Cancer early screening model construction method and device based on cfDNA difference statistics

Abstract

The invention discloses a method and a device for constructing a cancer early screening model based on cfDNA difference statistics. The construction method of the cancer early screening model comprises the following steps: performing difference statistics on tumor-derived cfDNA and healthy individual-derived cfDNA in a low-depth whole-genome sequencing mode, and establishing the cancer early screening model to realize non-invasive early screening of cancers; the difference statistics comprises cfDNA fragment size characteristic difference statistics, cfDNA upstream and downstream continuous three groups of 3bp base combination category distribution frequency change trend difference statistics at a fracture position and sequencing data coverage depth mode difference statistics of the cfDNA near TSS. The method has the advantages of being minimally invasive, high in early diagnosis sensitivity, capable of achieving dynamic monitoring and the like, the limitation of cfDNA gene mutation detection in the early stage of the tumor is overcome, and information lag or deletion caused by tumor heterogeneity is avoided.

Description

Technical Field

[0001] The present invention belongs to the field of biotechnology and relates to a method and device for constructing a cancer early screening model based on cfDNA difference statistics. Background Art

[0002] Biliary tract cancer is the fifth most common gastrointestinal malignancy, mainly including gallbladder cancer and bile duct cancer. Pancreatic cancer is the most invasive and lethal digestive tract malignancy, and the gallbladder, bile duct and pancreas are anatomically close. Early lesions of bile and pancreatic malignancies are usually asymptomatic, and the disease only becomes obvious after the tumor invades surrounding tissues or metastasizes to distant organs. Usually, patients are already in the late stage of the tumor when diagnosed, and late-stage patients often have a poor prognosis due to untimely treatment intervention. The five-year survival rate of pancreatic cancer patients is only 5%, while the five-year survival rate of bile duct cancer patients is less than 5%. Even if some patients can be treated with surgical intervention, the tumor is likely to recur after resection. Therefore, finding reliable early diagnostic biomarkers is more clinically valuable for the diagnosis and treatment of bile and pancreatic malignancies.

[0003] At present, imaging, biopsy, cytology and serum tumor marker detection are the main methods for diagnosing gallbladder and pancreatic malignancies. Among them, imaging methods mainly include ultrasound, MRI (magnetic resonance imaging), CT (computed tomography), ERCP (endoscopic retrograde pancreaticobiliary angiography), etc. Ultrasound is the preferred screening method for gallbladder and pancreatic malignancies, and MRI and CT are the main methods for diagnosis and staging of gallbladder and pancreatic malignancies. MRI can better show the anatomy of the liver, bile duct and pancreas and the scope of tumor involvement, but the sensitivity is poor for tumors with a diameter of less than 1 cm. The brushing exfoliated cell examination of puncture biopsy and ERCP is the gold standard for the diagnosis of gallbladder and pancreatic malignancies, but the sensitivity of brushing is low, and puncture biopsy is prone to abdominal infection and tumor spread and metastasis. CA19-9 is a unique prognostic biomarker approved by the FDA for resectable gallbladder and pancreatic malignancies, but its sensitivity and specificity are limited when used for diagnosis, and it is easily interfered by factors such as chronic pancreatitis, cholestasis or cholangitis.

[0004] Liquid biopsy is widely used for early diagnosis, postoperative monitoring, and companion diagnosis of tumors based on its minimally invasive or non-invasive sampling methods, economic convenience, and other advantages. When an organ or tissue suffers from a pathological state that leads to increased cell death, more free DNA (cfDNA) will appear in the circulating blood. Therefore, plasma cfDNA carries genetic and epigenetic information of pathological cells and their tissues of origin. However, not all patients can detect genetic variation information because the concentration of tumor-related cfDNA in plasma is low, especially in the early stages of the tumor. Plasma cfDNA fragmentomics is an emerging field that covers the fragment size, end point, nucleosome footprint, and transcription start site of plasma cfDNA. Studies have shown that there are differences in the genome-wide cfDNA fragment omics profiles between cancer patients and non-cancer populations, and the cfDNA fragment size and terminal sequence characteristics can be used for early diagnosis of tumors. Cristiano et al.'s study showed that the sensitivity of the genome-wide cfDNA fragment size characteristics for diagnosing cholangiocarcinoma is 88% (PMID: 31142840). However, studies using cfDNA fragmentomics from low-coverage whole-genome sequencing (LP-WGS) data to establish a detection model for anatomically proximal biliopancreatic malignancies are still lacking.

[0005] Therefore, there is an urgent need for research data from low-depth cfDNA whole-genome sequencing to assist in the clinical transformation of biomarkers in early diagnosis, postoperative detection, and companion diagnosis of biliary and pancreatic malignancies. Summary of the invention

[0006] In view of the deficiencies in the prior art and actual needs, the present invention provides a method and device for constructing a cancer early screening model based on cfDNA difference statistics. By performing whole genome sequencing on the blood cfDNA of patients with bile duct cancer and pancreatic cancer, and taking healthy people as the control group, the omics characteristics of plasma cfDNA fragments are analyzed, and liquid biopsy biomarkers that can be used for the diagnosis of bile duct cancer are discovered and verified, so as to facilitate the clinical transformation of biomarkers for early diagnosis of bile duct cancer, prolong the survival time of patients with bile duct cancer, and improve the quality of life. With the advantages of minimally invasive, high sensitivity for early diagnosis and dynamic monitoring, the method overcomes the limitations of cfDNA gene mutation detection in the early stages of tumors and avoids information lag or loss caused by tumor heterogeneity.

