A multi-class and multi-functional intelligent classification method for whole genome expression profile data

Abstract

The embodiment of the present invention discloses a multifunctional intelligent classification method applied to multi-classification inference of whole genome expression spectrum data. The present invention belongs to the intersection of medical detection and intelligent algorithm. The core technology covers the construction of "special qualitative gene element" set Ω, identification function f v (u|x) design and scoring function S(u|x) design. The present invention discloses how to use bulk RNA-seq to construct Ω, and then demonstrates how to calculate f based on tissue or single-cell RNA-seq (scRNA-seq). v (u|x) and S(u|x), and the category of the tissue sample or single cell sample is inferred based on this. The embodiments of the present invention clearly illustrate how to implement the present invention for multi-classification inference of whole genome expression profile data, and its potential medical detection applications. For example, identifying cancer cells in the tumor microenvironment, synchronously identifying whether the captured circulating tumor cells (clusters) are tumor cells (clusters) and their tumor sources, distinguishing benign and malignant tumors, and identifying primary tumor lesions.

Description

Technical Field

[0001] The present invention relates to the intersection of genomic medical detection and intelligent algorithm technology. Specifically, it relates to a multifunctional intelligent classification method applied to multi-classification inference of whole genome expression spectrum data. Compared with the current machine learning classification algorithm, it more specifically relates to a method that realizes the category inference of a single sample in a simpler way, which can be widely used in multi-classification medical detection of tumors. Background Art

[0002] Using single-cell transcriptome sequencing (scRNA-seq) to study the tumor microenvironment helps tumor diagnosis, drug development, precision medicine, and personalized medication guidelines. In order to explore the complexity of the tumor microenvironment, a technology that can effectively distinguish tumor cells from non-tumor cells is needed. Using circulating tumor cells to study the discovery, progression, and metastasis of tumors also helps precision medicine and personalized treatment.

[0003] The biggest problem with the current technology for capturing and identifying circulating tumor cells is that standardized testing cannot be achieved, and repeated testing of the same sample is easily biased due to factors such as reagent composition, operating procedures, and interpretation processes. Circulating tumor cells are very rare in peripheral blood, and they vary in size, morphology, molecular markers, gene expression, gene mutations, and other aspects. Therefore, there is still a lack of mature methods to completely remove other cells in peripheral blood and to completely enrich circulating tumor cells.

[0004] In addition, the current clinical application scenarios of circulating tumor cells are usually the primary site of a known tumor. However, circulating tumor cells can enter the vascular system before forming visible solid tumors. But even if real circulating tumor cells are detected, there is currently a lack of technology or markers that can be widely applied to a variety of cancer types to reveal the organ and tissue origin of circulating tumor cells. This also brings confusion to the fact that when clinical symptoms are unclear or other detection methods show no abnormalities, a positive circulating tumor cell test cannot be specifically pointed out.

[0005] In a nutshell, circulating tumor cells extracted and detected by different methods actually require a technology to reconfirm their true identity and the tumor source they point to.

[0006] On the other hand, recent studies have attempted to use circulating tumor cells to find the primary site of unknown primary cancer. Unknown primary cancer is a rare disease in which tumor cells spread to other parts of the body, but conventional tests cannot determine their origin. However, as mentioned above, in fact, if the primary site of circulating tumor cells cannot be determined or is not given in advance, the origin of circulating tumor cells itself is still a problem. Therefore, identifying the primary site of unknown primary cancer has not been effectively solved.

[0007] Although machine learning and artificial intelligence algorithms have been widely used in medical tests based on data such as images, signals, and semantic text, it is still difficult to break through the inherent limitations of the aforementioned biomedical and clinical conditions, and to effectively and more simply and conveniently solve the multi-classification problem of tumor medical testing by using genomic expression profile data and single sample category inference; and the established classification model is also difficult to be applied to both tissue and cell levels at the same time.

