Construction and use of tissue-of-origin prediction model

Abstract

Provided are construction and use of a tissue-of-origin prediction model. Specifically, a model for tissue-of-origin prediction or diagnosis is developed, including prediction of primary tissue of cancer; and a tissue-of-origin prediction model-based system is developed. The tissue-of-origin prediction model provides an efficient and accurate auxiliary diagnostic result for determining an origin site of a tissue to be tested, or a primary site of a tumor tissue to be tested.

Description

Construction and application of organization source prediction models Technical Field

[0001] This invention belongs to the field of medical diagnostics, specifically relating to the construction and application of tissue origin prediction models. Background Technology

[0002] Tumor pathology diagnosis is currently one of the gold standards for diagnosing many types of tumors, boasting a high degree of accuracy. Through tumor pathology diagnosis, the nature and origin of the tumor can be clearly identified, and its stage and type can be determined, along with its degree of differentiation. By observing and analyzing the histological characteristics, grading, and staging of the tumor, patient survival rates and recurrence risks can be predicted, providing crucial references for treatment decisions and patient management, thereby guiding individualized treatment strategies.

[0003] However, different types of tumors may have similar histological features. Sometimes patients may have multiple primary tumors or multiple independent malignant lesions at the same time. In addition, the origin of some metastatic tumors may not be determined by routine histological examination, such as cancer of unknown primary (CUP). Pathology has certain limitations in diagnosing primary and / or metastatic tumors in some cases, and molecular genetics and other techniques are needed to assist in diagnosis.

[0004] CUP refers to metastatic malignant tumors that are histologically confirmed but cannot be detected after a comprehensive and detailed examination of the primary site. CUP is characterized by early metastasis, rapid progression, multiple organ involvement in more than half of the patients, unknown metastatic patterns, and high mortality. Clinical treatment for CUP patients is mainly conventional chemotherapy, but empirical chemotherapy is ineffective, lacks specificity, and has failed to improve the survival prognosis of patients.

[0005] Identifying the primary tumor (including tumor type and tissue origin) is crucial for selecting patients with CUP who may benefit from targeted therapy. Pathological diagnosis of CUP is a complex task. For example, the location of the primary tumor is often difficult to pinpoint, with multiple potential primary sites involving different organs and tissues, and the clinical presentation is nonspecific. To address these diagnostic challenges, physicians typically employ a combination of diagnostic methods, including detailed medical history taking, physical examination, imaging studies, pathological analysis, immunohistochemical staining, and molecular biological testing. Furthermore, collaboration with a multidisciplinary team, including pathologists, radiologists, internists, and oncologists, is essential to improve the accuracy of diagnosis and treatment outcomes for tumors with unknown primary sites.

[0006] Furthermore, tumor pathological diagnosis is inherently subjective, requiring a high level of professional knowledge and experience from physicians. Different pathologists may interpret and judge the same tissue specimen differently. This can lead to inconsistencies in diagnoses between different hospitals and specialists, thereby affecting the comparability of treatment decisions and outcomes. Therefore, those skilled in the art recognized the need to obtain more tissue-specific biomarkers or biological characteristics at the molecular biological level to improve diagnostic accuracy.

[0007] Therefore, there is an urgent need in this field to develop a new, more accurate and efficient method or device for cancer-aided diagnosis, including the diagnosis of cancer tissue origin. Summary of the Invention

[0008] This invention provides a novel method for predicting tissue origin.

[0009] In a first aspect of the present invention, a method for constructing an organization origin prediction model is provided, comprising the steps of:

[0010] (s1) Process the third-generation sequencing data and RNA sequencing (RNA-seq) data of cancer tissue samples, adjacent normal tissue samples and normal tissue samples from different tissue sources to obtain reference GTF files of full-length RNA transcripts and tissue-specific RNA transcripts (Ti-SRT, Tissue Specific RNA Transcript) of different tissues.

[0011] (s2) Provide training set, test set and validation set data for model construction, the data including the expression information of full-length RNA transcripts obtained from RNA-seq data of tissue samples from different tissue sources after RNA sequencing, based on the reference GTF file; the expression information includes the expression information of tissue-specific RNA transcripts from the different tissues;

[0012] Specifically, for a certain type of tissue, the expression information of tissue-specific RNA transcripts of that tissue is the expression information of tissue-specific RNA transcripts of tissue-positive samples, while the expression information of tissue-specific RNA transcripts of other tissues for that tissue is the expression information of tissue-specific RNA transcripts of tissue-negative samples.

[0013] (s3) Construction of tissue origin sub-models: Different sub-models are constructed for different tissues to determine whether the sample to be tested originates from that tissue; including the following steps: For any tissue, the expression information of tissue-specific RNA transcripts of tissue-positive samples and tissue-negative samples of that tissue are input into different machine learning models for training; thereby obtaining tissue origin prediction sub-models for determining whether the sample to be tested originates from that tissue; and using this step, tissue origin prediction sub-models for all tissues are obtained;

[0014] The machine learning models include: Random Forest, Decision Tree, XGBoost, LightGBM, and CatBoost.

[0015] (s4) Construction of the organization source prediction model: Based on the organization source prediction sub-models of all organizations obtained in (s3), the test set and validation set data are used for optimization and selection to construct the organization source prediction model.

[0016] In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of adrenal tissue, the machine learning model is a random forest model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of bladder tissue, the machine learning model is an XGBoost model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of brain tissue, the machine learning model is a random forest model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of breast tissue, the machine learning model is a random forest model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of cervical tissue, the machine learning model is a LightGBM model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of intestinal tissue, the machine learning model is a CatBoost model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of B cells, the machine learning model is a LightGBM model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of esophageal tissue, the machine learning model is a random forest model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of head and neck tissue, the machine learning model is an XGBoost model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of kidney tissue, the machine learning model is a CatBoost model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of bone marrow tissue, the machine learning model is a CatBoost model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of liver tissue, the machine learning model is a CatBoost model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of lung tissue, the machine learning model is a Random Forest model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of ovarian tissue, the machine learning model is a CatBoost model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of pancreatic tissue, the machine learning model is a LightGBM model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of ganglion tissue, the machine learning model is a CatBoost model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of prostate tissue, the machine learning model is a CatBoost model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of bone and / or muscle tissue, the machine learning model is a random forest model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of skin tissue, the machine learning model is a LightGBM model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of stomach tissue, the machine learning model is a random forest model.In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of testicular tissue, the machine learning model is a CatBoost model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of thyroid tissue, the machine learning model is a random forest model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of thymus tissue, the machine learning model is an XGBoost model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of uterine tissue, the machine learning model is a random forest model. In another preferred embodiment, for the tissue origin prediction sub-model used to determine the origin of eye tissue, the machine learning model is a random forest model.

[0017] In another preferred embodiment, the transcript expression information includes the transcript expression level as measured in TPM (Transcript per million).

[0018] In another preferred embodiment, step (s3) includes the following sub-steps:

[0019] (s3a) Construct different tissue origin prediction sub-models for different tissues to determine whether the sample to be tested originates from that tissue; for any tissue, input the expression information of tissue-specific RNA transcripts of tissue-positive samples and tissue-negative samples for that tissue into the following five machine learning models for training: Random Forest, Decision Tree, XGBoost, LightGBM, and CatBoost; thereby obtaining the hyperparameters of the five machine learning models corresponding to the tissue; and based on this, obtain five trained machine learning models (or the first tissue origin prediction sub-model or the first sub-model) for determining the origin of any tissue;

[0020] (s3b) For the five first organization source prediction sub-models for determining any organization source, the first organization source prediction sub-models are scored using test set and validation set data, and the following scoring indicators are used: AUC value, accuracy, sensitivity and specificity; for any organization, the first organization source prediction sub-model with the highest score is taken as the final organization source prediction sub-model (or second organization source prediction sub-model or second sub-model) used for determining the organization.

[0021] (s3c) Obtain the organization source prediction sub-models for all organizations, i.e., all the second organization source prediction sub-models used for organization source prediction;

[0022] In another preferred embodiment, step (s4) includes the following steps:

[0023] (s4a) Based on the organization source prediction sub-models of all organizations obtained in (s3), use the validation set data and the roc auc function to score the organization source sub-models of all organizations;

[0024] (s4b) Select all sub-models for predicting the origin of an organization with AUC ≥ 0.9 and construct the organization origin prediction model.

[0025] In another preferred embodiment, the tissue samples include tissue samples from m different tissue sources, where m is a positive integer ≥10.

[0026] In another preferred embodiment, m ≥ 15; more preferably, n is 15-50; even more preferably, n is 18-40.

[0027] In another preferred embodiment, the m different tissue sources include: adrenal glands, B cells, bladder, brain, breast, bone and muscle, bone marrow, cervix, intestine, esophagus, eye, ganglion, kidney, liver, lung, head and neck, ovary, pancreas, prostate, skin, stomach, testis, thymus, thyroid gland, and uterus.

[0028] In another preferred embodiment, the tissue origin prediction sub-model is used to determine whether the tissue of the sample to be tested originates from the following sites: adrenal gland, bladder, bone marrow, brain, breast, intestine, eye, ganglion, liver, pancreas, prostate, thyroid, B cells, kidney, head and neck, skin, stomach, and uterus.

[0029] In another preferred embodiment, the tissue-specific RNA transcripts include tissue-specific RNA transcripts for the following tissues: adrenal glands, B cells, bladder, brain, breast, bone and muscle, bone marrow, cervix, intestine, esophagus, eye, ganglia, kidneys, liver, lungs, head and neck, ovaries, pancreas, prostate, skin, stomach, testes, thymus, thyroid gland, and uterus.

[0030] In another preferred embodiment, step (s1) includes: establishing a full-length transcript reference GTF file by processing third-generation sequencing data and filtering transcript annotations; the GTF file is used for quantifying RNAseq transcripts obtained from sample RNA sequencing; and

[0031] By comparing RNAseq transcript expression information from different tissue samples, tissue-specific RNA transcripts from different tissues can be obtained.