[0007] In order to achieve the purpose of the invention, the present invention adopts the following technical solutions:

[0008] In the first aspect, the present invention provides a method for constructing a cancer early screening model based on cfDNA difference statistics, the construction method comprising: performing difference statistics on tumor-derived cfDNA and cfDNA from healthy individuals through a relatively low-depth whole-genome sequencing method, establishing an early screening model for cancer, and realizing non-invasive early screening for cancer; the difference statistics include difference statistics on cfDNA fragment size characteristics, difference statistics on distribution frequency change trends of three consecutive groups of 3bp base combination categories upstream and downstream of the break position of cfDNA, and difference statistics on sequencing data coverage depth patterns of cfDNA near TSS.

[0009] The method of the present invention overcomes the limitations of cfDNA gene mutation detection in the early stages of tumors with the advantages of minimal invasiveness, high sensitivity for early diagnosis and dynamic monitoring, and avoids information lag or loss caused by tumor heterogeneity.

[0010] Preferably, the cfDNA fragments include cfDNA short fragments and cfDNA long fragments, the length of the cfDNA short fragments is [130, 177] bp, the length of the cfDNA long fragments is [177, 237] bp, and the ratio of the cfDNA short fragments to the cfDNA long fragments is the characteristic difference value of the cfDNA fragment size;

[0011] The statistical method for the distribution frequency of three consecutive groups of 3bp base combination categories of cfDNA upstream and downstream of the break position includes: counting 9 bases from the sixth base at the 5' end of the cfDNA and grouping three consecutive bases into three groups of 3bp motifs; and counting the occurrence frequency of the three groups of 3bp motifs for each sample;

[0012] The parameter calculation method for the sequencing coverage depth of cfDNA near TSS is as follows: the area 500bp [-250bp, 250bp] upstream and downstream of TSS is defined as the central area, and the upstream [-2000bp, -1000bp] and downstream [1000bp, 2000bp] of TSS are defined as the surrounding area; the NF value of the gene is the average coverage of the central area divided by the average coverage of the surrounding area.

[0013] Preferably, the early cancer screening method further comprises: performing statistical analysis on the differences between tumor-derived cfDNA and cfDNA from healthy individuals, detecting the serum CA19-9 expression level, and then constructing a model.

[0014] Preferably, the cancer comprises biliopancreatic malignancies.

[0015] Preferably, the biliopancreatic malignancy includes any one of pancreatic cancer, gallbladder cancer or bile duct cancer.

[0016] In a second aspect, the present invention provides a non-invasive early screening device for cancer, the non-invasive early screening device for cancer comprising: a feature extraction module, a machine learning classification model building module and an independent verification cohort evaluation module;

[0017] The feature extraction module is used to perform cfDNA fragment feature extraction, terminal Motif feature extraction and TSS data feature extraction;

[0018] The independent validation cohort evaluation module is used to perform validation of the predictive performance of the established machine learning classification model through an independent validation cohort.

[0019] Preferably, the cfDNA fragment feature extraction includes: a sequencing data comparison unit, a cfDNA fragment statistics unit and a cfDNA fragment feature determination unit;

[0020] The sequencing data comparison unit is used to perform the following steps: after removing the sequencing data sequencing adapter, aligning the sequencing data to the human reference genome hg19;

[0021] The cfDNA fragment statistical unit is used to perform the following steps: statistically analyzing cfDNA fragment length data information; dividing the hg19 autosome into 504 adjacent window fragments without intersection, each window fragment being 5 Mb in length; statistically analyzing the ratio of the number of cfDNA with a length greater than 130 bp and less than 177 bp to the number of cfDNA with a length greater than 177 bp and less than 237 bp in each window region; and finally obtaining the number of cfDNA long and short fragments in each 5 Mb interval.

[0022] The cfDNA fragment feature determination unit is used to perform the following steps: determining the interval with the largest difference in fragment distribution between cancer patients and healthy controls according to the difference distribution of fragment distribution between cancer patients and healthy controls; defining a short fragment range [130, 177] and a long fragment range [177, 237], and then calculating the standardized z-score of the number of short fragment cfDNA and total fragments in each window of 504 windows as feature input values ​​for model training;

[0023] The terminal motif feature extraction includes: a sequencing data comparison unit, a reads filtering unit, a combined motif calculation unit and a combined motif statistics unit;

[0024] The sequencing data comparison unit is used to perform the following steps: after removing the sequencing data sequencing adapter, aligning the sequencing data to the human reference genome hg19;

[0025] The reads filtering unit is used to perform the following steps: filtering and screening the sequencing data, wherein the filtering criteria are: only considering reads aligned to autosomes 1-22, with a quality score greater than 20; the insert length is between 150-600, and both ends must be properly paired; the reference region of the read does not contain degenerate bases;

