Application of gene marker in early screening of multiple cancer species in digestive tract, early screening model construction method and detection device

Abstract

The invention relates to the technical field of molecular biology, in particular to application of a gene marker in early screening of multiple alimentary canal cancer species, an early screening model construction method and a detection device.WGS low-depth sequencing is conducted on plasma sample cfDNA, and three difference characteristics including copy number change, copy number change, cfDNA fragment change, cfDNA fragment change, cfDNA fragment change, cfDNA fragment change, cfDNA fragment change, cfDNA fragment change, cfDNA fragment change, cfDNA fragment change, cfDNA fragment change, cfDNA fragment change and cfDNA fragment change are analyzed according to a high-throughput sequencing result. The method comprises the following steps of: performing training modeling by using a convolutional neural network according to a DNA fragment size ratio and a nucleosome covering mode, realizing high-specificity and high-sensitivity accurate early detection on a plurality of cancers under low sequencing depth, and performing training modeling by using five algorithms, namely a generalized linear model, gradient boosting, a random forest, deep learning and extreme gradient boosting respectively, so as to realize accurate early detection on the plurality of cancers under low sequencing depth. The generalized linear model is subjected to secondary set training to construct multi-feature algorithm integration, and the purpose of tissue traceability detection of multiple cancer species is achieved.

Description

Technical Field

[0001] The present invention relates to application of a gene marker in early screening of multiple digestive tract cancers, an early screening model construction method and a detection device, and belongs to the field of molecular biotechnology. Background Art

[0002] Colorectal cancer, liver cancer and gastric cancer are all common digestive tract cancers. In fact, the treatment effect of digestive tract malignancies depends on the timing of discovery. For example, the 5-year survival rate after surgery for early-stage gastric cancer is over 90%, 66% for stage II, and only 14% for late-stage cancer.

[0003] Colorectal cancer begins in the colon, or large intestine, as small colon growths called polyps. While they are mostly noncancerous growths, they can sometimes develop into cancer.

[0004] Most cancers have a good chance of being treated when they are caught early. It can be easy to ignore the symptoms of gastrointestinal cancer or brush them off as easily treatable conditions such as indigestion, thinking they will go away on their own. This can lead to delays in diagnosis and treatment. Early detection improves the chances of successful treatment, so vigilance and regular screening are vital.

[0005] At present, the most common screening method is digestive endoscopy. Endoscopy can not only visually observe whether there are tumorous lesions in the gastrointestinal tract, but also obtain pathological diagnosis through biopsy. However, due to the low comfort level of gastroenteroscopy, large-scale gastroenteroscopy screening has the disadvantage of low compliance. There are certain limitations for large-scale population screening. The digestive tract multi-cancer detection model has obvious advantages over traditional screening methods. The digestive tract multi-cancer detection model detects blood, which is easy to sample, can reduce patient trauma and risks, improve compliance, increase the proportion of early patient detection, and achieve treatment earlier, thereby improving the patient's five-year survival rate. Summary of the invention

[0006] The present application solves the problems raised in the above-mentioned background technology by proposing an application of a gene marker in the early screening of multiple digestive tract cancers, an early screening model construction method and a detection device.

[0007] To achieve the above object, the present invention provides the following technical solutions: The use of gene markers in the preparation of early screening reagents for multiple digestive tract cancers, wherein the early screening reagents for multiple digestive tract cancers are used to distinguish between patients with multiple digestive tract cancers and healthy people; or, the early screening reagents for multiple digestive tract cancers are used to trace and distinguish the cancer types in patients with multiple digestive tract cancers; The digestive tract cancers include liver cancer, intestinal cancer, and gastric cancer. The gene markers include: First marker: copy number in different windows on chromosomes in WGS data; Second marker: the proportion of short reads and long reads in different windows of cfDNA fragments aligned to the reference genome; The third marker: nucleosome coverage pattern; The first marker is obtained by the following steps: dividing the reference genome into a plurality of windows, and obtaining copy number data in different windows on chromosomes 1 to 22 in the WGS data respectively; The second marker is obtained by the following steps: aligning the read data results to the reference genome, dividing the reference genome into multiple windows, and obtaining the proportion of the number of short reads and the number of ultra-long reads in each window respectively; the short read refers to the read with a length of 60-150bp; the long read refers to the read with a length of 151-220bp; The third marker is obtained by the following steps: selecting transcription factors with known transcription sites from the GTRD database and excluding those transcription factors that are not recorded in the CIS-BP database; then, taking the transcription sites of these selected transcription factors as the center, defining a region from 5 kb upstream to 5 kb downstream of the site as the analysis window, and identifying comparable fragments with a length ranging from 100 to 220 bp within the window; next, performing GC content correction and deep smoothing on the sequencing reads of these fragments to draw a coverage pattern curve for each transcription factor; for each transcription factor, calculating the following three features to construct a nucleosome coverage pattern: (1) Calculate the average sequencing depth within the 1 kb interval above and below the transcription site; (2) Extracting the depth value of the trough from the coverage pattern curve, which represents the central depth of the transcription factor; (3) Perform a fast Fourier transform on the coverage pattern curve to determine the amplitude of the nucleosome amplitude signal peak.