[0032] In another preferred embodiment, step (s1) includes the following steps:

[0033] (a) Collect third-generation sequencing data and RNA sequencing data of cancer tissue samples, adjacent normal tissue samples and normal tissue samples from different tissue sources, and generate non-redundant transcriptome GTF files;

[0034] (b) Optimize non-redundant transcriptome GTF files, wherein, for a certain tissue RNA-seq, if all splicing sites of a transcript are detected in ≥m samples, the transcript is retained, where m is 1-90%, preferably 2-50%;

[0035] (c) Each transcript was compared with the reference transcriptome (GENCODE v.35). The transcripts were mainly divided into four groups: fully spliced, partially spliced, novel within the catalog, and novel outside the catalog. After filtering, a reference GTF file for full-length RNA transcripts was established.

[0036] Specifically, for all incompletely spliced ​​matching transcripts, these transcripts are filtered out from the entire transcriptome;

[0037] Unqualified novel transcripts, both in-directory and out-of-directory, are filtered out (e.g., those with Cage peaks and polyA sites annotated with SQANTI3 are filtered out if there are 16 or more adenines within 20 bases downstream of the annotated transcription termination site).

[0038] For similar transcripts, the longest transcript is retained, and other transcripts are filtered out;

[0039] (d) Based on the reference GTF file for full-length RNA transcripts, perform full-length transcript quantification on RNA-seq data from tissue samples;

[0040] (e) By comparison, RNA transcripts that are significantly different from other tissue samples are obtained, thereby obtaining tissue-specific RNA transcripts for different tissues.

[0041] In another preferred embodiment, the significant difference is that it is expressed in a particular tissue but not expressed or expressed at very low levels in other tissues.

[0042] In another preferred embodiment, the tissue-specific RNA transcript is obtained through the following steps:

[0043] (a) Collect third-generation sequencing data and RNA sequencing data of cancer tissue, adjacent normal tissue and normal tissue samples from different tissue sources;

[0044] (b) A full-length transcript reference GTF file is established by processing third-generation sequencing data and filtering transcript annotations; the GTF file is used for quantification of RNA-seq transcripts obtained from sample RNA-seq sequencing; and

[0045] (c) Using Salmon software, the RNA-seq short read sequences obtained from RNA sequencing are aligned to the transcript library corresponding to the GTF file, and the expression level of transcripts in each sample is measured in the form of TPM (Transcript per million, the number of transcripts per million transcripts) to generate transcript expression information of the tissue to be tested; the transcript expression information includes the expression information of tissue-specific RNA transcripts of different tissues.

[0046] In another preferred embodiment, when the tissue to be tested is tumor tissue, the tissue origin prediction model is used to assess the primary tissue of the tumor.

[0047] In a second aspect of the invention, a reference GTF file for a full-length RNA transcript is provided, obtained by comprising the following steps:

[0048] (i) Collect third-generation sequencing data and RNA-seq sequencing data of cancer tissue samples, adjacent normal tissue samples and normal tissue samples from different tissue sources, and generate non-redundant transcriptome GTF files;

[0049] (ii) Optimize non-redundant transcriptome GTF files, wherein, for a certain tissue RNA-seq, if all splicing sites of a transcript are detected in ≥m samples, the transcript is retained, where m is 1-90%, preferably 2-50%;

[0050] (iii) Each transcript was compared with the reference transcriptome (GENCODE v.35). The transcripts were mainly divided into four groups: fully spliced, partially spliced, novel in-directory, and novel out-of-directory. After filtering, a reference GTF file for full-length RNA transcripts was established.

[0051] Specifically, for all incompletely spliced ​​matching transcripts, these transcripts are filtered out from the entire transcriptome;

[0052] Unqualified novel transcripts, both in-directory and out-of-directory, are filtered out (e.g., those with Cage peaks and polyA sites annotated with SQANTI3 are filtered out if there are 16 or more adenines within 20 bases downstream of the annotated transcription termination site).

[0053] For similar transcripts, the longest transcript is retained, while other transcripts are filtered out.

[0054] In another preferred embodiment, the full splice matching, incomplete splice matching, in-catalog novel, and out-of-catalog novel are respectively: a transcript that is completely identical to a known isoform in the reference transcriptome database compared to the reference transcriptome; a transcript that is completely contained by a known transcript but has fewer exons than the known transcript, and whose intron sequences completely correspond to the known transcript; a transcript that shows a known splice site but presents a new splice pattern or exon combination; and a transcript that has at least one splice site that is not annotated in existing databases.

[0055] In a third aspect of the invention, a method for determining the expression information of full-length RNA transcripts is provided, comprising the steps of:

[0056] (i) Based on the reference GTF file for full-length RNA transcripts described in the second aspect of the present invention, full-length transcript quantification is performed on RNA-seq sequencing data of tissue samples to obtain full-length RNA transcript expression information.

[0057] In another preferred embodiment, the full-length RNA transcript expression information includes quantitative results of the full-length RNA transcript expression profile.

[0058] In a fourth aspect of the invention, a tissue origin prediction system is provided, the system comprising:

[0059] (a) An input module configured to input full-length RNA transcript expression information of the sample to be tested;

[0060] (b) Evaluation module (or prediction module): The evaluation module (or prediction module) is configured to input the expression information of the full-length RNA transcript into a tissue origin prediction model constructed using the method described in the first aspect of the present invention, thereby obtaining a tissue origin prediction result for the sample to be tested; and

[0061] (d) Output module, which is configured to output the prediction result.

[0062] In another preferred embodiment, the system further includes (a0) a preprocessing module configured to analyze and determine the full-length RNA transcript expression information of the test sample based on the RNA-seq sequencing information of the test sample.

[0063] In another preferred embodiment, the system further includes (a0) a preprocessing module configured to analyze and determine the full-length RNA transcript expression information of the test sample using the method described in the third aspect of the present invention, based on the RNA-seq sequencing information of the test sample.

[0064] In another preferred embodiment, the system further includes (d) a control module configured to control the operation of the modules.

[0065] In another preferred embodiment, the system further includes (e) a storage module configured to store data including: hyperparameters, prediction probabilities, and preset reference thresholds of each tissue origin prediction sub-model in the tissue origin prediction model.

[0066] In another preferred embodiment, the preset reference threshold includes the preset reference thresholds of each sub-model in the tissue origin prediction model.

[0067] In another preferred embodiment, the judgment in the evaluation module includes:

[0068] (i) Input the full-length RNA transcript expression information into the tissue origin prediction model, and then input it into the sub-models of each tissue origin prediction model in sequence to obtain the prediction values ​​of all sub-models; wherein, each of the sub-models is used to independently determine whether the sample to be tested originates from a specific tissue;

[0069] (ii) Determine whether the predicted value of all sub-models is greater than or equal to the preset reference threshold C1 for the predicted tissue origin of its corresponding sub-model; the determination includes: if the predicted value of a sub-model is greater than or equal to the preset reference threshold C1 for the predicted tissue origin of its corresponding sub-model, then determine that the tissue is a candidate tissue of the sample to be tested; thereby obtaining all candidate tissues.

[0070] (iii) Sort all candidate tissues from largest to smallest according to their corresponding predicted values ​​to obtain the ranking results of candidate tissues, and obtain the candidate tissue with the largest predicted value as the tissue source of the test sample; that is, the tissue of the test sample comes from the candidate tissue with the largest predicted value.

[0071] In another preferred embodiment, when the sample to be tested is a tumor sample, the tissue origin prediction is the prediction result of the primary origin tissue of the tumor tissue.

[0072] In another preferred embodiment, the tissue origin prediction model comprises 18 sub-models, which are used to determine whether the tissue of the sample to be tested originates from the following sites: adrenal gland, bladder, bone marrow, brain, breast, intestine, eye, ganglion, liver, pancreas, prostate, thyroid, B cells, kidney, head and neck, skin, stomach, and uterus.

[0073] In a fifth aspect of the invention, a computer storage medium is provided for storing a computer program corresponding to the algorithm of the tissue origin prediction model constructed by the method described in the first aspect of the invention.

[0074] In a sixth aspect of the invention, a method for predicting the source of data organization is provided, comprising the steps of:

[0075] (a) Provide data: Provide RNA sequencing information for the sample to be tested;

[0076] (b) Data preprocessing: For the sequencing information, analyze and determine the expression information of full-length RNA transcripts of the sample to be tested, or use the method described in the third aspect of the present invention to analyze and determine the expression information of full-length RNA transcripts of the sample to be tested;

[0077] (c) Analysis and Evaluation: The expression information of the full-length RNA transcripts is input into the cancer prediction model constructed using the method described in the first aspect of this invention to obtain the evaluation results of the sample to be tested; and

[0078] (d) Output results: Output the evaluation results.

[0079] In another preferred embodiment, the evaluation process in the analysis and evaluation is as follows:

[0080] (i) Input the full-length RNA transcript expression information into the tissue origin prediction model, and then input it into each sub-model of the tissue origin prediction model in sequence to obtain the prediction values ​​of all sub-models; wherein, each sub-model is used to independently determine whether the sample to be tested originates from a specific tissue;

[0081] (ii) Determine whether the predicted value of all sub-models is greater than or equal to the preset reference threshold C1 for the predicted tissue origin of its corresponding sub-model; the determination includes: if the predicted value of a sub-model is greater than or equal to the preset reference threshold C1 for the predicted tissue origin of its corresponding sub-model, then determine that the tissue is a candidate tissue of the sample to be tested; thereby obtaining all candidate tissues.

[0082] (iii) Sort all candidate tissues from largest to smallest according to their corresponding predicted values ​​to obtain the ranking results of candidate tissues, and obtain the candidate tissue with the largest predicted value as the tissue source of the test sample; that is, the tissue of the test sample comes from the candidate tissue with the largest predicted value.

[0083] In another preferred embodiment, the preset reference threshold includes the preset reference thresholds of each sub-model in the tissue origin prediction model.