[0026] The combination motif calculation unit is used to perform the following steps: counting 9 bases from the sixth base at the 5' end and dividing them into three groups as the combination motif; for the 3' end read, counting 9 bases from the sixth base at the 3' end and dividing them into three groups in the order of 3' to 5', and then reverse complementing them as the combination motif;

[0027] The combined motif statistics unit is used to count the frequencies of the three groups of 3bp motifs for each sample, each group of 3bp motifs has 64 different combinations, and a total of 192 frequencies are used as its training variables;

[0028] The TSS data feature extraction includes a sequencing data comparison unit, a reads filtering unit, a gene screening and TSS determination unit and a NF value calculation unit;

[0029] The sequencing data comparison unit is used to perform the following steps: after removing the sequencing data sequencing adapter, aligning the sequencing data to the human reference genome hg19;

[0030] The reads filtering unit is used to perform the following steps: filtering and screening the sequencing data, wherein the filtering criteria are: only considering reads aligned to autosomes 1-22, with a quality score greater than 20, an insert length between 150-600, and both ends must be properly paired; and the reference region of the read does not contain degenerate bases;

[0031] The gene screening and TSS determination unit is used to perform the following steps: determining the TSS of gene transcription according to the transcript annotation of the UCSC hg19 genome; for genes with multiple TSSs, only those with a difference of less than 50 bp between different TSSs are retained, and their average value is used as the TSS of the gene, and only genes on autosomes are considered;

[0032] The NF value calculation unit is used to perform the following steps: obtaining parameters for the sequencing coverage depth of cfDNA near TSS, and the calculation method is: defining the region 500bp [-250bp, 250bp] upstream and downstream of TSS as the central region; defining 2000bp upstream [-2000bp, -1000bp] and downstream [1000bp, 2000bp] of TSS as the surrounding region; and the NF value of a gene is the average coverage of the central region divided by the average coverage of the surrounding region.

[0033] Preferably, the machine learning classification model building module includes: a model parameter acquisition unit and a model effectiveness evaluation unit;

[0034] The model parameter acquisition unit is used to perform the following steps: using a machine algorithm in a training set and using a 5-fold or 10-fold cross validation method to obtain model parameters;

[0035] The model effectiveness evaluation unit is used to perform the following steps: drawing a receiver operating characteristic curve of the training set according to the model prediction value and pathological detection result of each sample in the training set.

[0036] In a third aspect, the present invention provides a cancer early screening model, wherein the input of the cancer early screening model is the score value of the risk scoring model and Log2(1+CA19-9); and the output is the model score.

[0037] Preferably, the scoring value of the risk scoring model is the scoring value of the test set.

[0038] Preferably, when the model score is lower than 0, it is judged as healthy, and when it is higher than or equal to 0, it is judged as cancer.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) The present invention has developed a blood cfDNA biomarker liquid biopsy technology that can be used for the early diagnosis of gallbladder and pancreatic malignancies. The risk prediction model with high specificity and high sensitivity is constructed based on the cfDNA fragment size characteristics (independently defined size fragment range), 5' end Motif characteristics (independently developed 3x3bp combined Motif characteristics) and deep pattern characteristics of sequencing data near the transcription start site (TSS). The blood cfDNA fragment biomarkers and serum CA19-9 expression levels are combined to establish a more effective diagnostic model.

[0041] (2) The present invention overcomes the limitations of cfDNA gene mutation detection in the early stages of tumors with its advantages of minimal invasiveness, high sensitivity for early diagnosis, and dynamic monitoring, and avoids information lag or loss caused by tumor heterogeneity. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1A This is the result of blood cfDNA fragment size characteristic analysis;

[0043] Figure 1B This is the result of the analysis of the 5' end Motif characteristics of blood cfDNA;

[0044] Figure 1C This is a graph showing the results of deep pattern feature analysis of sequencing data near the transcription start site (TSS) of blood cfDNA;

[0045] Figure 2 This is the ROC curve diagram for the training cohort;

[0046] Figure 3 ROC curve diagram for the validation cohort;

[0047] Figure 4 This is the ROC curve of the blood cfDNA fragment omics combined with CA19-9 diagnostic model. DETAILED DESCRIPTION

[0048] To further illustrate the technical means and effects of the present invention, the present invention is further described below in conjunction with the embodiments and drawings. It should be understood that the specific implementation methods described herein are only used to explain the present invention, rather than to limit the present invention.

[0049] If no specific techniques or conditions are specified in the examples, the techniques or conditions described in the literature in the field or the product instructions are used. If no manufacturer is specified for the reagents or instruments used, they are all conventional products that can be purchased through regular channels.

[0050] Example 1

[0051] This embodiment provides a method and model for detecting malignant tumors of the gallbladder and pancreas.

[0052] (1) Study cohort

[0053] A total of 147 patients (55 cholangiocarcinoma, 30 gallbladder cancer, and 62 pancreatic cancer) suspected of having malignant tumors of the gallbladder and pancreas by tumor markers and imaging examinations and 71 healthy subjects were included in this study. Blood samples were collected from all patients before surgery. Each enrolled patient was given an accurate pathological diagnosis after surgery.