[0008] A method for constructing a digestive tract multi-cancer early screening model, wherein the multi-cancer early screening reagent is used to distinguish between tumor patients with digestive tract multi-cancer and healthy people; the method comprises the following steps: S1: Extraction and sequencing of cfDNA samples: Extract circulating free DNA (cfDNA) from samples of the positive group and the control group, and sequence them to obtain read data; S2: Collect chromosome copy number data: Divide the reference genome into multiple windows, and collect the copy number information of whole genome sequencing (WGS) data in these windows to form the first set of features; S3: Analyze read data and collect window features: align the read data obtained by sequencing with the reference genome, divide the genome into multiple windows, and count the proportion of short reads and ultra-long reads in each window to form the second set of features; S4: Screening and analyzing transcription factors: Transcription factor information was obtained from the GTRD database, and those transcription factors without known transcription sites in the CIS-BP database were excluded. For the remaining transcription factors, the range of -5kb to +5kb near their transcription sites was defined as a window, and comparable fragments with a length of 100-220bp were collected within these windows. These fragments were GC-corrected and sequenced with depth smoothing to obtain the coverage pattern curve of each transcription factor. The following three features were further extracted to form the third set of features: Calculate the average depth of all transcription factors within the range of 1 kb from the upper end to 1 kb from the lower end. The amplitude value of the trough is extracted from the coverage pattern curve as the central depth of the transcription factor; Perform fast Fourier transform on the coverage pattern curve to extract the amplitude value of the peak value of the nucleosome amplitude signal; S5: Model training and construction of early screening model for multiple digestive tract cancers: The above three sets of feature sets are used as initial feature values ​​to form the feature vectors of the model, and the probability of cancer is used as the output value. These feature vectors are input into the convolutional neural network model for training, and finally the early screening model for multiple digestive tract cancers is obtained; Preferably, the window size in S2 is 1Mb; S3 includes: S3-1, dividing the reference genome into multiple windows, and obtaining the number of long reads and the number of short reads in each window respectively; S3-2, standardizing the number of short reads and the number of long reads in all windows of S3-1, and taking the ratio of the normalized number of short reads to the number of long reads as the first eigenvalue; the window size in S3-1 is 5Mb; the short reads are 100-150bp in length, and the long reads are 151-220bp in length.

[0009] Preferably, the multi-cancer early screening model is also used to trace the origin of cancer in patients with multiple cancers; it also includes step S6, taking the first, second, and third groups of features as initial feature values, and inputting them into different classifier models as model feature vectors to train the models, and obtain multiple sub-classifier models for the first, second, and third groups of features respectively; then the sub-classifier models are subjected to secondary set training to obtain a tracing and differentiation model; the different classifier models are one or a combination of several of a generalized linear regression model, a gradient boosting model, a random forest, a deep learning neural network model, and an extreme gradient boosting model; a generalized linear model is used in the secondary set training.