[0008] In general, if there is a technology that can cover and solve all the various problems mentioned above, this technology will bring huge potential value and applications. Summary of the invention

[0009] The present invention aims to provide a novel one-stop multifunctional intelligent classification method, breaking through the bottleneck of existing technologies and the limitations of clinical congenital conditions, using whole genome expression profile data to perform single sample category inference with simpler, more convenient and effective technology, in which the established model can be applied to inference at both tissue and cell levels. The implementation of the present invention will be widely used in multi-classification medical detection of tumors, such as: synchronously determining whether a single cell is a tumor cell and its tumor source, early diagnosis of tumors, tumor typing, identification of primary tumor lesions, separation of cancer cells in the tumor microenvironment, etc. The concepts and practices provided by the present invention are described in detail as follows.

[0010] (1) Database access procedure. Extract K (≥2) stored categories Each type of c u Gene expression profiles of samples, creating a construction set Where n = ∑ u n u , n u It is c u The number of samples in , and for any sample d i ∈D,y i ∈C is the class label, x i =[x i1 ...x ip ] is an expression vector containing p features, whose features are determined by the gene set defined.

[0011] (2) Operational analysis constructor. For each type of c u Construct this type of "special qualitative gene element" ω = (h, h'), element ω consists of two different genes h, Composition, in any two different categories c u , c v ∈C, the parsing constructor creates the element ω according to two conditions, where condition 1: at least 90% of the labels are y i =c u Sample di Medium ih >x ih′ Established, condition 2: at least 90% of the labels are y i =c v Sample d i Medium ih <x ih′ Established, according to this, we can get c u , c v All m uv A set of ω elements,

[0012] (3) Operate the concatenation connector. For each class c established by the analysis constructor u All elements ω of K classes are merged into a confluence connector to reorganize all elements ω of K classes to construct a set of "special qualitative gene elements" for K classes. and stored in the internal storage device.

[0013] (4) Read a sample of the category to be inferred Where x is the sample The gene expression profile of is the category to be inferred.

[0014] (5) Read the set of “special qualitative gene elements” Ω in the internal storage device.

[0015] (6) Calculate the recognition function f v (u|x) program. This indicates that the v predict c u The likelihood of is calculated as

[0016]

[0017] In equation 1, Describe the sample Gene sets for expression profiling, Represents the ineffective element ω∈Ω uv The number of components whose genes (h or h′) are not in is defined in uv (x) is a meaningless element ω∈Ω uv The number of x ih =x ih′ .

[0018] (7) Calculate the score function S(u|x). c u The likelihood of is calculated as

[0019]

[0020] In Equation 1 and Equation 2, [Q] represents an Iverson bracket stating Q.

[0021] (8) Operation of decision operator. Based on the result obtained by the scoring function calculation program, the decision operator finally To infer the sample Category

[0022] Beneficial Effects

[0023] (1) The present invention provides a method for solving the problem that it is difficult to determine whether a single cell in a tumor microenvironment is a tumor cell or a non-tumor cell.

[0024] (2) The present invention provides a method for solving two major bottlenecks in the current biomedical engineering technology for capturing circulating tumor cells. That is, the circulating tumor cells captured and initially identified may not be authentic circulating tumor cells, and the tissue source of the captured circulating tumor cells cannot be determined.

[0025] (3) Estimation of the primary site of metastatic cancer with unknown primary site.

[0026] (4) Early cancer detection and estimation of cancer type.

[0027] (5) Estimation of tumor type or subtype. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A conceptual diagram of the multi-category multi-functional intelligent classification method applied to whole genome expression profile data and its components of the present invention.

[0029] Figure 2 This is an overall evaluation of the inference performance of a specific embodiment of the present invention. (A) Single cell application set: the inference performance of whether a single cell and circulating tumor cells are tumor cells. (B) Tumor tissue application set: the inference performance of the origin of tumor tissue. (C) Single cell application set: the inference performance of the tumor origin of single cells and circulating tumor cells.

[0030] Figure 3 Confusion matrix evaluation of the inference performance of specific embodiments of the present invention. (A) Model construction set: confusion matrix for tumor tissue origin inference. (B) Tumor tissue application set: confusion matrix for tumor tissue origin inference. (C) Single cell application set: confusion matrix for single cell and circulating tumor cell tumor origin inference. (D) Single cell application set: confusion matrix for false positive circulating tumor cells captured from prostate cancer patients.