[0084] In another preferred embodiment, the method for obtaining the expression information of the full-length RNA transcript of the tissue to be tested based on the RNA sequencing information of the tissue to be tested is as follows:

[0085] The RNA-seq short read sequences of the tissue to be tested are aligned to the transcript library corresponding to the GTF file described in the second aspect of the present invention, and the expression level of each transcript in the sample is measured in the form of TPM (Transcript per million) to generate the full-length RNA transcript expression information of the tissue to be tested.

[0086] In another preferred embodiment, when the sample to be tested is a tumor sample, the tissue origin prediction is the prediction result of the primary origin tissue of the tumor tissue.

[0087] In another preferred embodiment, the method is an in vitro method.

[0088] In another preferred embodiment, the method is non-diagnostic and non-therapeutic.

[0089] It should be understood that, within the scope of this invention, the above-described technical features of this invention and the technical features specifically described below (such as in the embodiments) can be combined with each other to form new or preferred technical solutions. Due to space limitations, they will not be described in detail here. Attached Figure Description

[0090] Figures 1A-B show the ROC curves of the tissue origin sub-model in Embodiment 3 of the present invention when predicting different tissue origins.

[0091] Figure 2 shows an example schematic diagram of the tissue origin prediction model of the present invention (the result obtained in the figure is that the tissue originates from B cells), and the "Feat" in the figure refers to the feature.

[0092] Figure 3 shows a schematic diagram of a prediction module of an example of the present invention.

[0093] Figure 4 shows a schematic diagram of the tissue origin prediction system of the present invention, where "Feat" refers to a feature.

[0094] Figure 5 shows a schematic flowchart of a process for predicting the tissue origin of a population to be tested, according to an example of the present invention. Detailed Implementation

[0095] Through extensive and in-depth research, the inventors obtained tissue-specific RNA transcripts from 25 different tissues based on sequencing information from normal tissues. They then developed, for the first time, a model for predicting the tissue origin of test samples (tissue origin prediction model) (including prediction of the primary site of cancer samples). Specifically, the model can be used to diagnose or predict the tissue origin of 18 tissues: adrenal gland, bladder, bone marrow, brain, breast, intestine, eye, ganglion, liver, pancreas, prostate, thyroid, B cells, kidney, head and neck, skin, stomach, and uterus. Furthermore, when the test sample is cancerous tissue, the predicted tissue origin is the primary site or tissue of the cancer. The tissue origin prediction model of this invention has high predictive accuracy and can be used for early clinical auxiliary diagnosis. Based on this, the present invention was completed.

[0096] the term

[0097] To facilitate a clearer understanding of this disclosure, certain terms are first defined. As used herein, unless otherwise expressly specified herein, each of the following terms shall have the meaning given below. Other definitions are set forth throughout the application.

[0098] The term “about” can refer to a value or composition within an acceptable margin of error for a particular value or composition as determined by a person skilled in the art, depending in part on how the value or composition is measured or determined. For example, as used herein, the expression “about 100” includes all values ​​between 99 and 101 (e.g., 99.1, 99.2, 99.3, 99.4, etc.).

[0099] As used herein, the terms “containing” or “including (comprise)” can be open-ended, semi-closed, or closed. In other words, the terms also include “consistently made of” or “composed of”.

[0100] As used herein, unless otherwise stated, any concentration range, percentage range, proportion range, or integer range shall be understood to include any integer value within the range and, where appropriate, its fractional value (e.g., one-tenth and one-hundredth of an integer).

[0101] As used herein, the term “and / or” refers to and covers any and all possible combinations of one or more of the related listed items.

[0102] As used in this article, “transcription expression information,” “transcription quantification results,” “transcription expression level,” and “transcription expression value” have the same meaning and can be used interchangeably. They all refer to the expression level of each transcript in the sample, measured in TPM (Transcript per million), by aligning the RNA-seq short read sequence of the tissue sample to the transcript library corresponding to the GTF file constructed in this invention.

[0103] As used in this article, "full-length RNA transcript" and "full-length transcript" have the same meaning and can be used interchangeably. Both refer to the complete RNA molecule produced from a gene, including all parts of the coding and non-coding regions. It contains a series of information such as the 5' untranslated region (5'UTR), exons, introns (which are cut out in mature mRNA but retained in the full-length transcript), and the 3' untranslated region (3'UTR).

[0104] Tissue-specific RNA transcripts

[0105] This invention calculates normal tissue-specific RNA transcripts (Tissue-SRTs) and uses them to specifically identify the tissue origin of tumor tissue or normal tissue.

[0106] Normal tissue-specific RNA transcripts are defined as specific RNA transcripts that are highly expressed only in a single tissue and not expressed or expressed at very low levels in other normal tissues. These RNA transcripts are expressed only in a specific tissue.

[0107] Data sources used for calculating tissue-specific RNA transcripts include next-generation sequencing RNAseq data from 28 normal tissue samples in the public databases GTEx and GEO datasets, including adrenal glands, B cells, bladder, brain, breast, bone & muscle, bone marrow, cervix, intestine, esophagus, eye, ganglia, kidneys, liver, lungs, head and neck, ovaries, pancreas, prostate, skin, stomach, testes, thymus, thyroid gland, uterus, heart, spleen, and small intestine.

[0108] To obtain tissue-specific RNA transcripts, a specificity score for each transcript was calculated using expression profiles from multiple tissue types. Shannon entropy was used to measure tissue specificity, and its formula is as follows:

[0109] Where H t p represents the Shannon entropy of transcript t, N is the total number of tissue types, and p it This represents the expression proportion of transcript t in tissue type i. The expression proportion of each transcript across all tissue types is calculated as follows:

[0110] Where x it This represents the expression value of transcript t in tissue type i, and is the median value of transcript t in each tissue type i. H t The maximum value is log2(N), when H tWhen = log2(N), it indicates that the transcript t is expressed identically in all tissue types. H t The minimum value of 0 indicates that transcript t is specifically expressed only in a certain tissue type. Therefore, the Shannon entropy is converted into a specificity score, which is calculated as follows: S t =log2(N)-H t

[0111] After transformation, S t The maximum value is log2(N), and the minimum value is 0. Finally, the maximum expression ratio p of the transcript in tissue / cancer / cell type is calculated. mt It is the second largest expression ratio p nt More than twice the specificity score S t When the value is greater than 1, it is defined as a transcript specifically expressed in a certain tissue type.

[0112] Using analytical methods, 25 tissue-specific RNA transcripts were identified (3 of which were normal tissues without associated tumors; tissue-specific RNA transcripts from the small intestine, heart, and spleen were not counted). These included tissue-type-specific RNA transcripts from the adrenal gland, B cells, bladder, brain, breast, bone & muscle, bone marrow, cervix, intestine, esophagus, eye, ganglion, kidney, liver, lung, head and neck, ovary, pancreas, prostate, skin, stomach, testis, thymus, thyroid, and uterus (a list of 25 tissue-specific RNA transcripts expressed only in various normal tissues, unique to this invention).

[0113] The present invention provides a method for quantifying full-length RNA transcripts in tissue samples.

[0114] The method for quantifying full-length RNA transcripts in tissue samples according to the present invention includes the following steps:

[0115] (St1): 177 tumor tissue specimens from patients with hepatocellular carcinoma, colorectal cancer, ovarian cancer, breast cancer, nasopharyngeal carcinoma, neuroendocrine tumors, gastric cancer, gastrointestinal stromal tumors, renal cancer, and cervical cancer were collected. RNA was extracted from the collected tissues and PacBio Iso-seq third-generation sequencing was performed. The raw PacBio sequencing data were processed using the Iso-Seq workflow.

[0116] (St2): Collected third-generation PacBio and ONT transcriptome sequencing data from 21 cancer tissue samples (including esophageal cancer, lung cancer, glioblastoma, leukemia, lymphoma, melanoma, myeloma, and sarcoma) and 153 normal tissue samples (including adipose tissue, adrenal gland, brain, breast, heart, liver, lung, muscle, ovary, pancreas, testis, fibroblasts, GM12878, H1, H9, HEK293T, and WTC11 cells) from public databases.

[0117] (St3): FASTQ sequences from PacBio and ONT data were aligned using minimap2 and processed using a standard pipeline to obtain merged GTF transcriptomes. Full-length transcripts from all samples were merged into non-redundant transcriptome GTF files using gffcompare software.

[0118] (St4): If all splicing sites of the transcript are detected in at least 5 samples in TCGA and GTEx for a certain cancer / tissue RNA-seq in the merged GTF file, then the transcript is retained.

[0119] (St5): Using SQANTI3 and gffcompare software, these transcripts were compared with the reference transcriptome (GENCODE v.35). The transcripts were mainly divided into four groups: fully spliced ​​matched, partially spliced ​​matched, novel within the catalog, and novel outside the catalog. All partially spliced ​​matched transcripts, which may be the result of RNA degradation and incomplete reverse transcription, were filtered out from the entire transcriptome.

[0120] (St6): Cage peaks and polyA sites were annotated using SQANTI3. For novel transcripts both in-directory and out-of-directory, transcripts with 16 or more adenines within 20 bases downstream of the annotated transcription termination site were filtered.

[0121] (St7): The full-length transcript sequence was obtained using the gffread software based on the transcript location provided by the GTF. ORFs in the transcripts from the third-generation sequencing data were predicted using GeneMark. When the stop codon was located more than 50 nucleotides upstream of the last exon-exon splicing site, the predicted transcript would induce meaningless mediated decay.

[0122] (St8): Transposon elements from the entire human hg38 genome were downloaded from the UCSC Genome Browser, and BEDTools software was used to determine whether transcripts overlapped with TEs within a TSS ± ​​100 bp. Similar transcripts were merged according to the following criteria: i) A maximum of three exon differences were allowed. ii) Differences per exon did not exceed four bases. iii) The maximum allowed dissimilarity for transcription start and termination sites was 100. iv) For coding transcripts, it was ensured that there were no differences in ORFs between transcripts.

[0123] (St9): For similar transcripts, the longest one is retained and other transcripts are filtered out. A reference GTF file containing 1,069,895 full-length RNA transcripts is created for full-length transcript quantification of RNAseq data from tissue samples.