[0010] A multi-cancer early screening detection device is obtained by the construction method described in any of the above technical features; the device consists of the following parts: Sequencing unit: responsible for extracting circulating free DNA (cfDNA) from positive group and control group samples and sequencing them to obtain read segment data; Feature extraction unit 1: divides the reference genome into multiple windows, and collects the copy numbers of WGS data on chromosomes in different windows to form the first set of features; Feature extraction unit 2: align the read data obtained by sequencing with the reference genome, divide the genome into multiple windows, and count the proportion of short reads and ultra-long reads in each window to form the second set of features; Feature extraction unit 3: Filter out transcription factors from the GTRD database and exclude those that have no known transcription sites in the CIS-BP database, a total of 854 transcription factors; for these transcription factors, set the -5kb to +5kb region around their transcription sites as windows, and collect comparable fragments with a length of 100-220bp in these windows; perform GC correction and sequencing depth smoothing on these fragments to generate a coverage pattern curve for each transcription factor; further, extract the following three features for each transcription factor to form the third set of features: 1) Calculate the average depth of the interval from 1 kb above to 1 kb below the transcription site of all transcription factors; 2) Obtain the amplitude value of the trough from the coverage pattern curve as the central depth of the transcription factor; 3) Perform fast Fourier transform on the coverage pattern curve to extract the amplitude value of the highest point of the nucleosome amplitude signal; Sample classification unit: The first, second, and third groups of features are used as initial feature values ​​and input into the convolutional neural network model for training, and the probability of cancer is used as the output value to construct an early screening model for multiple gastrointestinal cancers.

[0011] Compared with the prior art, the present invention has the following beneficial effects: The early screening model constructed in the present invention can perform early screening for multiple types of digestive tract cancers, and can simultaneously distinguish digestive tract tumors (liver cancer, gastric cancer, and intestinal cancer); the early detection model for multiple types of digestive tract cancers can effectively distinguish between cancer and healthy people, with a sensitivity of 69.2% and a specificity of 99.0% in the training set; the performance of the model in the test set: a sensitivity of 75.8% and a specificity of up to 99.6%, with no difference in results between sets, which helps improve the accuracy of early screening for multiple types of cancer. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a schematic diagram of the model building process.

[0013] Figure 2 is a schematic diagram of the construction process of the gastrointestinal cancer early screening detection model (Gastrointestinal Cancer, GI DOC Model).

[0014] Figure 3 is a schematic diagram of the construction process of the gastrointestinal cancer tissue tracing model (GI TOO Model).

[0015] Figure 4 shows the AUC performance of the gastrointestinal multi-cancer early screening model (GI DOC Model) in the training set and validation set.

[0016] Figure 5 shows the score distribution of the digestive tract multi-cancer early screening model (GI DOC Model) in the validation set.

[0017] Figure 6 shows the sensitivity performance of the validation set of the gastrointestinal multi-cancer tissue model (GI TOO Model).

[0018] FIG7 shows the score distribution of the gastrointestinal multi-cancer early screening model (GI DOC Model) in the validation set after grouping according to stages I-IV.

[0019] FIG8 is a confusion matrix of the top 1 cancer detection results of the digestive tract multi-cancer tissue model (GI TOO Model).

[0020] Fig. 9 It is the confusion matrix of the top 2 cancer detection results of the digestive tract multi-cancer tissue model (GI TOO Model). DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] The low-depth WGS (sequencing depth of about 5x) cfDNA sequencing results of 837 cancer patients (suffering from liver cancer, gastric cancer, and intestinal cancer) and 1,113 healthy people were analyzed. Three features, namely, 1MB window fragment copy number variation (CNV), DNA fragment size ratio (FSC), and nucleosome coverage pattern (NP), were included in the statistics. A deep learning convolutional neural network was used to construct a gastrointestinal cancer detection model (GI DOC Model), and five different algorithms were used to train a gastrointestinal cancer tissue of origin model (GI TOO Model).

[0023] The present invention provides a multi-molecule, multi-feature, and multi-training algorithm diagnostic model based on high-throughput, low-depth whole-genome sequencing of plasma cfDNA. A convolutional neural network is used to construct an early detection model for multiple digestive tract cancers, and a variety of complex gene feature markers are screened and trained. The model can not only diagnose three early digestive tract cancers and trace their tissue origins, but also has the characteristics of non-invasive detection, low throughput, and high detection specificity and sensitivity. The model uses a validation set consisting of 837 cancer patients (suffering from one of the three cancers involved in the present invention) and 1113 healthy people to verify the model performance, and the validation set can still achieve high specificity and high sensitivity. In the DOC model, 837 patients and 1113 healthy people samples were used to construct the model, while in the TOO model, 837 patient samples were used to construct the model.