[0031] Figure 4The area under the curve for inferring tumor origin of single cells and circulating tumor cells in a specific embodiment of the present invention is evaluated. (A) Area under the receiver operating characteristic curve (AUROC). (B) Area under the precision-recall curve (AUPRC).

[0032] Figure 1 (Designated abstract figure) Symbol description: 10 represents program_access database, 101 represents stored data_gene expression profile class c1, 102 represents stored data_gene expression profile class c2, 103 represents stored data_gene expression profile class c u , 104 represents the stored data_gene expression profile class c K , 20 represents the operator_analysis constructor, 201 represents the transitional data_qualitative gene element class c1, 202 represents the transitional data_qualitative gene element class c2, 203 represents the transitional data_qualitative gene element class c u , 204 represents transitional data_qualitative gene element class c K , 30 represents the operation operator_joiner, 40 represents the internal storage device and data_special qualitative gene element set Ω, 50 represents the read data_to be inferred sample gene expression profile 60 represents the program_calculate the recognition function f v (u|x), 70 represents the program_calculation score function S(u|x), 80 represents the operation operator_decision operator, 90 represents the inference terminal point_ DETAILED DESCRIPTION

[0033] The following examples will publicly demonstrate how to implement the multi-class multi-functional intelligent classification method of the present invention to perform multi-classification inference of tumor genome expression profile data. Through the display of this example and related drawings, the purpose, technical solution, advantages, effectiveness and feasibility of the present invention will be more clearly understood. Figure 1 (Designated Abstract Attachment) as shown.

[0034] Step 1: Access the bulk RNA-seq gene expression quantitative data of tumor tissue and leukocytes of the cancer type to be identified (10). Extract the stored K (=8) classes Each type of c u The sample gene expression profile (101, 102, ..., 103, ..., 104) is created to construct a set Where n = ∑ u n u , n u It is c u The number of samples in , and for any sample d i ∈D,y i∈C is the class label, x i =[x i1 ...x ip ] is an expression vector containing p features, whose features are determined by the gene set The embodiment of the present invention only takes 7 types of cancer as examples, including ovarian cancer, lung cancer, liver cancer, colorectal cancer, breast cancer, prostate cancer and melanoma. For each type of cancer tumor tissue, only the in situ tissue is used, and the tissue of the tumor metastasis is discarded. For white blood cells, white blood cells of healthy people and non-tumor patients are used. Table 1 shows the created construction set D.

[0035] Step 2: Create tumor tissue bulk RNA-seq and single cell (including circulating tumor cells) scRNA-seq application sets to simulate the use of tumor tissue samples and single cell sample inference (50). Table 2 is the tumor tissue application set. Table 3 is the single cell application set.

[0036] Step 3: Zero-value filling preprocessing of scRNA-seq data. The embodiment of the present invention uses but is not limited to the SCRABBLE algorithm to fill zero values ​​of scRNA-seq data. The data preprocessing of this step can be considered whether to perform data preprocessing according to the standard process of omics data output adopted. A major advantage of the present invention is that it has a certain degree of antagonism to whether the data is preprocessed.

[0037] Step 4: Operate the parsing constructor (20). For each class c in D u Construct this type of "special qualitative gene element" ω = (h, h'), element ω consists of two different genes h, Composition, in any two different categories c u , c v ∈C, the parsing constructor (20) builds the component ω according to two conditions, where condition 1: at least 90% of the labels are y i =c u Sample d i Medium ih >x ih′ Established, condition 2: at least 90% of the labels are y i =c v Sample d i Medium ih <x ih′ Established, according to this, we can get c u , c v All m uv A set of ω elements, Each type of c u Qualitative genetic elements (201, 202, ..., 203, ..., 204).

[0038] Step 5: Operate the join connector (30). For each class c established by the analysis constructor (20), u All elements ω(201, 202, ..., 203, ..., 204) of K classes are merged into a reorganization by a confluence connector (30) to construct a set of "special qualitative gene elements" for K classes. And stored in the internal storage device (40). Table 4 shows the number of elements ω between any two categories in the whole genome integrated element set Ω.