[0124] (St10): During data processing, Salmon software can be used to create an index for the GTF file containing 1,069,895 full-length RNA transcripts obtained above.

[0125] (St11): By aligning short reads of RNA-seq to the transcript library corresponding to the GTF file and measuring the expression level of transcripts in each sample in the form of TPM (Transcript per million), a quantitative method for quantifying the expression profile of full-length RNA transcripts in a sample using RNA-seq data was developed.

[0126] Machine learning-based organization origin prediction model

[0127] The tissue origin prediction model based on machine learning in this invention relies on the specific biological characteristics (tissue-specific transcripts, Tissue-SRTs) of different normal tissues. Machine learning is used to set positive and negative thresholds to construct the prediction model. Based on the expression characteristics of each tissue-specific transcript, a tissue origin prediction model is established to determine the tissue origin of cancer. The tissue origin prediction system constructed based on the tissue origin prediction model predicts or diagnoses the origin of the tissue by using the specific transcript (SRT) data expressed by the tissue under test. When the sample or tissue under test is tumor tissue, the tissue origin prediction model of this invention can be used to predict the primary site of the sample.

[0128] Machine learning, a branch of artificial intelligence, uses statistical and computer science methods to enable computer systems to learn from data and automatically improve. It analyzes and identifies patterns and regularities in data to make predictions and decisions. In the field of oncology, it can analyze patient data such as clinical symptoms, medical imaging, laboratory tests, and histopathology to discover new patterns and correlations between variables and generate predictive models. By applying machine learning techniques, the field of oncology diagnosis can extract valuable information from large amounts of complex data, helping doctors provide more accurate and faster diagnostic results and improving treatment outcomes and survival rates for cancer patients.

[0129] Machine learning models were used to construct the final model of this invention during the model building process. These machine learning models include: Random Forest model, Decision Tree model, XGBoost model, LightGBM model, and CatBoost model.

[0130] Tissue origin prediction of the present invention

[0131] This invention provides several applications of the tissue origin prediction model, including the development of tissue origin-assisted diagnostic software, devices, or systems. Figure 5 shows a schematic flowchart of the tissue origin prediction process for the population to be tested based on the tissue origin-assisted diagnostic model of this invention.

[0132] As shown in Figure 5, the first step is to obtain the full-length transcript expression information of the target population. This involves: obtaining biopsy tissue samples from the target population; establishing a full-length transcript reference GTF file through third-generation sequencing data processing and transcript annotation filtering; using this file for quantifying RNA-seq transcripts in the tissue samples; aligning short RNA-seq sequences to a transcript library already obtained through long-read sequencing; and measuring the expression level of transcripts in each sample in the form of TPM (Transcripts per Million) to generate the expression information of the full-length RNA transcripts of the target tissue. This transcript expression information includes tissue-specific RNA transcripts from the target tissue.

[0133] The expression information of the full-length RNA transcript is input into the tissue origin prediction model. A sub-model corresponding to a specific tissue origin is then determined from this model. The expression information of the full-length RNA transcript is input into this sub-model to obtain its output, which is the predicted value indicating that the tested tissue belongs to the corresponding tissue origin. The next sub-model corresponding to a tissue origin is determined by inputting the expression information of the full-length RNA transcript into it, and so on, until the outputs of all sub-models are determined. If the tested tissue is determined to include cancerous tissue, the tissue origin of the tested tissue is determined based on the output of the tissue origin prediction model.

[0134] In another preferred embodiment, when predicting tissue origin, each time a determination is made as to whether the tissue origin is one of the specified tissue types, for example, when determining whether the tissue origin is adrenal gland, a sub-model for adrenal gland origin is obtained. Based on this sub-model, a set of adrenal gland feature names is determined, and corresponding features are obtained from tissue-specific transcripts and input into the classifier of the corresponding sub-model. The corresponding classifier can be determined based on its name. The classifier has a preset classifier name and parameters, which can be determined based on the corresponding sub-model in the preset tissue origin model. The classifier parameters include pre-determined hyperparameters to determine whether the tissue origin is adrenal gland tissue. The output of the sub-model is the predicted value of the tissue origin of the tested tissue corresponding to the sub-model. If the predicted value of the tested tissue belonging to the corresponding tissue origin of the sub-model is greater than a preset tissue origin prediction threshold, then the tissue origin of the tested tissue is determined to include that tissue origin; if the predicted value of the tested tissue belonging to the corresponding tissue origin of the sub-model is less than the preset tissue origin prediction threshold, then the tissue origin of the tested tissue is determined to not include that tissue origin.

[0135] Continuing in the same manner, determine whether the tissue origin is from other tissues, obtaining the outputs of all sub-models. Finally, generate the output of the tissue origin prediction model based on the tissue origin corresponding to one or more sub-models whose predicted output value is greater than the tissue origin prediction threshold. For example, if the output of one sub-model shows that the tissue origin of the tested tissue includes adrenal glands, and the output of another sub-model shows that the tissue origin includes B cells, then the output of the tissue origin prediction model includes both adrenal glands and B cells.

[0136] In another preferred embodiment, the tissue origin prediction device, software, or system of the present invention includes modules selected from the group consisting of: an input module, a preprocessing module, a prediction module (or an evaluation module), and an output module.

[0137] In another preferred embodiment, the tissue origin prediction system of the present invention makes predictions based on the expression information of full-length RNA transcripts of patient tissues (including tissue-specific transcript (SRT) data) for the diagnosis or prediction of the tissue origin of the sample to be tested, and the prediction indicators are objective.

[0138] The technical solution of this invention is based on a pre-trained tissue origin prediction model that predicts the probability of the tissue origin of a current sample (including the probability of the primary site of cancer tissue) based on the genetic information of the sample tissue. The genetic feature information includes ti-SRT (tissue-specific transcripts). All these features are obtained by analyzing tissue samples through long-read and short-read RNA sequencing technologies and integrating third-generation RNA sequencing data from various normal and cancerous human tissues and cells. Tissue-specific RNA transcripts are obtained by analyzing a large number of case data. Five models, namely Random Forest, Decision Tree, XGBoost, LightGBM, and CatBoost, are used to obtain the hyperparameters of each model using grid search for each feature. The roc_auc of the five models under their hyperparameters is compared, and the model with the highest roc_auc and its hyperparameters are selected to obtain five prediction models for all types of tumors after training. Thus, the probability of diagnosing the sample as cancer and the probability of predicting the origin of cancer are used as outcome indicators, which is very objective. Tissue origin prediction systems can assist in diagnosing the primary and / or metastatic status of tumors, and can also improve the diagnostic accuracy of tumors with unknown primary sites.

[0139] Furthermore, the system is independent of pathologists, requires less human intervention, and can provide standardization and consistency, offering pathologists highly stable tumor auxiliary diagnosis. This reduces the burden on physicians and improves work efficiency, demonstrating high clinical translational value and market potential.

[0140] It should be noted that, in the apparatus or device of this application, the storage medium (computer-readable medium) used to store the model or algorithm of the present invention can be a computer-readable signal medium or a non-transitory computer-readable storage medium, or any combination thereof. A non-transitory computer-readable storage medium can be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a non-transitory computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0141] In this application, a non-transitory computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a non-transitory computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination thereof.

[0142] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0143] The main advantages of this invention include:

[0144] (a) The tissue origin prediction system of the present invention can assist in the diagnosis of the primary and / or metastatic status of tumors, and can also improve the diagnostic accuracy of tumors with unknown primary sites and multiple primary tumors, helping doctors to provide more accurate and faster diagnostic results, improve diagnostic efficiency, and has good clinical application value and market prospects.

[0145] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Experimental methods in the following embodiments, unless otherwise specified, are generally performed under conventional conditions, such as those described in Sambrook et al., Molecular Cloning: A Laboratory Manual (New York: Cold Spring Harbor Laboratory Press, 1989), or as recommended by the manufacturer. Unless otherwise stated, percentages and parts are weight percentages and parts by weight.

[0146] Example 1: Construction of GTF Files

[0147] 1.1 Data Acquisition

[0148] We collected 177 tumor tissue specimens from patients with hepatocellular carcinoma, colorectal cancer, ovarian cancer, breast cancer, nasopharyngeal carcinoma, neuroendocrine tumors, gastric cancer, gastrointestinal stromal tumors, renal cancer, and cervical cancer. RNA was extracted from the collected tissues and then subjected to PacBio Iso-seq third-generation sequencing and second-generation sequencing (this part of the data was generated by the sequencing of our own tissues). The raw PacBio sequencing data was processed using the Iso-Seq workflow.

[0149] We collected third-generation PacBio and ONT transcriptome sequencing data and second-generation sequencing data from 21 cancer tissue samples (including esophageal cancer, lung cancer, glioblastoma, leukemia, lymphoma, melanoma, myeloma, and sarcoma) and 153 normal tissue samples (including adipose tissue, adrenal gland, brain, breast, heart, liver, lung, muscle, ovary, pancreas, testis, fibroblasts, GM12878, H1, H9, HEK293T, and WTC11 cells) from public databases.

[0150] 1.2 Building GTF Files

[0151] FASTQ sequences from PacBio and ONT data were aligned using minimap2 and then processed using a standard pipeline to obtain merged GTF transcriptomes. Full-length transcripts from all samples were merged into non-redundant transcriptome GTF files using gffcompare software.

[0152] Non-redundant transcriptome GTF files were optimized using various methods. In the merged GTF file, if at least 5 samples in TCGA and GTEx for a certain cancer / tissue RNA-seq showed all splicing sites of the transcript, the transcript was retained.

[0153] These transcripts were compared with the reference transcriptome (GENCODE v.35) using SQANTI3 and gffcompare software. The transcripts were mainly divided into four groups: fully spliced, partially spliced, novel in the catalog, and novel out of the catalog.

[0154] All incomplete splice-matched transcripts, which may be the result of RNA degradation and incomplete reverse transcription, should be filtered out from the entire transcriptome.