[0024] The present invention first requires the steps of extracting, building a library, and sequencing cfDNA from a blood sample. The extraction and library building methods are not particularly limited and can be adjusted from the extraction methods in the prior art. During the sequencing process, the sequencing technology in the prior art can be used to obtain the base information of cfDNA. The data set used in the model building process of the present invention is as follows: The sample data included in this patent are derived from the data collected by the patent applicant during the testing work. The sample data of the training set and validation set used in this patent are all derived from actual clinical samples collected by multiple hospitals that have independent cooperation with the patent applicant.

[0025] Table 1: Dataset grouping information

[0026] Table 2: Cancer type information of the dataset

[0027] Extraction and sequencing method of plasma cfDNA samples: Use purple blood collection tubes (EDTA anticoagulant tubes) to collect 8ml of whole blood samples from patients, centrifuge and separate plasma in time (within 2 hours), and after transporting to the laboratory, use QIAGEN plasma DNA extraction kit to extract cfDNA according to the instructions. After the collection of cfDNA samples is library-built, WGS sequencing is performed, and the sequencing depth is about 5X. After obtaining the offline data, the data is aligned to the human reference genome to obtain the base data information of the corresponding read segment.

[0028] Data processing The marker data in the present invention mainly utilizes three molecular features: 1 Mb-Bin Copy Number Variation (CNV) It has been widely recognized by the scientific community that copy number variation plays an important role in the occurrence of cancer. At present, by analyzing the copy number of certain key genes or genomic regions related to cancer, it is possible to distinguish cancers to a certain extent. However, there are still some rare or undiscovered genes and genomic regions that may hide important information about copy number variation.

[0029] During the data collection process, the following method was used: First, for the whole genome sequencing (WGS) data of each sample, the reference genome of chromosomes 1 to 22 was divided into non-overlapping windows of size 1Mb. Then, the bedtoolscoverage tool was used to calculate the read depth within each window. In addition, it is necessary to correct it according to the GC content and average alignment ability of each window (recorded in the UCSC BigWig file). Through this process, individual read depth data for 2475 windows can be collected. For further analysis, the hidden Markov model (HMM) and the control baseline depth of each window population were used to construct the logarithm of copy number changes for each window, calculated as log2 (normalized depth after correction of the sample to be tested / normalized depth after correction of the population baseline). In this way, we can obtain the copy number variation information of each sample to be tested.

[0030] 2. Fragmentation Size Coverage (FSC) The ratio of DNA fragment sizes reflects the distribution characteristics of circulating free DNA (cfDNA) fragment lengths. Through machine learning technology, the coverage depth ratio of DNA fragment size can be used to build a predictive model that can help identify different types of cancer. When studying patients with one of the three cancers covered by the present invention, significant differences were observed in the distribution of cfDNA fragments of 100-150 base pairs (bps) and 151-220bps in length on chromosomes, which can serve as important features for distinguishing different types of cancer.

[0031] 3. Nucleosome Profile (NP) Transcription factors were screened from the GTRD database (https: / / gtrd.biouml.org / #!) (v21.12), while those that were not recorded in the CIS-BP database (http: / / cisbp.ccbr.utoronto.ca / ) (v2.00) were excluded. After screening, 854 transcription factors were identified, all of which had more than 10,000 highly matched sites.

[0032] For these selected transcription factors, a window region was defined, i.e., -5kb to +5kb around the transcription site, and aligned fragments with a length of 100-220bp were searched within this window. For these fragments, GC content correction was performed, and a cubic polynomial Savitzky-Golay filter was applied to the sequencing depth curve for smoothing, and finally the coverage pattern curve of each transcription factor was obtained.

[0033] After obtaining the coverage pattern curve, we extracted three features for each transcription factor (a total of 854×3=2563 features were extracted) Calculate the average depth of all transcription factors within the range of 1 kb from the upper end to 1 kb from the lower end. Determine the depth of transcription factor central sites; The coverage pattern curve was subjected to fast Fourier transform to extract the amplitude value of the peak value of the nucleosome amplitude signal.