[0039] Step 6: Read a sample of the category to be inferred from the construction set created in step 1 and the application set established in step 2 one by one Where x is the sample The gene expression profile of is the category to be inferred.

[0040] Step 7: Read the set of “special qualitative gene elements” Ω (40) in the internal storage device.

[0041] Step 8: Calculate the recognition function f v (u|x) program (60). Its calculation method is

[0042]

[0043] In equation 1, Describe the sample Gene sets for expression profiling, Represents the ineffective element ω∈Ω uv The number of components whose genes (h or h′) are not in is defined in uv (x) is a meaningless element ω∈Ω uv The number of x ih =x ih′ .

[0044] Step 9: Calculate the score function S(u|x) program (70). The calculation method is:

[0045]

[0046] In Equation 1 and Equation 2, [Q] represents an Iverson bracket stating Q.

[0047] Step 10: Operate the decision operator (80). Based on the result obtained by the score function calculation program (70), the decision operator (80) finally determines the score according to To infer the sample Category

[0048] Step 11: Overall evaluation of the inference performance of the embodiment of the present invention. For the samples in the model building set of step 1, the tumor tissue application set of step 2, and the single cell application set Inferred category The following 9 evaluation indicators are used to evaluate the feasibility of the practical application of the present invention. 9 evaluation indicators: accuracy, sensitivity, specificity, F-score, precision, recall, area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), confusion matrix. The results are shown in Figures 2 to 4 .

[0049] This embodiment further defines the method of "tumor tissue origin" and "simultaneously determining whether a single cell is a tumor cell and its tumor origin" in the present invention for multi-classification medical detection of tumors, wherein:

[0050] The biological sample in this embodiment can be arbitrarily selected according to actual conditions, and can be a single cell, a cell cluster, or a tissue.

[0051] When it is uncertain whether cells in the tumor microenvironment are tumor cells, the model established in this embodiment is suitable for determining whether cells in the tumor microenvironment are tumor cells.

[0052] When it is uncertain whether the captured circulating tumor cells are tumor cells, the model established in this embodiment is suitable for re-identification of whether the captured circulating tumor cells are tumor cells.

[0053] When the source of circulating tumor cells is unknown, the model established in this embodiment is suitable for identifying the tumor source of circulating tumor cells.

[0054] When the primary focus of the circulating tumor cell cluster is unknown, the model established in this embodiment is suitable for identifying the primary focus of the circulating tumor cell cluster tumor.

[0055] When it is uncertain whether the captured circulating tumor cell clusters are tumor cell clusters, the model established in this embodiment is suitable for re-identification of whether the captured circulating tumor cell clusters are tumor cells.

[0056] When the primary tumor site of the patient is unknown, the model established in this embodiment is suitable for identifying the primary tumor site.

[0057] When it is unknown whether the patient has a tumor, the model established in this embodiment is suitable for distinguishing benign or malignant tumors.

[0058] Schedule:

[0059] Table 1. Model building set (bulk RNA-seq)

[0060]

[0061] Table 2. Tumor tissue application set (bulk RNA-seq)

[0062]

[0063] Table 3. Single cell application set (scRNA-seq)

[0064]

[0065] *Single tumor cells derived from tissue.

[0066] +Circulating tumor cells.

[0067] - False positive circulating tumor cells.

[0068] Table 4. Number of "special qualitative gene elements" between any two categories of the integrated element set Ω

[0069]