[0155] Cage peaks and polyA sites were annotated using SQANTI3. For novel transcripts both in-directory and out-of-directory, transcripts with 16 or more adenines within 20 bases downstream of the annotated transcription termination site were filtered.

[0156] The full-length transcript sequence was obtained using the gffread software based on the transcript location provided by the GTF. GeneMark was used to predict ORFs in the transcripts from the third-generation sequencing data. When the stop codon was located more than 50 nucleotides upstream of the last exon-exon splicing site, the predicted transcript would induce meaningless mediated decay.

[0157] Transposon elements from the entire human hg38 genome were downloaded from the UCSC Genome Browser, and BEDTools software was used to determine whether transcripts within the TSS ± ​​100 bp overlapped with TEs. Similar transcripts were merged according to the following criteria: i) a maximum of three exon differences were allowed. ii) differences per exon did not exceed four bases. iii) the maximum allowed dissimilarity for transcription start and termination sites was 100. iv) for coding transcripts, it was ensured that there were no differences in ORFs between transcripts.

[0158] For similar transcripts, the longest one was retained, and other transcripts were filtered out, creating a reference GTF file containing 1,069,895 full-length RNA transcripts. This file was used for full-length transcript quantification of RNAseq data from tissue samples (this newly generated GTF file was developed by the inventors based on their own sequencing data and analysis of public data).

[0159] During data processing, Salmon software can be used to create an index on the GTF file containing 1,069,895 full-length RNA transcripts obtained above (creating an index based on the newly generated GTF file is unique to this invention).

[0160] The RNA-seq short read sequences of the above samples (data sources include public databases and second-generation sequencing data obtained from the above self-test) were aligned to the transcript library corresponding to the GTF file, and the expression level of transcripts in each sample was measured in the form of TPM (Transcript per million). A quantitative method for quantifying the expression profile of full-length RNA transcripts in samples using RNA-seq data was developed.

[0161] Transcript quantification was performed on RNAseq data from 33 TCGA tumors and 28 normal tissues in public databases (TCGA, GTEX, and GEO). The FLIBase database based on full-length RNA transcripts was developed, creating a comprehensive resource library of full-length transcript information from human cancer tissues and normal tissues.

[0162] This database covers multiple cancer types and different normal tissue types, identifies a large number of unannotated tumor-specific RNA transcripts, and provides a panoramic view of tissue-specific RNA transcripts in different normal tissues. FLIBase has a significant advantage in discovering a large number of previously unannotated isoforms and tumor-specific RNA transcripts. Utilizing this database for in-depth transcriptome mining and research at the transcript level is expected to yield new insights into precision diagnosis and treatment of cancer.

[0163] Example 2: Calculation of tissue-specific RNA transcripts

[0164] Transcript quantification and analysis were performed on RNA-seq data from 28 normal tissues in public databases (GTEx and GEO) to determine tissue-specific RNA transcripts for each tissue sample.

[0165] The number of samples for each normal tissue in the public databases GTEx and GEO datasets is shown in Table 1.

[0166] Table 1

[0167] To obtain tissue-specific RNA transcripts, a specificity score for each transcript was calculated using expression profiles from multiple tissue types. Shannon entropy was used to measure tissue specificity, and its formula is as follows:

[0168] Where H t p represents the Shannon entropy of transcript t, N is the total number of tissue types, and p it This represents the expression proportion of transcript t in tissue type i. The expression proportion of each transcript across all tissue types is calculated as follows:

[0169] Where x it This represents the expression value of transcript t in tissue type i, and is the median value of transcript t in each tissue type i. H t The maximum value is log2(N), when H t When = log2(N), it indicates that the transcript t is expressed identically in all tissue types. H t The minimum value of 0 indicates that transcript t is specifically expressed only in a certain tissue type. Therefore, the Shannon entropy is converted into a specificity score, which is calculated as follows: S t =log2(N)-H t

[0170] After transformation, S t The maximum value is log2(N), and the minimum value is 0. Finally, the maximum expression ratio p of the transcript in tissue / cancer / cell type is calculated. mt It is the second largest expression ratio p nt More than twice the specificity score S t When the value is greater than 1, it is defined as a transcript specifically expressed in a certain tissue type.

[0171] Using analytical methods, tissue-specific RNA transcripts were identified for 25 normal tissues (3 of which had no associated tumors; tissue-specific RNA transcripts for the small intestine, heart, and spleen were not counted). These included tissue-specific RNA transcripts (Tissue-SRT) from the adrenal gland, B cells, bladder, brain, breast, bone & muscle, bone marrow, cervix, intestine, esophagus, eye, ganglion, kidney, liver, lung, head and neck, ovary, pancreas, prostate, skin, stomach, testis, thymus, thyroid, and uterus. (This list of tissue-specific RNA transcripts for 25 tissues expressed only in various normal tissues was calculated by the inventors and is a first and unique invention of this invention.)

[0172] Example 3: Construction of a Tissue Origin Prediction Model

[0173] The tissue origin prediction model of the present invention is used to distinguish which tissue a given tissue sample originates from, or which type of cancer (including the type of carcinoma in situ that has metastasized) is in a cancer patient.

[0174] 3.1 Sample Data

[0175] Sample data source: The sample data includes RNA-seq data of 33 cancer tissues from the TCGA dataset in the public database, as shown in Table 2.

[0176] Table 2 shows the number of cancer tissue samples for 33 types of cancer in the TCGA dataset from the public database.

[0177] 3.2 Model Construction

[0178] 3.2.1 Construction of Sub-models

[0179] In the model construction process, cancer samples from each specific tissue origin in the TCGA dataset were designated as positive samples in the training set, while cancer samples from other tissue origins were designated as negative samples. First, tissue origin prediction sub-models were constructed for each of the 25 different tissues; that is, for any given tissue, a corresponding sub-model was built to determine whether the sample originated from that tissue. Then, the optimized sub-models were combined to construct the tissue origin prediction model.

[0180] Therefore, model construction includes the creation of sub-models for prediction of each tissue. Specifically, this modeling process involves building sub-models for prediction of the following 25 tissue origins: adrenal gland, B cell, bladder, brain, breast, bone & muscle, bone marrow, cervix, intestine, esophagus, eye, ganglion, kidney, liver, lung, head and neck, ovary, pancreas, prostate, skin, stomach, testis, thymus, thyroid, and uterus.

[0181] Based on organ type, the 32 types of cancer in TCGA (excluding mesothelioma (MESO) due to the lack of corresponding RNAseq data from normal peritoneum and pleura) were categorized into 25 organ-derived cancers (this classification is based solely on organ tissue origin, involving a total of 25 calculations, each distinguishing one tissue origin per run), as shown in Table 3. The feature count in Table 3 represents the number of tissue-specific RNA transcripts for each tissue calculated in Example 2.

[0182] Table 3. List of tissue origins and corresponding cancer types

[0183] All cancer tissue samples in the TCGA dataset were classified according to the list of tissue origin and cancer type, namely: adrenal gland, B cell, bladder, brain, breast, bone & muscle, bone marrow, cervix, intestine, esophagus, eye, ganglion, kidney, liver, lung, head and neck, ovary, pancreas, prostate, skin, stomach, testis, thymus, thyroid, uterus, and MESO, totaling 26 types.

[0184] The classification method shown in Table 3 is as follows: Based on 25 tissue-specific RNA transcripts, a binary classification method is used, with one feature used for each classification, for a total of 25 calculations. (After 25 training iterations, the total number of features corresponding to the feature counts of different tissues represents the number of tissue-specific RNA transcripts.)

[0185] The first time, using 1947 adrenal gland characteristics, positive results were identified as ACC cancer tissue, while negative results were identified as other 25 types of cancer tissue, thus differentiating adrenal cancer from other 25 types of cancer.

[0186] The second time, using 1292 B-cell characteristics, positive results were identified as DLBC cancer tissue, while negative results were identified as other 25 types of cancer tissue, to differentiate lymphoma from the other 25 types of cancer.

[0187] The third time, using 573 bladder characteristics, positive results were identified as BLCA cancer tissue, while negative results were identified as 25 other types of cancer tissue, thus differentiating bladder cancer from the other 25 types of cancer.

[0188] The fourth time, using 5466 brain features, positive results were identified as cancerous tissues of GBM and LGG, while negative results were identified as cancerous tissues of the other 25 types, thus distinguishing brain tumors from the other 25 types of cancer.

[0189] The fifth time, using 786 breast features, positive results were identified as BRCA-positive cancer tissue, while negative results were identified as 25 other types of cancer tissue, thus differentiating breast cancer from the other 25 types of cancer.

[0190] The sixth time, using 17,668 bone and muscle features, positive results were identified as SARC cancer tissue, while negative results were identified as other 25 types of cancer tissue, thus differentiating sarcoma from the other 25 types of cancer.

[0191] In the 7th study, 24,759 bone marrow characteristics were used to distinguish between leukemia and 25 other types of cancer tissues, with positive results indicating LAML and negative results indicating other types of cancer tissues.

[0192] The 8th time, cervical characteristics were used to distinguish cervical cancer from 25 other types of cancer by identifying positive cervical cancer tissue and negative cervical cancer tissue.

[0193] In the 9th study, 209 intestinal characteristics were used, with positive results indicating cancerous tissues of COAD and READ, and negative results indicating 25 other types of cancerous tissues, to differentiate intestinal cancer from 25 other types of cancer.

[0194] In the 10th study, 103 esophageal features were used to distinguish esophageal cancer from ESCA cancer tissue (positive results) and other 25 types of cancer tissue (negative results).

[0195] In the 11th study, 13,535 eye features were used to distinguish between uveal melanoma and 25 other types of cancer tissues, with positive results identified as UVM-positive cancer tissues and negative results identified as negative results.

[0196] In the 12th study, using 69,670 ganglion features, positive results were identified as cancerous tissues of PCPG, while negative results were identified as cancerous tissues of 25 other types, thus differentiating paragangliomas from the other 25 types of cancer.

[0197] In the 13th study, 1482 kidney characteristics were used to distinguish between kidney cancer and 25 other types of cancer, with positive results for KICH, KIRC, and KIRP cancer tissues and negative results for other types of cancer tissues.