[0034] Through the above data acquisition, the initial data vectors of these three types of data can be obtained respectively. The three types of features are input into the convolutional neural network to construct the digestive tract multi-cancer early screening detection model (GI DOC Model), and the model used can be used to distinguish between tumor patients (liver cancer, gastric cancer, intestinal cancer) and healthy people; further, on this basis, the above three types of data are input into five different algorithms to integrate and construct the digestive tract multi-cancer tissue tracing model (DOC TOO Model), which can further distinguish the three types of digestive tract tumors in patients known to have digestive tract tumors.

[0035] In the construction of the above two models, the classification models used are as follows: Convolutional neural network (CNN); Generalize Linear Model (GLM); Random Forest (RF); Deep Learning (DL); Extreme Gradient Boosting (XGBoost); In the process of modeling, the Random Grid Search Parameters algorithm can also be used to optimize the model according to the existing parameter optimization method.

[0036] The construction framework of the digestive tract multi-cancer early screening model (GI DOC Model) is as follows: The Digestive Tract Multi-cancer Early Screening Model (GI DOC Model) combines convolutional neural networks and stacking ensemble learning technology. The basic convolutional model is relearned through the meta-learning process to integrate various features and optimize their combination. The structural design of the model is as follows: Data input part: responsible for receiving the input feature vector.

[0037] Feature extraction part: It consists of three convolutional layers, each of which uses a convolution kernel to extract features.

[0038] Feature integration and classification part: contains three fully connected layers for higher-level integration and classification of features.

[0039] Result output part: output the probability of the sample suffering from cancer.

[0040] Input details of feature vector: The first, second, and third features are directly fed into the input layer of the model as input vectors.

[0041] The processing flow of the convolutional layer: The feature vector is first processed by a convolution kernel.

[0042] Then the ReLU activation function is applied to introduce nonlinearity and enhance the expressiveness of the model.

[0043] Batch normalization is then performed to speed up the training process and improve the stability of the model.

[0044] Finally, the maximum pooling operation is used to reduce the computational burden and extract key features.

[0045] The processing flow of the fully connected layer: The feature vector processed by the convolutional layer is further activated by ReLU.

[0046] It then passes through three fully connected layers, which are responsible for integrating and classifying the features.

[0047] Function of the output layer: Output a predicted value, indicating the probability that the sample has cancer.

[0048] The three gene marker feature vectors are input into the convolutional neural network respectively to form three basic convolution models. The results of these models are then subjected to secondary ensemble learning to obtain the final output results.

[0049] In addition, by analyzing the impact of model output on input gradients, the importance of input features can be predicted. The importance of the three features is analyzed, and the eigenvalues ​​are simplified. The eigenvalue with the largest contribution in each feature set is selected as the component of the final model input vector. The specific performance of each feature is shown in Table 3 (each feature selects the most important variable): Table 3. Important characteristic variables of copy number variation (CNV)

[0050] Table 4. Important characteristics of DNA fragment size fraction (FSC)

[0051] In Table 4, the numbers after the names indicate the sequential number of the windows, long refers to long reads, and short refers to short reads. Table 5. Important characteristic variables of nucleosome coverage pattern (NP)

[0052] Mean in Table 5 refers to the average depth of the upper 1Kb and the lower 1kb of all transcription sites of the transcription factor, central refers to the central depth of the transcription factor, amplitude refers to the amplitude value of the highest point of the nucleosome amplitude signal, and the last letter string represents the name of the transcription factor.

[0053] The design framework of the multi-cancer tissue tracing model (TOO Model) is mainly aimed at those samples that are initially judged as positive in the early screening model for multiple gastrointestinal cancers, and further tracing predictions of cancer types are performed. The model also uses three gene marker features as input and is analyzed by the following five algorithms: generalized linear model, gradient boosting algorithm, random forest, deep learning, and XGBoost (extreme gradient boosting model). Each algorithm optimizes hyperparameters through random search to adapt to sub-model training of different features. Finally, the optimal sub-model for each feature is screened by the mean square error (MSE) criterion, and 3 best models are selected for each gene marker, resulting in a total of 9 optimal sub-models. At the same time, the eigenvalues ​​are streamlined, and the eigenvalues ​​with the largest contribution in each feature set are selected as the input of the final sub-model.