Claims

1. A multi-category and multi-functional intelligent classification method, characterized in that: Create a construction set D, use D to construct the "special qualitative gene element" set Ω, and then use the identification function f v (u|x) design and score function S(u|x) design to infer samples with unknown category labels The specific method is described as follows: (1) Access database program to extract K (≥2) stored categories Each type of c u Gene expression profiles of samples, creating a construction set Where n = ∑ u n u , n u It is c u The number of samples in , and for any sample d i ∈D,y i ∈C is the class label, x i =[x i1 …x ip ] is an expression vector containing p features, Gene set defined; (2) Operate the analysis constructor to analyze each type of c from D u Construct this type of "special qualitative gene element" ω=(h,h′), element ω consists of two different genes h, Composition, in any two different categories c u , c v ∈C, the parsing constructor creates the element ω according to two conditions, where condition 1: at least 90% of the labels are y i =c u Sample d i Medium ih >x ih′ Established, condition 2: at least 90% of the labels are y i =c v Sample d i Medium ih <x ih′ Established, according to this, we can get c u , c v All m uv A set of ω elements, (3) Operate the join connector to analyze each class of c established by the constructor u All elements ω of K classes are merged into a confluence connector to reorganize all elements ω of K classes to construct a set of "special qualitative gene elements" for K classes. and stored in an internal storage device; (4) Read a sample of the category to be inferred Where x is the sample The gene expression profile of is the category to be inferred; (5) Reading the set of "special qualitative gene elements" Ω in the internal storage device; (6) Calculate the recognition function f v (u|x) program, indicating that the category c v predict c u The likelihood of is calculated as In equation 1, Describe the sample Gene sets for expression profiling, Represents the ineffective element ω∈Ω uv The number of components whose genes (h or h′) are not in is defined in uv (x) is a meaningless element ω∈Ω uv The number of x ih =x ih′ ; (7) The program for calculating the score function S(u|x) shows that c u The likelihood of is calculated as In Equations 1 and 2, [Q] represents the Iverson brackets for statement Q; (8) Operate the decision operator. Based on the result obtained by the scoring function calculation program, the decision operator finally To infer the sample Category 2. The multi-category multi-function intelligent classification method according to claim 1, characterized in that: The method for establishing the construction set D in (2) uses bulk RNA sequencing (bulk RNA-seq) or microarray (microarray) expression profile data of tumor tissues and leukocytes of multiple cancer types.

3. The multi-category multi-function intelligent classification method according to claim 1, characterized in that: In the method for establishing the sample set constructed in (2), only the tissue of the primary tumor focus is obtained, and the tissue of the tumor metastasis focus is discarded; the white blood cells are selected from healthy people and non-tumor patients.

4. The multi-category multi-function intelligent classification method according to claim 1, characterized in that: The scope and method of "special qualitative gene elements" used for inference of samples with unknown category labels do not use conventional feature-based screening algorithms to pre-screen a fixed number of genome sets for sample inference, but use a whole genome integrated element set Ω; provide recognition functions to dynamically screen "special qualitative gene elements" automatically and intelligently according to the quality of the whole genome expression spectrum of each sample as the basis for inference.

5. The multi-class multi-functional intelligent classification method according to any one of claims 1 to 4, applied to multi-classification inference of single-cell / cluster RNA sequencing (scRNA-seq) expression profile data without zero-filling, characterized in that: The samples whose categories are to be inferred are from single-cell / cluster samples, and their gene expression profiles have not been preprocessed.

6. The multi-class multi-functional intelligent classification method according to any one of claims 1 to 4, applied to multi-classification inference of single-cell / cluster RNA sequencing (scRNA-seq) expression profile data pre-processed with zero-value filling, characterized in that: The samples whose categories are to be inferred are from single cell / cluster samples, and their gene expression profiles are preprocessed by zero-filling.

7. The multi-classification and multi-functional intelligent classification method according to any one of claims 1 to 4 is applied to multi-classification inference of whole genome expression profile data bulk RNA-seq or microarray at the tissue level, characterized in that: The source of the sample to be inferred is a tissue sample.

8. The multi-category multi-function intelligent classification method according to claim 1, characterized in that: The decision operator is applied to the inference of the binary classification. The decision operator is finally based on or Choose one or the other to infer the sample Category 9. The multi-category multi-function intelligent classification method according to claim 1, characterized in that ； Applied to multi-classification, including binary classification, inference detection problems, the gene expression spectrum data is replaced with other omics data, including DNA mutation spectrum, DNA methylation spectrum, protein expression spectrum, molecular mass spectrum, to create the construction set D.

Images