[0198] In the 14th study, 21,988 liver characteristics were used to distinguish liver cancer from 25 other types of cancer, with positive results for LIHC and CHOL cancer tissues and negative results for other types of cancer tissues.

[0199] In the 15th study, 1449 lung characteristics were used to differentiate between lung cancer and 25 other types of cancer tissues, with positive results for LUAD and LUSC and negative results for other types of cancer tissues.

[0200] In the 16th study, 1357 head and neck features were used to differentiate between head and neck squamous cell carcinoma and other 25 cancers. Positive results were identified as HNSC cancer tissue, while negative results were identified as other 25 cancer tissues.

[0201] In the 17th study, 2,615 ovarian characteristics were used to distinguish between ovarian cancer and 25 other types of cancer tissues, with positive results indicating OV and negative results indicating other types of cancer tissues.

[0202] In the 18th study, 1135 pancreatic characteristics were used to distinguish pancreatic cancer from PAAD cancer tissue (positive results) and other 25 types of cancer tissue (negative results).

[0203] In the 19th study, 2146 prostate characteristics were used to distinguish prostate cancer from PRAD cancer tissue (positive results were identified as cancerous tissue of PRAD) and other 25 types of cancer tissue (negative results were identified as cancerous tissue of other types).

[0204] In the 20th study, 2273 skin characteristics were used to differentiate between skin melanoma and 25 other types of cancer. Positive results indicated cancerous tissue of SKCM, while negative results indicated cancerous tissue of 25 other types of cancer.

[0205] In the 21st study, 384 gastric characteristics were used to distinguish between gastric cancer and other 25 types of cancer tissues, with positive results indicating STAD and negative results indicating STAD.

[0206] In the 22nd study, 53,097 testicular characteristics were used to distinguish testicular cancer from 25 other types of cancer by identifying positive TGCT findings and negative findings from other types of cancer tissue.

[0207] In the 23rd study, 1685 thymic characteristics were used to distinguish thymic carcinoma from other 25 types of cancer tissues, with positive results indicating THYM cancer and negative results indicating other 25 types of cancer tissues.

[0208] In the 24th instance, 2821 thyroid characteristics were used to distinguish thyroid cancer from 25 other types of cancer tissues, with positive results indicating THCA cancer tissues and negative results indicating 25 other types of cancer tissues.

[0209] In the 25th study, 989 uterine characteristics were used to differentiate uterine tumors from 25 other types of cancer tissues. Positive results were identified as cancerous tissues of UCEC and UCS, while negative results were identified as cancerous tissues of 25 other types.

[0210] Taking the tissue origin prediction sub-model (first sub-model) used to determine whether a test sample originates from the adrenal gland as an example, the construction process is as follows:

[0211] (a) Dataset distribution: RNAseq data from various tissue sources in the TCGA dataset were divided into training and test sets in a 3:1 ratio.

[0212] (b) Sample classification: The samples in the TCGA dataset are classified as follows: cancers originating from the adrenal gland (such as adrenocortical carcinoma) are positive samples, and cancers originating from non-adrenal glands (i.e., cancers from other tissue sources) are negative samples.

[0213] (c) Obtain the expression information (or quantitative results) of full-length RNA transcripts in the positive and negative samples according to the full-length RNA transcript quantification method of tissue samples; specifically, this includes aligning the RNA-seq short read sequences of the samples to the transcript library corresponding to the GTF file constructed in Example 1, and measuring the expression level of transcripts in each sample in the form of TPM (Transcript per million) to obtain the expression information (or quantitative results) of full-length RNA transcripts in tissue samples.

[0214] (d) Using the expression information of the full-length RNA transcripts of the tissue samples, the hyperparameters of the machine learning model to be trained are determined by grid search. During the grid search, each parameter combination is cross-validated, such as using 5-fold cross-validation with ROC AUC as the scoring function. Finally, the optimal hyperparameters are determined.

[0215] The machine learning models include: Random Forest, Decision Tree, XGBoost, LightGBM, and CatBoost. When determining hyperparameters, the expression information of the full-length RNA transcripts of the tissue sample is input into each of these five machine learning models to obtain the optimal hyperparameters for each model. Based on this, five tissue origin prediction sub-models (or the first sub-model) are obtained to determine whether the sample to be tested originates from the adrenal gland.

[0216] (e) Validate and evaluate the five tissue origin prediction sub-models trained above using test set data, including by scoring using the ROC AUC function.

[0217] (f) The five tissue origin prediction sub-models trained above were validated and evaluated using validation set data, including methods such as scoring using the ROC AUC function. Finally, the best-performing model was selected from the five sub-models as the final tissue origin prediction sub-model (second sub-model) used to determine whether the test sample originated from the adrenal gland. The validation set consisted of RANSeq data of cancer tissues from 33 cancers obtained from the public databases GEO and dbGAP independent datasets, as shown in Table 4.

[0218] Table 4. Cancer tissue sample data for 33 types of cancer from the public databases GEO and dbGAP.

[0219] The other 24 sub-models, which predict different tissue origins, all adopt the same construction process as the tissue origin prediction sub-model used to determine whether the test sample originates from the adrenal gland. Finally, they are used to determine whether the test sample originates from any other tissue.

[0220] Finally, a total of 25 tissue origin prediction sub-models were obtained, which were used to determine whether the sample to be tested originated from the following 25 tissues: adrenal gland, B cell, bladder, brain, breast, bone & muscle, bone marrow, cervix, intestine, esophagus, eye, ganglion, kidney, liver, lung, head and neck, ovary, pancreas, prostate, skin, stomach, testis, thymus, thyroid gland, and uterus.

[0221] The AUC values ​​of the five machine learning models corresponding to each organization source prediction, namely the five organization source prediction sub-models (or the first sub-model) corresponding to each organization prediction, on the training set, test set, and validation set are shown in Table 2 below.

[0222] Table 5

[0223] As shown in Table 5, the five first sub-models performed well across all evaluation metrics in both the training and test sets. Further validation of model accuracy can be achieved using a validation set; other datasets can be used to verify whether the model possesses the same predictive ability on an independent dataset.

[0224] The results showed that, in the validation set, among the five first-order models, each tissue had its own best-performing algorithm model in predicting various evaluation metrics for 25 tissue origins (covering 32 types of cancer); the best-performing algorithm model would be used as the final tissue origin prediction sub-model (second sub-model) for that tissue.

[0225] Among them, the AUC of the best algorithm for the prediction sub-models of 18 cancer origins (covering 24 types of cancer) was greater than 0.9, including adrenal gland, bladder, bone marrow, brain, breast, intestine, eye, ganglion, liver, pancreas, prostate, thyroid, B cell, kidney, head and neck, skin, stomach and uterus, which proved the effectiveness of this tissue origin prediction model in predicting the tissue origin of most cancers.

[0226] Figures 1A and 1B show the ROC curves of a tissue origin prediction model according to an example of the present invention when predicting different tissues.

[0227] 3.2.2 Model Construction and Validation

[0228] In section 3.2.1, the final selection of the prediction sub-models or second sub-models for each organization's source is shown in Table 6.

[0229] Table 6: Selection of Sub-models in the Organization Origin Prediction Model

[0230] According to Table 6, this invention selected 18 high-performing sub-models with AUC ≥ 0.9 for constructing the tissue origin prediction model. This invention constructs the tissue origin prediction model by using specific transcripts in the training set samples for modeling. The algorithm model is then validated on the test set and / or validation set to determine its final form.

[0231] The tissue origin prediction model of this invention predicts the tissue origin of a patient's cancer based on the expression information of full-length RNA transcripts (including tissue-specific transcript (SRT) data) from the patient's tumor tissue. The optimal algorithm for each of the 18 cancer tissue origin prediction models (covering at least 24 cancers) has an AUC greater than 0.9, including adrenal glands, bladder, bone marrow, brain, breast, intestine, eye, ganglia, liver, pancreas, prostate, thyroid, B cells, kidney, head and neck, skin, stomach, and uterus. Therefore, in the application of this model, tissue origin prediction only involves predicting the origin of 18 tissues (including cancerous tissue) (covering at least the 24 cancers verified by this invention).

[0232] A schematic diagram of the tissue origin prediction model of the present invention is shown in Figure 2.

[0233] When predicting tissue origin, each step determines whether the tissue origin is one of the specified tissue types. For example, when determining whether the tissue origin is adrenal gland, the corresponding features are obtained from tissue-specific transcripts based on the adrenal gland feature name set and input into the classifier of the corresponding sub-model. The classifier can be determined based on its name. The classifier has a preset name and parameters, which can be determined based on the corresponding sub-model in the preset tissue origin model. The classifier parameters include pre-determined hyperparameters to determine whether the tissue origin is adrenal gland tissue. If the tissue origin is not adrenal gland, the determination of whether the origin is B cell is continued in the same manner, and if so, it is determined to be B cell. As an example of this invention, Figure 2 shows the prediction result for determining the tissue origin as B cell.

[0234] Example 4: Tissue-based diagnostic system and device

[0235] Based on the tissue origin prediction model constructed in this invention, this invention has developed a tissue origin prediction system or device. The development process includes the development of software for tissue origin prediction.

[0236] In addition to the input and output modules, the aforementioned tissue origin prediction system also includes a preprocessing module and a prediction module (or evaluation module).

[0237] The preprocessing module quantifies the expression values ​​of full-length RNA transcripts in the sample to be tested.

[0238] Figure 3 shows a schematic diagram of the prediction module of an example of the present invention. This module extracts the normal tissue-specific RNA transcript values ​​of the sample to be tested based on a set of feature names and inputs them into the corresponding classifiers. The classifier (sub-model) and its parameters for each tissue category are determined by a fixed configuration file. Each classifier predicts the probability of whether the input sample belongs to its category and its threshold.

[0239] The prediction module is configured to respond to receiving the full-length RNA transcript expression information of the test tissue and input it into a pre-trained tissue origin prediction model; the tissue origin of the test tissue is determined based on the output of the tissue origin prediction model.