[0054] In order to further improve the predictive ability of the classifier, these trained and screened optimal sub-model results are trained for secondary integration, using Stacking ensemble learning technology. Stacking improves the predictive performance of the model by meta-learning multiple primary classifiers, collecting the characteristics of each classifier, and finding the best integration method. In this patent, Stacking uses the generalized linear model (GLM) as a training algorithm, and establishes the relationship between the mathematical expectation of the response variable and the linear combination of the predictor variables through the link function.

[0055] The digestive tract multi-cancer tissue tracing model (GI TOO Model) is a multi-classification model whose output is the probability of different cancer types. The model predicts three types of cancer, including liver cancer, intestinal cancer, and gastric cancer, for the samples to be predicted, covering the cancer types included in this patent. The final judgment result is based on the highest probability (Top1) and the second highest probability (Top2) of the three cancer prediction results. The important features of the digestive tract multi-cancer tissue tracing model are shown in Table 6: Table 6. Important characteristic variables of copy number variation (CNV) feature Bowel cancer Gastric cancer Liver cancer Average contribution CNV.1.100000001.101000000 1.2786896827630700 0.3690773808255740 0.00041149670093140000 0.5493928534298580 CNV.1.10000001.11000000 1.9019219311885500 0.4724324187360830 0.00058745875683375 0.79164726956049 CNV.1.101000001.102000000 0.6284261612294360 0.33785481032046900 0.000336229242128125 0.32220573359734400 CNV.1.102000001.103000000 0.592727802018635 0.3045084358318950 0.000306218320055 0.29918081872352800 CNV.1.105000001.106000000 1.007315464131530 0.35406608491052200 0.00039842030444937500 0.45392665644883500 CNV.1.106000001.107000000 1.3625074437198500 0.36659439496104600 0.00043764499740475 0.5765131612261020 CNV.1.107000001.108000000 1.1705616099759900 0.4441364602027850 0.00051681590610965 0.5384049620282960 CNV.1.109000001.110000000 1.2744682926684600 0.45954817051738500 0.000546600063524275 0.5781876877497900 CNV.1.110000001.111000000 0.6738882806152110 0.38274446133715000 0.000408934704726175 0.35234722555236200 CNV.1.11000001.12000000 2.4520381782203900 0.4844059367735650 0.000591732776644275 0.9790119492568670 CNV.1.111000001.112000000 1.3761742256581800 0.40358368841620200 0.00051587934556825 0.5934245978066500 CNV.1.112000001.113000000 0.9938814794644710 0.40234709113607800 0.0004292909635031 0.46555262052135100 CNV.1.113000001.114000000 1.9610709873959400 0.5243085911286660 0.00062172426383815 0.8286671009294820 CNV.1.114000001.115000000 0.7658918304368850 0.38491388557226300 0.000415638117165025 0.38374045137543800 CNV.1.115000001.116000000 1.2109789448635000 0.3740746162628300 0.000428586710441725 0.528494049278925 CNV.1.116000001.117000000 1.0938933945260900 0.44999119008037100 0.0004925945526268000 0.5147923930530310 CNV.1.117000001.118000000 1.1841086926870000 0.5214428658765290 0.000584877095794475 0.568712145219776 CNV.1.118000001.119000000 0.8090228088200090 0.3814082387901140 0.000396768899165075 0.39694260550309600 CNV.1.119000001.120000000 1.2095829464495200 0.40782406536966100 0.000451537533228475 0.5392861831174690 CNV.1.12000001.13000000 1.6938699384219900 0.5186626740807600 0.000580694604954 0.7377044357025690 CNV.1.14000001.15000000 1.5924279242754000 0.4542066472902520 0.00053494799524145 0.6823898398536300 CNV.1.150000001.151000000 1.2392472738400100 0.28651487455574600 0.000338911776406625 0.508700353390721 CNV.1.15000001.16000000 2.248176991629100 0.47034329076305700 0.000755339292281975 0.9064252072281480 CNV.1.151000001.152000000 2.2673115594952800 0.33946535069570200 0.00042980211863035000 0.8690689041032050 CNV.1.152000001.153000000 1.3887670935131600 0.31897099468655400 0.00037438822895240000 0.5693708254762230 CNV.1.153000001.154000000 1.348928533308210 0.3365491000998730 0.00040002683852440000 0.5619592200822020 CNV.1.154000001.155000000 3.2504698380827900 0.28421529724399600 0.00042175992877325 1.178368965085190 CNV.1.155000001.156000000 5.710675969564670 0.4395912582452550 0.000706129895248025 2.0503244525683900 CNV.1.156000001.157000000 5.367686677879340 0.418513275399357 0.001430818163890130 1.9292102571475300