[0054] (2) Blood sample processing of patients with pancreatic and biliary malignancies and healthy subjects

[0055] Blood samples from patients with pancreatic and biliary malignancies and healthy controls were collected in 10 mL free nucleic acid storage tubes (BEAVERCell Free DNATubes, 43803, China) and shipped at room temperature. After receiving the samples, the blood was first centrifuged at 1,600 g for 10 min at 4 °C to remove cells and debris. After centrifugation, the supernatant was transferred to a new 1.5 mL centrifuge tube, and the supernatant was centrifuged again at 16,000 g for 15 min at 4 °C. Finally, the supernatant was transferred and divided into 1.5 mL centrifuge tubes at 1 mL per tube and stored in a -80 °C refrigerator.

[0056] (3) Plasma cfDNA extraction from patients with pancreatic and biliary malignancies and healthy subjects

[0057] After the plasma sample was taken out from -80℃, it was transferred to a 37℃ water bath to melt, and then centrifuged at 4℃, 16,000g for 10min, and 0.99mL of supernatant was aspirated into a new 1.5mL centrifuge tube. The QIAamp Circulating Nucleic Acid Kit (55114, Qiagen, Shanghai, China) was used for subsequent cfDNA extraction. After sample lysis, protein digestion, column, washing, elution and other steps, plasma cfDNA was collected and then placed in a -20℃ or -80℃ refrigerator for long-term storage.

[0058] (4) Detection of plasma cfDNA concentration and fragment size

[0059] The concentration of plasma cfDNA was detected using the Qubit 3.0 fluorescence quantifier and supporting reagents (Q32854, Thermo Fisher, USA). The distribution of extracted plasma cfDNA fragments was analyzed using the 2100 analyzer and supporting chips and reagents (5067-4626, Agilent, USA). The specific detection process is referred to the kit instructions.

[0060] (5) Plasma cfDNA micro-library construction and LP-WGS sequencing

[0061] The plasma cfDNA library was constructed using the KAPA DNA HyperPrep (KK8504, KAPA, USA) kit. The input amount of each cfDNA was 3-10 ng. The final library was obtained after end repair and A-tailing, adapter ligation and purification, library amplification and purification. The specific operation process is referred to the product manual. After the constructed library passed the quality control, the Illumina Nova Seq6000 was used for double-end sequencing, and the sequencing data volume of each sample was 10G.

[0062] (6) Extraction of plasma cfDNA fragment genomic features

[0063] 1) Extraction of plasma cfDNA fragment size characteristics

[0064] 1. Sequencing data alignment. After removing the adapter (fastq) from the original fastq data obtained by LP-WGS sequencing of plasma cfDNA from patients with pancreatic and biliary malignancies and healthy controls, the sequencing data were aligned to the human reference genome hg19 (genome download link: ftp: / / ftp-trace.ncbi.nih.gov / 1000genomes / ftp / technical / reference / human_g1k_v37.fas ta.gz) using BWA software (version: 0.7.12-r1039), and low-quality sequences were removed from the obtained BAM files and duplicate sequences were filtered out;

[0065] 2. Statistics of cfDNA fragment length;

[0066] 3. Whole-genome fragment size distribution map. The low-coverage regions of the hg19 reference genome and the Duke black box regions were excluded; the hg19 autosomes were then divided into 504 adjacent, non-overlapping window fragments, each with a length of 5 Mb; in each window region, the ratio of the number of cfDNAs with a length greater than 130 bp and less than 177 bp to the number of cfDNAs with a length greater than 177 bp and less than 237 bp was counted; finally, the number of long and short cfDNA fragments in each 5 Mb interval was obtained, and the ratio was used to visualize the cfDNA fragmentation size map of the entire genome.

[0067] 2) 5' end motif feature extraction

[0068] 1. Sequencing data alignment. After removing the sequencing adapters from the plasma cfDNA sequencing data of patients with pancreatic and biliary malignancies and healthy controls, the sequencing data were aligned to the human reference genome hg19 (genome download link: ftp: / / ftp-trace.ncbi.nih.gov / 1000genomes / ftp / technical / reference / human_g1k_v37.fas ta.gz) using BWA software (version: 0.7.17-r1188);

[0069] 2. Read filtering. Only reads that are aligned to autosomes 1-22 are considered; the quality score is greater than 20; the insertion size is between 150 and 600 bp; both ends must be properly paired; the reference region of the read does not contain degenerate bases;

[0070] 3. Combined motif calculation. For the read at the 5' end, count 9 bases from the 5' end and divide three consecutive bases into three groups, which is its combined motif. Since the sequencing data has UMI added, there is a UMI sequence in front of each read. The actual calculation is to count 9 bases from the 6th base (inclusive) at the 5' end and divide them into three groups as its combined motif. 4. Combined motif statistics. Each group of 3bp motifs has 64 different combinations. For each sample, the frequencies of its three groups of 3bp motifs are counted, and a total of 192 frequencies are used as its training variables.