[0240] Taking the adrenal gland as an example, after inputting the expression information of the full-length RNA transcript of the tissue to be tested into the tissue origin prediction model, candidate prediction results are obtained, including prediction probability and preset reference threshold. When the prediction probability in the prediction results is greater than or equal to the preset probability threshold, it is judged as positive, which is used to characterize that the source of the target sample (or the sample to be tested) is the adrenal gland.

[0241] Figure 4 shows a schematic diagram of a tissue origin prediction system according to an example of the present invention.

[0242] As shown in Figure 4, when a sample requires diagnosis, the prediction module reads the sample file and divides it into 18 test sets based on features. It then performs tissue origin prediction.

[0243] After inputting the expression information of the full-length RNA transcript of the tissue to be tested into the tissue origin prediction model, a sub-model corresponding to a specific tissue origin is determined from the tissue origin prediction model. The expression information of the full-length RNA transcript is then input into the sub-model corresponding to the tissue origin, and the output of the sub-model is obtained. The output of the sub-model is the predicted value of the tissue to be tested belonging to the corresponding tissue origin of the sub-model. The next sub-model corresponding to the tissue origin is determined, and the expression information of the full-length RNA transcript is input into the next sub-model corresponding to the tissue origin, and the output of the next sub-model is obtained, until the outputs of all sub-models are determined.

[0244] The output of the tissue origin prediction model is generated based on the tissue origin corresponding to one or more sub-models whose output predicted value is greater than the tissue origin prediction threshold, thus obtaining the predicted probabilities corresponding to multiple candidate primary sites (or tissue origins). The predicted probabilities in the candidate prediction results are then sorted to obtain a prediction probability ranking result.

[0245] The tissue origin prediction result is determined based on the candidate primary site corresponding to the highest predicted probability in the predicted probability ranking results. When the highest predicted probability is greater than or equal to a preset probability threshold, the candidate primary site corresponding to the highest predicted probability is taken as the tissue origin prediction result. Alternatively, when the highest predicted probability is less than the preset probability threshold, the candidate primary sites corresponding to the top three predicted probabilities in the predicted probability ranking results are taken as the tissue origin prediction results. The predicted probability ranking results are arranged in descending order of probability value.

[0246] The following is information related to software development:

[0247] Software Name: Machine Learning-Based Tissue-Based Diagnostic System for Specific RNA Transcripts

[0248] Hardware environment for development

[0249] CPU AMD Ryzen 7 6800U with Radeon Graphics

[0250] 64GB RAM (2100MHz)

[0251] The operating system used to develop this software

[0252] Distributor ID: Ubuntu

[0253] Description: Ubuntu 22.04.3LTS

[0254] Release: 22.04

[0255] Codename: jammy

[0256] Software development environment / tools: Vim 8.2

[0257] Programming language: Python 3.

[0258] The output module of the tissue origin prediction device, software, or system of the present invention is used to output the tissue origin of the sample to be tested based on the prediction results of the tissue origin prediction model. The tissue origin of the tissue to be tested may include one or more. If the tissue to be tested is cancerous tissue, the tissue origin of the cancer may include one or more.

[0259] Example 5 External Verification

[0260] The present invention also obtains a first external detection set and a second external detection set, and performs first external verification and second external verification respectively.

[0261] 5.1 First External Validation

[0262] The first external validation set is used to evaluate the accuracy of the tissue origin prediction model, device, or system of the present invention in diagnosing primary and metastatic cancers. The data in this first external validation set is proprietary detection data of the present invention, specifically including: 128 cases of hepatocellular carcinoma, 16 cases of colorectal cancer, 1 case of renal cancer, 1 case of gastric cancer, and 34 cases of liver metastases (12 cases of breast cancer liver metastases, 19 cases of colorectal cancer liver metastases, and 3 cases of gastric cancer liver metastases).

[0263] By inputting tissue-specific RNA transcripts from cancer samples of different origins or primary sites into the tissue origin prediction model, device, or system of the present invention, prediction or diagnostic results are obtained as shown in Tables 7 and 8 below.

[0264] Table 7 Diagnostic accuracy rate of tumor origin tracing (tissue source)

[0265] Table 8. Diagnostic or predictive results for some patients with multiple cancers. Note: NA is below the threshold. The score (or predicted probability) is calculated using its chosen machine model.

[0266] As shown in Table 7, in terms of performance in predicting tumor origin, the overall accuracy rate of the tissue origin prediction model is 80.6%, meaning that in more than 80.6% of cases, the cancer prediction model can correctly predict the origin of cancer on the first attempt. Furthermore, its overall accuracy rate for the three most likely options is as high as 98.4%, indicating that even if the first-choice prediction is incorrect, the top three most likely options listed have a 98.4% chance of containing the correct answer.

[0267] As shown in Table 8, the primary tumor and metastatic lesion tissue samples of patients with liver cancer, colorectal cancer, breast cancer, gastric cancer, breast cancer liver metastasis, colorectal cancer liver metastasis, and gastric cancer liver metastasis were predicted. The top three most likely options in the prediction results included the patient's histopathological examination results.

[0268] 5.2 Second External Verification

[0269] The second external validation set was used to evaluate the accuracy of the tissue origin prediction model of this invention in diagnosing the primary site of metastatic cancer. Transcriptome sequencing data from 500 tumor patients and 22 tissue / organ metastases were selected from the second external validation dataset (65 cases of prostate cancer, 12 cases of head and neck cancer, 6 cases of adrenal cancer, 5 cases of brain tumors, 4 cases of bladder cancer, 2 cases of pancreatic cancer, 2 cases of melanoma, and 1 case of thyroid cancer) for prediction. The results are shown in Tables 9 and 10.

[0270] Table 9. Accuracy Rate of Tumor Source Tracing Diagnosis

[0271] Table 10 Diagnostic or predictive results for some patients with multiple cancers Note: NA means below the threshold.

[0272] As shown in Table 9, among the eight types of cancer metastatic tissues with known primary tumor origins, the overall accuracy rate of the tissue origin prediction model was 81.3%, meaning that in over 81.3% of cases, the cancer prediction model was able to correctly predict the origin of the cancer on the first attempt. Furthermore, its overall accuracy rate was as high as 96.9%, indicating that even if the first-choice prediction failed (or the most probable result was not correct), the top three most likely options listed had a 96.9% chance of containing the correct answer.

[0273] As shown in Table 10, the top three most likely prediction results for patients with metastatic prostate cancer, head and neck cancer, adrenal cancer, glioma, melanoma, and thyroid cancer were all included in the patient's histopathological examination results.

[0274] The present invention further collects more primary and metastatic cancer tissue samples, including more cancer types (covering 18 types) and more samples, and performs RNA sequencing to test the accuracy of the model.

[0275] The above tests demonstrate the potential of this invention as a valuable adjunct tool in clinical practice, which may guide optimal treatment strategies for CUP patients and further prolong overall survival. It is worthwhile to conduct in-depth research in prospective randomized trials in the future.

[0276] discuss

[0277] Tumor pathology diagnosis is currently one of the gold standards for diagnosing many types of tumors, boasting high accuracy. Through tumor pathology diagnosis, the nature and origin of the tumor can be clearly identified, and it can be staged, classified, and its differentiation level evaluated. By observing and analyzing the histological characteristics, grade, and stage of the tumor, patient survival rates and recurrence risks can be predicted, providing important references for treatment decisions and patient management, thereby guiding individualized treatment strategies. However, different types of tumors may exhibit similar histological features. Sometimes, patients may have multiple primary tumors or multiple independent malignant lesions simultaneously. Furthermore, the origin of certain metastatic tumors may not be determined through routine histological examination, such as cancer of unknown primary (CUP). In some cases, pathology has limitations in diagnosing primary and / or metastatic tumors, necessitating the use of techniques such as molecular genetics to assist in diagnosis.

[0278] CUP refers to metastatic malignant tumors that are histologically confirmed but whose primary site cannot be detected after a comprehensive and detailed examination. CUP is characterized by early metastasis, rapid progression, multiple organ involvement in more than half of patients, unknown metastatic patterns, and high mortality. Clinical treatment for CUP patients is mainly conventional chemotherapy, but empirical chemotherapy is ineffective, lacks specificity, and fails to improve patient survival. Identifying the primary tumor (including tumor type and tissue origin) is crucial for selecting CUP patients who may benefit from targeted therapy. Pathological diagnosis of CUP is a complex task; for example, the location of the primary tumor is difficult to determine, there may be multiple potential primary sites, including different organs and tissues, and the clinical presentation is nonspecific. Faced with the diagnostic challenges of CUP, physicians typically use a combination of diagnostic methods, including detailed medical history, physical examination, imaging examinations, pathological analysis, immunohistochemical staining, and molecular biological testing. Simultaneously, collaboration with a multidisciplinary team, including pathologists, radiologists, internists, and oncologists, is essential to improve the diagnostic accuracy and treatment outcomes of tumors with unknown primary sites. Furthermore, tumor pathological diagnosis is inherently subjective, requiring a high level of professional knowledge and experience from physicians. Different pathologists may interpret and judge the same tissue specimen differently. This can lead to inconsistencies in diagnoses between different hospitals and specialists, thereby affecting the comparability of treatment decisions and outcomes. Therefore, those skilled in the art recognized the need to obtain more tissue-specific biomarkers or biological characteristics at the molecular biological level to improve diagnostic accuracy.

[0279] In the human genome, over 95% of genes have more than one transcript, and the same gene often expresses different transcripts in different tissues or disease states. These different transcripts differ in structure and function. This transcript diversity is particularly evident in different tissues, cell types, and disease states, especially in human cancer research, where transcript expression patterns exhibit significant tissue and pathological state preferences. Specific RNA transcripts (SRTs) are transcripts of genes specifically expressed in a particular tissue or disease.