[0056] Table 7. Important characteristics of DNA fragment size fraction (FSC) feature Bowel cancer Gastric cancer Liver cancer Average contribution FSC.longA1 1.3919353146106000 0.3703590072839980 0.0010642220663246300 0.5877861813203080 FSC.longA10 1.602135417982940 0.2562746265714910 0.000823978717392075 0.6197446744239400 FSC.longA100 1.0763407358899700 0.3185651565172540 0.00096272991887795 0.4652895407753680 FSC.longA101 2.233802641183140 0.28710405872492600 0.0009634589685507000 0.8406233862922050 FSC.longA102 0.9847581611946230 0.2887674944826910 0.000816236963097875 0.4247806308801370 FSC.longA103 2.527864120900630 0.33158019546037700 0.00111891789232185 0.9535210780844430 FSC.longA104 1.940945841372010 0.33301300755594000 0.001111844282229130 0.7583568977367270 FSC.longA105 1.3959414243872700 0.2827649823587820 0.00080087090393365 0.5598357592166630 FSC.longA106 3.551663603633640 0.37367877995328700 0.001307444824728330 1.308883276137220 FSC.longA107 2.2814438892528400 0.2934277854882420 0.000954949854783 0.8586088748652890 FSC.longA108 1.9865540970737700 0.29791221710431500 0.00093663860544485 0.7618009842611780 FSC.longA109 1.9837863172094000 0.3163787243289280 0.00102981015029985 0.7670649505628770 FSC.longA11 1.40243571329241 0.22604794450125700 0.0006629821943778 0.5430488799960150 FSC.longA110 5.82168640041103 0.308732265995345 0.0012914883415246 2.0439033849159700 FSC.longA111 1.6702717719599600 0.2918945523006330 0.000893547474223875 0.6543532905782730 FSC.longA112 3.197900162388880 0.45748248246422500 0.001648158169242980 1.2190102676741200 FSC.longA113 2.264299506942430 0.3362049532937720 0.0010446647968239 0.8671830416776760 FSC.longA114 1.4135142628413900 0.32252247251192900 0.0009573263084713250 0.5789980205539310 FSC.longA115 1.4157111238067300 0.3588890907033530 0.0011027593774041 0.5919009912958310 FSC.longA116 0.9151425984843320 0.31785949961152500 0.000899204926615725 0.41130043434082400 FSC.longA117 1.4635597279605800 0.24232219493228000 0.000759151327878725 0.5688803580735780 FSC.longA118 3.4827224280064300 0.3530971977749190 0.00126940674097355 1.2790296775074400 FSC.longA119 1.5938994421934100 0.34031133507682500 0.00107402650600225 0.6450949345920790 FSC.longA12 2.7441196400672200 0.3270133543573660 0.001035418242177530 1.0240561375555900 FSC.longA120 2.1367661754290300 0.48221638226498600 0.001641009746685230 0.8735411891468990 FSC.longA121 1.306786835193630 0.26107434704631100 0.00076360387484395 0.5228749287049300 FSC.longA122 2.8476286359131300 0.36480120497116300 0.001125637219994580 1.071185159368100 FSC.longA123 0.8727637966318680 0.2761954067761380 0.000831976441061175 0.3832637266163560 FSC.longA124 1.3460130129009500 0.3494403623899790 0.0010797059313186 0.5655110270740820