[0071] 3) Deep pattern feature extraction of sequencing data near the transcription start site (TSS)

[0072] 1. Sequencing data alignment. After removing the sequencing adapters from the plasma cfDNA sequencing data of patients with pancreatic and biliary malignancies and healthy controls, the sequencing data were aligned to the human reference genome hg19 (genome download link: ftp: / / ftp-trace.ncbi.nih.gov / 1000genomes / ftp / technical / reference / human_g1k_v37.fas ta.gz(;

[0073] 2. Read filtering. Only reads aligned to autosomes 1-22 are considered; the quality score is greater than 20; the insertion size is between 150 and 600; both ends must be properly paired; the reference region of the read does not contain degenerate bases;

[0074] 3. Gene screening and TSS determination. The transcription start site (TSS) of the gene was determined based on the transcript annotation of the UCSC hg19 genome. For genes with multiple TSSs, only those with a difference of less than 50 bp were retained, and their average value was used as the TSS of the gene. At the same time, only genes on autosomes were considered;

[0075] 4. Nucleosome footprint (NF) value calculation. Define the region 500bp upstream and downstream of TSS [-250bp, 250bp] as the central region. Define the 2000bp upstream [-2000bp, -1000bp] and downstream [1000bp, 2000bp] of TSS as the surrounding region. The NF of a gene is calculated as the average coverage of the central region divided by the average coverage of the surrounding region.

[0076] (7) Establishment and evaluation of a risk scoring model for pancreatic and biliary tumors

[0077] 1) Model establishment: Using the training cohort and the pathological test results, the Stacking algorithm is used to construct an integrated scoring model that uses three types of features as input to distinguish cancer patients from healthy people. The model consists of three parts: variables, primary model formulas, secondary model formulas, and predicted values. The process is as follows:

[0078] 1. Primary model variables and parameters of three types of features:

[0079] a.Fragment prediction value:

[0080] a) Model variables and parameters: The LinearSVC model was used, with 504 short fragments within a 5 Mb region and 1008 standardized z-scores of the number of all fragments as feature variables (Table 1), and the model parameters were obtained using a 10-fold cross-validation method.

[0081] Table 1

[0082]

[0083]

[0084] b) Scoring model. The scoring model formula is as follows:

[0085] Score = w T x i -b

[0086] where x i As input variables, the calculation formulas of model coefficients w and b are as follows:

[0087]

[0088] Where λ is the penalty parameter, n is the number of samples, and y i is the true value of the sample, 1 for cancer and -1 for health.

[0089] Using the classification model and the fragment distribution of different regions at the whole genome level of each sample, the category prediction result (Score) of each sample can be obtained.

[0090] b. Motif prediction value:

[0091] a) Model variables and parameters: The RandomForest model was used, with the frequencies of 192 3bp motifs as feature variables (Table 2), and the 10-fold cross validation method was used to obtain the model parameters.

[0092] Table 2

[0093] Serial number Model variables 3bp motif frequency 1 AAA Freq1 2 AAT Freq2 3 …… …… 191 CCG Freq191 192 CCC Freq192

[0094] Scoring model. The scoring model formula is as follows:

[0095]

[0096] where x i is the input variable, P k (x i ) is the probability that the k-th decision tree predicts that sample i is cancer. Each decision tree is constructed based on Giniimpurity. The value I G The calculation formula is:

[0097] I G =2p c (1-p c )

[0098] where p c is the proportion of cancer samples in a node after each branch of the decision tree.

[0099] Using the classification model and the combined motif frequency of each sample, the category prediction result (Score) of each sample can be obtained.

[0100] c. NF predicted value:

[0101] a) Model variables and parameters: The LinearSVC model was used, with the NF values ​​of 21,334 genes as feature variables (Table 3), and the 10-fold cross validation method was used to obtain the model parameters.

[0102] Table 3

[0103] Serial number Model variables NF value 1 A1BG NF1 2 A1CF NF2 … …… …… 21333 ZYX NF21333 21334 ZZEF1 NF21334

[0104] Scoring model. The scoring model formula and coefficient calculation formula are the same as those of primary model a. Using the classification model and the NF value of each sample, the category prediction result of each sample can be obtained.

[0105] 2. Secondary model variables and parameters The input variables of the secondary model are the 5-fold cross-prediction values ​​of the three primary models (Table 4), and the 10-fold cross-validation method is used to obtain the parameters and thresholds of the secondary model.

[0106] Table 4

[0107] Serial number Model variables Predicted value 1 Fragment prediction value Prediction value 1 2 Motif prediction value Prediction value 2 3 NF predicted value Prediction value 3

[0108] 3. Secondary model formula. The secondary model uses LogisticRegression, and the model formula is as follows:

[0109] Score = w T x i +b

[0110] where x i As input variables, the calculation formulas of model coefficients w and b are as follows:

[0111]

[0112]

[0113] Where yi is the true value of the sample, 1 is cancer, and 0 is healthy. The threshold is 0.104, that is, the model scores below 0.104 are healthy, and those above or equal to 0.104 are cancer.

[0114] 2) Model performance evaluation: Based on the reference value of 0.104, the training cohort samples were divided into healthy people (i.e. predicted to be healthy) and patients with pancreatic and biliary malignancies (i.e. predicted to have pancreatic and biliary malignancies). The pathological test results were used as the true value to evaluate the model prediction performance. The model prediction performance evaluation methods included the area under the ROC curve (AUC, Area Under Curve, range 0-1), positive predictive value (PPV, Positive Predictive Value, range 0-1), specificity (range 0-1) and sensitivity (range 0-1). The higher the value, the better the effect.