[0280] With the continuous development of high-throughput RNA sequencing technology and significant improvements in computational analysis methods, researchers are now able to conduct more in-depth analyses of transcriptional regulatory mechanisms under normal physiological and pathological conditions, and are expected to identify new transcripts. Although short-read RNA-seq technology has been widely used in transcriptomics research, it has inherent limitations in reconstructing transcripts, especially when multiple transcriptomic isoforms of the same gene share exons, making accurate differentiation and identification difficult. Considering the high complexity of the transcriptome and the aforementioned challenges, the task of comprehensively reconstructing all transcripts relying solely on short-read data is particularly arduous. In contrast, long-read sequencing (LR-Seq), also known as third-generation sequencing technology, including Pacific Biosciences (PacBio) and Oxford Nanopore Technologies (ONT), has advantages due to its ability to generate longer reads (>10kb), accurately capturing full-length isoforms. Third-generation sequencing technology promotes comprehensive and accurate characterization of transcript structure by eliminating the need for transcript assembly. Third-generation sequencing technology, thanks to its high accuracy in capturing complete transcriptomic isoforms, provides a comprehensive molecular basis for studying the transcriptomic characterization of tissues and cancers. Currently, numerous studies, including our own, have also used long-read RNA sequencing of tissues and samples to discover unannotated tissue-specific transcripts.

[0281] The expression of different transcripts of the same gene in different tissues is closely related to the specific functions of those tissues. Tissue-specific RNA transcripts are defined as transcripts of a gene that are expressed only in a specific tissue and not expressed or expressed at very low levels in other tissues. Specific RNA transcripts expressed in normal tissues are called tissue SRTs (ti-SRTs), also known as tissue-specific RNA transcripts.

[0282] Therefore, this invention analyzes tumor and normal tissue samples using long-read and short-read RNA sequencing technologies, and integrates third-generation RNA sequencing data from various normal and cancerous human tissues and cells, identifying a large number of previously unannotated tissue-specific RNA transcripts in a large sample set. These results lay a solid scientific foundation for further research and implementation of precision RNA diagnostic and therapeutic strategies.

[0283] All documents mentioned in this invention are incorporated herein by reference as if each document were individually incorporated by reference. Furthermore, it should be understood that after reading the foregoing teachings of this invention, those skilled in the art can make various alterations or modifications to this invention, and these equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A method for constructing an organizational origin prediction model, characterized in that, Including the following steps: (s1) Process the third-generation sequencing data and RNA sequencing (RNA-seq) data of cancer tissue samples, adjacent normal tissue samples and normal tissue samples from different tissue sources to obtain reference GTF files of full-length RNA transcripts and tissue-specific RNA transcripts (Ti-SRT) of different tissues. (s2) Provide training set, test set and validation set data for model construction, the data including the expression information of full-length RNA transcripts obtained from RNA-seq data of tissue samples from different tissue sources after RNA sequencing, based on the reference GTF file; the expression information includes the expression information of tissue-specific RNA transcripts from the different tissues; Specifically, for a certain type of tissue, the expression information of tissue-specific RNA transcripts of that tissue is the expression information of tissue-specific RNA transcripts of tissue-positive samples, while the expression information of tissue-specific RNA transcripts of other tissues for that tissue is the expression information of tissue-specific RNA transcripts of tissue-negative samples. (s3) Construction of tissue origin sub-models: Different sub-models are constructed for different tissues to determine whether the sample to be tested originates from that tissue; including the following steps: For any tissue, the expression information of tissue-specific RNA transcripts of tissue-positive samples and tissue-negative samples of that tissue are input into different machine learning models for training; thereby obtaining tissue origin prediction sub-models for determining whether the sample to be tested originates from that tissue; and using this step, tissue origin prediction sub-models for all tissues are obtained; The machine learning models include: Random Forest, Decision Tree, XGBoost, LightGBM, and CatBoost; and (s4) Construction of the organization source prediction model: Based on the organization source prediction sub-models of all organizations obtained in (s3), the test set and validation set data are used for optimization and selection to construct the organization source prediction model.

2. The method as described in claim 1, characterized in that, The transcript expression information includes the transcript expression level as measured in TPM (Transcript per million).

3. The method as described in claim 1, characterized in that, Step (s1) includes the following steps: (a) Collect third-generation sequencing data and RNA sequencing data of cancer tissue samples, adjacent normal tissue samples and normal tissue samples from different tissue sources, and generate non-redundant transcriptome GTF files; (b) Optimize non-redundant transcriptome GTF files, wherein, for a certain tissue RNA-seq, if all splicing sites of a transcript are detected in ≥m samples, the transcript is retained, where m is 1-90%, preferably 2-50%; (c) Each transcript was compared with the reference transcriptome (GENCODE v.35). The transcripts were mainly divided into four groups: fully spliced, partially spliced, novel within the catalog, and novel outside the catalog. After filtering, a reference GTF file for full-length RNA transcripts was established. Specifically, for all incompletely spliced ​​matching transcripts, these transcripts are filtered out from the entire transcriptome; Unqualified novel transcripts, both in-directory and out-of-directory, are filtered out (e.g., those with Cage peaks and polyA sites annotated with SQANTI3 are filtered out if there are 16 or more adenines within 20 bases downstream of the annotated transcription termination site). For similar transcripts, the longest transcript is retained, and other transcripts are filtered out; (d) Based on the reference GTF file for full-length RNA transcripts, full-length transcript quantification was performed on RNA-seq data from tissue samples to obtain expression information of full-length RNA transcripts; and (e) By comparison, RNA transcripts that show significant differences in expression information between each tissue sample and other tissue samples are obtained, thereby obtaining tissue-specific RNA transcripts for different tissues.

4. The method as described in claim 1, characterized in that, Step (s3) includes the following sub-steps: (s3a) Construct different tissue origin prediction sub-models for different tissues to determine whether the sample to be tested originates from that tissue; for any tissue, input the expression information of tissue-specific RNA transcripts of tissue-positive samples and tissue-negative samples of that tissue into the following five machine learning models for training: Random Forest, Decision Tree, XGBoost, LightGBM, and CatBoost; thereby obtaining the hyperparameters of the five machine learning models corresponding to the tissue; and based on this, obtain five trained machine learning models (or the first tissue origin prediction sub-model) for determining the origin of any tissue. (s3b) For the five first organization source prediction sub-models for determining any kind of organization source, the first organization source prediction sub-models are scored using test set and validation set data, and the following scoring indicators are used: AUC value, accuracy, sensitivity and specificity. For any organization, the highest-scoring first organization source prediction sub-model is used as the final organization source prediction sub-model (or second organization source prediction sub-model) for determining that organization; and (s3c) Obtain the organization source prediction sub-models for all organizations, i.e., all the second organization source prediction sub-models used for organization source prediction; 5. The method as described in claim 1, characterized in that, Step (s4) includes the following steps: (s4a) Based on the organization source prediction sub-models of all organizations obtained in (s3), use the validation set data and the roc auc function to score the organization source sub-models of all organizations; (s4b) Select all sub-models for predicting the origin of an organization with AUC ≥ 0.9 and construct the organization origin prediction model.

6. The method as described in claim 1, characterized in that, The tissue samples include tissue samples from the following tissues: adrenal glands, B cells, bladder, brain, breast, bone and / or muscle, bone marrow, cervix, intestine, esophagus, eye, nerve ganglia, kidneys, liver, lungs, head and neck, ovaries, pancreas, prostate, skin, stomach, testes, thymus, thyroid gland, and uterus.

7. A reference GTF file for full-length RNA transcripts, characterized in that, It is obtained by the following steps: (i) Collect third-generation sequencing data and RNA-seq sequencing data of cancer tissue samples, adjacent normal tissue samples and normal tissue samples from different tissue sources, and generate non-redundant transcriptome GTF files; (ii) Optimize non-redundant transcriptome GTF files, wherein, for a certain tissue RNA-seq, if all splicing sites of a transcript are detected in ≥m samples, the transcript is retained, where m is 1-90%, preferably 2-50%; (iii) Each transcript was compared with the reference transcriptome (GENCODE v.35). The transcripts were mainly divided into four groups: fully spliced, partially spliced, novel in-directory, and novel out-of-directory. After filtering, a reference GTF file for full-length RNA transcripts was established. Specifically, for all incompletely spliced ​​matching transcripts, these transcripts are filtered out from the entire transcriptome; Unqualified novel transcripts, both in-directory and out-of-directory, are filtered out (e.g., those with Cage peaks and polyA sites annotated with SQANTI3 are filtered out if there are 16 or more adenines within 20 bases downstream of the annotated transcription termination site). For similar transcripts, the longest transcript is retained, while other transcripts are filtered out.

8. A system for predicting the origin of an organization, characterized in that, The system includes: (a) An input module configured to input full-length RNA transcript expression information of the sample to be tested; (b) Evaluation module: The evaluation module is configured to input the expression information of the full-length RNA transcript into the tissue origin prediction model constructed using the method of claim 1, thereby obtaining the tissue origin prediction result of the sample to be tested; and (d) Output module, which is configured to output the prediction result.

9. The prediction system as described in claim 8, characterized in that, The evaluation in the evaluation module includes: (i) Input the full-length RNA transcript expression information into the tissue origin prediction model, and then input it into the sub-models of each tissue origin prediction model in sequence to obtain the prediction values ​​of all sub-models; wherein, each of the sub-models is used to independently determine whether the sample to be tested originates from a specific tissue; (ii) Determine whether the predicted values ​​of all sub-models are greater than or equal to a preset reference threshold C1 for their corresponding tissue origin prediction; the determination includes: if the predicted value of a sub-model is greater than or equal to the preset reference threshold C1 for its corresponding tissue origin prediction, then determine that the tissue is a candidate tissue of the sample to be tested; thereby obtaining all candidate tissues; and (iii) Sort all candidate tissues from largest to smallest according to their corresponding predicted values ​​to obtain the ranking results of candidate tissues, and obtain the candidate tissue with the largest predicted value as the tissue source of the test sample; that is, the tissue of the test sample comes from the candidate tissue with the largest predicted value.

10. A computer storage medium, characterized in that, The storage medium is used to store the computer program corresponding to the algorithm of the tissue origin prediction model constructed by the method described in the first aspect of the present invention.

Images