAs with SEQ ID NO: 1, 14, 25, 28, 48, 77, 89, 95, 114, 116, 120, 130, 140-141, 143, 147, 186, 194-195, 211, 218, 227-229, 231-232, 240, 244, 276, 293, and 336-355; prostate cancer, wherein the predetermined miRNA set comprises at least one of the miRNAs with SEQ ID NOS: 1, 10, 14, 23, 25, 77, 79, 83, 85-86, 89, 104, 120, 122, 143, 165- 166, 172, 217, 239, 248, 254, 263, 283, 285, 292, 299, 300, 315, 328, 330, and 356-374; or sarcoma, wherein the predetermined miRNA set comprises at least one of the miRNAs with SEQ ID NO: 4, 6, 12, 18, 25, 54, 58, 69-70, 78, 84-86, 89, 91, 95, 100, 108, 118, 120-121, 130, 132-133, 141, 143, 146-147, 149, 151, 154, 161, 166, 172, 181, 186-187, 194, 205, 217-218, 225, 227-232, 234, 236, 238-241, 243-245, 249, 253, 263, 272, 286, 291, 293, 299-300, 307, 313-314, 319, 322, 327, 369, and 375-426.

14. A method for determining a cancer tissue of origin for an individual, comprising the steps of:(a) obtaining an expression profile of a selected miRNA set based on a liquid biopsy sample from the individual;(b) inputting the expression profile of the selected miRNA set into a classifying algorithm comprising at least two predetermined classifiers, each for a unique cancer tissue of origin, thereby obtaining a probability of the individual having the unique cancer tissue of origin corresponding to each predetermined classifier; and(c) estimating the predicted cancer tissue of origin for the individual based on the probability of the individual having the unique cancer tissue of origin corresponding to the each predetermined classifier; wherein: each predetermined classifier is obtained based on, and the selected miRNA set comprises a combination of an optimal set of miRNAs corresponding to the each predetermined classifier as selected by, the method according to any one of claims 1-10; and the individual is of the same species as the population of subjects.

15. The method of claim 14, wherein step (c) comprises the sub-steps of:(cl) ranking the unique cancer tissue of origin corresponding to the each predetermined classifier based on a corresponding probability i, thereby obtaining ranked cancer tissue of origin, where n is the total number of the at least two predetermined classifiers, i corresponds to cancer tissues of origin z (0 < z < / ?); and(c2) estimating the predicted cancer tissue of origin for the individual based on the ranked cancer tissues of origin.

16. The method of claim 15, wherein sub-step (c2) of estimating the predicted cancer tissue of origin for the individual based on the probability of the individual having the unique cancer tissue of origin corresponding to the each predetermined classifier comprises: picking up top m cancer tissues of origin (m < ri) and determining that the individual has the cancer tissue of origin based on the top m cancer tissues of origin.

17. The method of claim 16, wherein m =1, wherein the cancer tissue of origin is bladder cancer, breast cancer, esophageal cancer, gastric cancer, glioma, lung cancer, or prostate cancer, and the method is capable of reaching an accuracy equal to or more than 70%.

18. The method of claim 17, wherein the cancer tissue of origin is breast cancer, gastric cancer, glioma, or prostate cancer, and the method is capable of reaching an accuracy equal to or more than 80%.

19. The method of claim 17, wherein the cancer tissue of origin is breast cancer or glioma, and the method is capable of reaching an accuracy equal to or more than 90%.

20. The method of claim 16, wherein m =2.

21. The method of claim 16, wherein m =3, wherein the cancer tissue of origin is bladder cancer, breast cancer, esophageal cancer, gastric cancer, glioma, liver cancer, lung cancer, or prostate cancer, and the method is capable of reaching an accuracy equal to or more than 90%.

22. The method of claim 21, wherein the cancer tissue of origin is bladder cancer, breast cancer, gastric cancer, glioma, lung cancer, or prostate cancer, and the method is capable of reaching an accuracy equal to or more than 95%.

23. The method of claim 21, wherein the cancer tissue of origin is gastric cancer, glioma, or lung cancer, and the method is capable of reaching an accuracy equal to or more than 98%.

24. The method of claim 14, wherein the selected miRNA set comprises at least one of the miRNAs whose sequences are set forth in SEQ ID NO: 1-426.

25. A system, comprising: a processor; and a non-transitory storage medium containing program instructions for execution by the processor, wherein the program instructions cause the processor to execute steps in the method according to any one of claims 1-24.

26. A non-transitory storage medium, storing computer-executable program instructionswhich, when executed by a processor, cause the processor to execute the method according to any one of claims 1-24.

Images