[0115] (8) Validation of the risk scoring model for bile and pancreatic malignancies

[0116] In the validation cohort, the effectiveness of the model prediction is verified based on the risk score model and reference values ​​determined in the training cohort. The process is as follows:

[0117] 1) Risk value. In the validation cohort, the risk value of each sample is calculated.

[0118] 2) Model performance verification. Based on the reference value of 0.104, the patients in the validation cohort were divided into healthy people (same as the training cohort) and patients with pancreatic and biliary malignancies. The pathological test results were used as the true value, and the ROC curve was drawn to evaluate the predictive performance of the model, including AUC, PPV, specificity, and sensitivity. The higher the value, the better the effect.

[0119] (9) Optimization of risk scoring model for pancreatic and biliary malignancies

[0120] Combining the fragmentomics diagnostic model with the CA19-9 level in serum can further improve the accuracy of the model. Combined classification model based on CA19-9 level:

[0121] 1) Model establishment: Using the training cohort, combined with the risk scoring model scores and CA19-9 levels, the LinearSVC model was used to build a joint classification model for cancer patients and healthy people. The model consists of three parts: variables, model formulas, and predicted values. The process is as follows:

[0122] 1. Model variables and parameters. The input variables of the model are the score value of the risk scoring model and the log2(1+CA19-9) level. The score value in the training set is the 10-fold cross-prediction value of the risk scoring model, and the score value in the test set is the prediction value of the final risk scoring model. The model parameters and thresholds are all default values. The variables and parameters of the joint classification model are shown in Table 5.

[0123] Table 5

[0124]

[0125] 2. Scoring model. The scoring model formula is the same as the formula of the primary model a in the risk scoring model. The model threshold is 0, that is, in the joint classification model, scores below 0 are healthy, and scores above or equal to 0 are cancer.

[0126] 2) Model performance evaluation: Taking the reference value of 0 as the standard, the training cohort samples were divided into healthy people (i.e. predicted to be healthy) and patients with pancreatic and biliary malignancies (i.e. predicted to have pancreatic and biliary malignancies). The pathological test results were taken as the true value to evaluate the model prediction performance. The model prediction performance evaluation methods included the area under the ROC curve (AUC, Area Under Curve, range 0-1), positive predictive value (PPV, Positive Predictive Value, range 0-1), specificity (range 0-1) and sensitivity (range 0-1). The higher the value, the better the effect.

[0127] 3) Model performance verification: In the validation cohort, the risk value of each sample was calculated. Based on the reference value of 0, the validation cohort patients were divided into healthy people (same as the training cohort) and patients with pancreatic and biliary malignancies. The pathological test results were used as the true value, and the ROC curve was drawn to evaluate the model prediction performance, including AUC, PPV, specificity and sensitivity. The higher the value, the better the effect.

[0128] (10) Application of risk scoring model for bile and pancreatic malignancies

[0129] 1) Collect peripheral blood from patients suspected of having malignant biliary and pancreatic tumors based on imaging test results to obtain plasma cfDNA, and use LP-WGS sequencing to obtain the expression of biomarkers in plasma cfDNA fragments;

[0130] 2) Using the risk scoring model, the cfDNA fragment omics biomarkers were used as input to obtain the model's score value, and the score value and serum CA19-9 expression level were used as input to obtain the risk value of each patient calculated by the joint classification model;

[0131] 3) Compare the risk value with the model reference value to give a predicted result of the risk of each patient suffering from bile and pancreatic malignancies.

[0132] Example 2

[0133] This example verifies the effectiveness of the risk scoring model for gallbladder and pancreatic malignancies.

[0134] (1) Study cohort and clinical information

[0135] A total of 147 patients with pancreatic and biliary malignancies and 71 healthy controls were included in this study and divided into two independent study cohorts. The training cohort included 31 healthy subjects, 28 pancreatic ductal adenocarcinomas, 14 gallbladder cancer patients, and 16 cholangiocarcinoma patients, and the validation cohort included 40 healthy subjects, 34 pancreatic ductal adenocarcinomas, 16 gallbladder cancer patients, and 39 cholangiocarcinoma patients. Clinical characteristics such as gender, TNM tumor stage, and serum tumor marker concentrations are shown in Table 6.

[0136] Table 6

[0137]

[0138]

[0139] (2) Omic characteristics of blood cfDNA fragments

[0140] The blood cfDNA fragment size characteristics, 5' end Motif characteristics and deep pattern characteristics of sequencing data near the transcription start site (TSS) were extracted. The results showed that there were differences in blood cfDNA fragment size, 5' end Motif and deep pattern of sequencing data near the transcription start site (TSS) between healthy people and patients with pancreatic and biliary malignancies, which can be used to distinguish healthy people from patients with pancreatic and biliary malignancies ( Figure 1A-Figure 1C ).

[0141] (3) Establishment and evaluation of a risk scoring model for pancreatic and biliary tumors

[0142] In order to construct a risk classification model for pancreatic and biliary malignancies, the training cohort was used, combined with the results of pathological examinations, and the Stacking algorithm was used to construct an integrated scoring model that distinguishes cancer patients and healthy people with three types of features as input. The model consists of three parts: variables, primary model formulas, secondary model formulas, and predicted values. In the training cohort, based on the cfDNAFragment prediction value, Motif prediction value, and NF prediction value of the plasma of each patient and healthy person in the primary model, combined with 10-fold cross-validation, the reference value was determined to be 0.104. When the risk value is less than the reference value, the sample is predicted to be healthy; otherwise, it is predicted to be pancreatic and biliary malignancies. According to the plasma cfDNA fragment omics biomarkers and pathological test results of each patient in the training cohort, the ROC curve of the training cohort was drawn ( Figure 2 The AUC, specificity, sensitivity and PPV of the model prediction efficiency were 0.978, 96.8%, 91.4% and 98.1%, respectively (Table 7). The results showed that in the training cohort, this risk classification model had a high AUC, specificity, sensitivity and PPV, and the model prediction efficiency was better.

[0143] (4) Validation of the risk scoring model for bile and pancreatic malignancies

[0144] In order to verify the effectiveness of the risk scoring model in predicting pancreatic and biliary malignancies, another independent cohort was selected as the validation cohort. The effectiveness of the model was verified based on the risk classification model and reference value determined in the training cohort. Based on the reference value of 0.104, the patients in the validation cohort were divided into healthy people (same as the training cohort) and patients with pancreatic and biliary malignancies. According to the plasma cfDNA fragment biomarkers and pathological test results of each patient in the validation cohort, the ROC curve of the validation cohort was drawn ( Figure 3 ). The AUC, specificity, sensitivity and PPV of the model prediction efficiency were evaluated to be 0.941, 80.0%, 84.3% and 90.4% respectively (Table 7). The results showed that in the validation cohort, this risk prediction model also had high AUC, specificity, sensitivity and PPV, that is, the model prediction efficiency was better.

[0145] Table 7

[0146]

[0147] (5) Optimization of the risk scoring model for pancreatic and biliary malignancies

[0148] Combining the fragmentomics diagnostic model with the CA19-9 level in serum can further improve the accuracy of the model. Figure 4 ), when the specificity of the training cohort remained unchanged, the sensitivity of the training cohort could be increased to 93.1%, and the sensitivity and specificity of the validation cohort could be increased to 98.9% and 85.0%, respectively. The positive predictive values ​​of the training cohort and validation cohort were 98.2% and 93.6%, respectively (Table 8). The results showed that the plasma cfDNA fragmentomics combined with the serum CA19-9 expression level model could further improve the sensitivity of prediction.

[0149] Table 8

[0150]

[0151] In summary, the method of the present invention overcomes the limitations of cfDNA gene mutation detection in the early stages of tumors with the advantages of minimal invasiveness, high sensitivity for early diagnosis and dynamic monitoring, and avoids information lag or loss caused by tumor heterogeneity.

[0152] The applicant declares that the present invention illustrates the detailed method of the present invention through the above-mentioned embodiments, but the present invention is not limited to the above-mentioned detailed method, that is, it does not mean that the present invention must rely on the above-mentioned detailed method to be implemented. Those skilled in the art should understand that any improvement of the present invention, equivalent replacement of various raw materials of the product of the present invention, addition of auxiliary components, selection of specific methods, etc., all fall within the protection scope and disclosure scope of the present invention.

Claims

1. A method for constructing a cancer early screening model based on cfDNA difference statistics, characterized in that: The construction method includes: performing statistical analysis of differences between tumor-derived cfDNA and cfDNA from healthy individuals by whole-genome sequencing at a relatively low depth, establishing an early screening model for cancer, and realizing non-invasive early screening for cancer; The difference statistics include the difference statistics of cfDNA fragment size characteristics, the difference statistics of the distribution frequency change trend of three consecutive groups of 3bp base combination categories upstream and downstream of the break position of cfDNA, and the difference statistics of the sequencing data coverage depth pattern of cfDNA near TSS.

2. The construction method according to claim 1, characterized in that: The cfDNA fragments include cfDNA short fragments and cfDNA long fragments, the length of the cfDNA short fragments is [130, 177] bp, the length of the cfDNA long fragments is [177, 237] bp, and the ratio of the cfDNA short fragments to the cfDNA long fragments is the characteristic difference value of the cfDNA fragment size; The statistical method for the distribution frequency of three consecutive 3bp base combination categories of cfDNA upstream and downstream of the break position includes: counting 9 bases from the 5' end of the cfDNA and grouping three consecutive bases into three groups of 3bp motifs; The frequency of occurrence of three groups of 3bp motifs was counted for each sample; The parameter calculation method for the sequencing coverage depth of cfDNA near TSS is as follows: the area 500bp [-250bp, 250bp] upstream and downstream of TSS is defined as the central area, and the upstream [-2000bp, -1000bp] and downstream [1000bp, 2000bp] of TSS are defined as the surrounding area; the NF value of the gene is the average coverage of the central area divided by the average coverage of the surrounding area.

3. The construction method according to claim 1 or 2, characterized in that: The method for early cancer screening also includes: performing statistical analysis on the differences between tumor-derived cfDNA and cfDNA from healthy individuals, detecting the serum CA19-9 expression level, and then constructing a model.

4. The construction method according to any one of claims 1 to 3, characterized in that: The cancer includes bile and pancreatic malignancies; Preferably, the biliopancreatic malignancy includes any one of pancreatic cancer, gallbladder cancer or bile duct cancer.

5. A non-invasive early screening device for cancer, characterized in that: The non-invasive early screening device for cancer comprises: a feature extraction module, a machine learning classification model building module and an independent verification cohort evaluation module; The feature extraction module is used to perform cfDNA fragment feature extraction, terminal Motif feature extraction and TSS data feature extraction; The independent validation cohort evaluation module is used to perform validation of the predictive performance of the established machine learning classification model through an independent validation cohort.

6. The non-invasive early screening device for cancer according to claim 5, characterized in that: The cfDNA fragment feature extraction includes: a sequencing data comparison unit, a cfDNA fragment statistics unit and a cfDNA fragment feature determination unit; The sequencing data comparison unit is used to perform the following steps: after removing the sequencing data sequencing adapter, aligning the sequencing data to the human reference genome hg19; The cfDNA fragment statistical unit is used to perform the following steps: statistically analyzing cfDNA fragment length data information; dividing the hg19 autosome into 504 adjacent window fragments without intersection, each window fragment being 5 Mb in length; statistically analyzing the ratio of the number of cfDNA with a length greater than 130 bp and less than 177 bp to the number of cfDNA with a length greater than 177 bp and less than 237 bp in each window region; and finally obtaining the number of cfDNA long and short fragments in each 5 Mb interval. The cfDNA fragment feature determination unit is used to perform the following steps: determining the interval with the largest difference in fragment distribution between cancer patients and healthy controls according to the difference distribution of fragment distribution between cancer patients and healthy controls; defining a short fragment range [130, 177] and a long fragment range [177, 237], and then calculating the standardized z-score of the number of short fragment cfDNA and total fragments in each window of 504 windows as feature input values ​​for model training; The terminal motif feature extraction includes: a sequencing data comparison unit, a reads filtering unit, a combined motif calculation unit and a combined motif statistics unit; The sequencing data comparison unit is used to perform the following steps: after removing the sequencing data sequencing adapter, aligning the sequencing data to the human reference genome hg19; The reads filtering unit is used to perform the following steps: filtering and screening the sequencing data, wherein the filtering criteria are: only considering reads aligned to autosomes 1-22, with a quality score greater than 20; the insert length is between 150-600, and both ends must be properly paired; the reference region of the read does not contain degenerate bases; The combined motif calculation unit is used to perform the following steps: counting 9 bases from the 5' end and dividing them into three groups as the combined motifs; The combined motif statistics unit is used to count the frequencies of the three groups of 3bp motifs for each sample, each group of 3bp motifs has 64 different combinations, and a total of 192 frequencies are used as its training variables; The TSS data feature extraction includes a sequencing data comparison unit, a reads filtering unit, a gene screening and TSS determination unit and a NF value calculation unit; The sequencing data comparison unit is used to perform the following steps: after removing the sequencing data sequencing adapter, aligning the sequencing data to the human reference genome hg19; The reads filtering unit is used to perform the following steps: filtering and screening the sequencing data, wherein the filtering criteria are: only considering reads aligned to autosomes 1-22, with a quality score greater than 20, an insert length between 150-600, and both ends must be properly paired; and the reference region of the read does not contain degenerate bases; The gene screening and TSS determination unit is used to perform the following steps: determining the TSS of gene transcription according to the transcript annotation of the UCSC hg19 genome; for genes with multiple TSSs, only those with a difference of less than 50 bp between different TSSs are retained, and their average value is used as the TSS of the gene, and only genes on autosomes are considered; The NF value calculation unit is used to perform the following steps: obtaining parameters for the sequencing coverage depth of cfDNA near TSS, and the calculation method is: defining the region 500bp [-250bp, 250bp] upstream and downstream of TSS as the central region; defining 2000bp upstream [-2000bp, -1000bp] and downstream [1000bp, 2000bp] of TSS as the surrounding region; and the NF value of a gene is the average coverage of the central region divided by the average coverage of the surrounding region.

7. The non-invasive early screening device for cancer according to claim 5 or 6, characterized in that: The machine learning classification model building module includes: a sample data classification unit, a model parameter acquisition unit and a model effectiveness evaluation unit; The sample data classification unit is used to perform the following steps: dividing the samples into a training set and a test set in a ratio of 4:1, and making the distribution ratios of healthy controls and samples of various cancer types in the two sets consistent; The model parameter acquisition unit is used to perform the following steps: using a machine algorithm in a training set and using a 5-fold or 10-fold cross validation method to obtain model parameters; The model effectiveness evaluation unit is used to perform the following steps: drawing a receiver operating characteristic curve of the training set according to the model prediction value and pathological detection result of each sample in the training set.

8. A cancer early screening model, characterized in that: The input of the cancer early screening model is the score value of the risk scoring model and Log2(1+CA19-9); the output is the model score.

9. The cancer early screening model according to claim 8, characterized in that: The scoring value of the risk scoring model is the scoring value of the test set.

10. The cancer early screening model according to claim 8 or 9, characterized in that: When the model score is lower than 0, it is judged as healthy, and when it is higher than or equal to 0, it is judged as cancer.

Images