[0057] Table 8. Important characteristic variables of nucleosome coverage pattern (NP) Feature Colorectal cancer Gastric cancer Liver cancer Average contribution value NP.854.amplitude.ADNP 2.1679138283540500 0.12244784350974100 0.0003497098784604 0.7635704605807490 NP.854.amplitude.AEBP2 1.5798581648381300 0.10229514040324900 0.00021816222344727200 0.560790489154943 NP.854.amplitude.AHR 1.6752744868087300 0.17394615922803400 0.0005186470846872000 0.6165797643738160 NP.854.amplitude.AHRR 2.206660765921700 0.14823007183478900 0.00038473955463545 0.7850918591037070 NP.854.amplitude.ALX4 1.5587990166503000 0.10752902265366600 0.00021603538417805800 0.5555146915627150 NP.854.amplitude.AR 1.2044503093362500 0.1285129463978360 0.00023820895864127500 0.44440048823091100 NP.854.amplitude.ARHGAP35 0.7349496383446250 0.1112786970319330 0.0001831552566525 0.28213716354440300 NP.854.amplitude.ARID2 1.2058325541050500 0.11144040691708300 0.000341729449325975 0.4392048968238190 NP.854.amplitude.ARID3A 2.4362364285237500 0.16312465541123700 0.00038222814258835 0.8665811040258580 NP.854.amplitude.ARID3B 2.511954737541060 0.13952973813560400 0.00031587200479765 0.8839334492271520 NP.854.amplitude.ARID5B 2.088676296174530 0.1812226450551460 0.000615641145625925 0.7568381941250990 NP.854.amplitude.ARNT 2.4022221951978300 0.18175970382607500 0.000588521780205875 0.8615234736013690 NP.854.amplitude.ARNT2 2.983468225691470 0.13788825754745100 0.0005168745207662 1.0406244525865600 NP.854.amplitude.ARNTL 1.9796264115884700 0.08696850942455420 0.0003226558167595260 0.6889725256099280 NP.854.amplitude.ASCL1 3.2656213613033000 0.13468292863852500 0.0003871211104051 1.1335638036840800 NP.854.amplitude.ASCL2 1.1480444050205800 0.1356911767949960 0.00024531919314750000 0.4279936336695740 NP.854.amplitude.ASH1L 2.276006334938990 0.10441808334297900 0.000373277018463375 0.793599231766811 NP.854.amplitude.ATF1 0.7767875546310620 0.0548722651353718 9.48907018943219E-05 0.2772515701561090 NP.854.amplitude.ATF2 2.9428095097343100 0.16327951560989000 0.0004885742584103 1.0355258665342000 NP.854.amplitude.ATF3 13.673710762678300 0.22913241895830000 0.0031447941324962 4.635329325256380 NP.854.amplitude.ATF4 1.4130508345902900 0.1280959846101290 0.00023729658629012500 0.5137947052622350 NP.854.amplitude.ATF5 1.741442037804520 0.1592175728467750 0.0004190844846313 0.6336928983786410 NP.854.amplitude.ATF6 2.4607145852409300 0.16042516612443500 0.0003425109448438 0.8738274207700700 NP.854.amplitude.ATF7 1.5755344420361000 0.1787433098117820 0.00035043009701040000 0.5848760606482980 NP.854.amplitude.ATOH1 1.699213701688370 0.11898117208892300 0.00037060877841945 0.6061884941852390 NP.854.amplitude.BACH1 2.532371500794400 0.1179842767040900 0.000382608388254025 0.8835794619622470 NP.854.amplitude.BACH2 6.786259239565290 0.15772564350529200 0.0008686695133387000 2.314951184194640 NP.854.amplitude.BARHL1 1.637048591790520 0.11892002174924900 0.0006194552715187 0.5855293562704280 NP.854.amplitude.BARX1 1.452544597924950 0.0850031575950884 0.00016474958156196400 0.5125708350338670

[0058] The early detection model for multiple digestive tract cancers can effectively distinguish cancer from healthy people. The sensitivity in the training set is 69.2% and the specificity is 99.0%. The performance of the model in the test set is: sensitivity 75.8%, specificity can reach 99.6%, and there is no difference in the results between the sets. The specific results are shown in Tables 9 to 12: Table 9 Cancer detection sensitivity of the digestive tract multi-cancer early screening model (GI DOC Model)

[0059] Table 10 Specific performance of the digestive tract multi-cancer early screening model (GI DOC Model) in healthy people

[0060] Table 11 Sensitivity of cancer detection by stage in the multi-cancer early screening model (DOC Model)

[0061] Table 12 Detection sensitivity of various cancers in the digestive tract multi-cancer early screening model (GI DOC Model)

[0062] The digestive tract multi-cancer tissue tracing model (GI TOO Model) is used to trace the cancer types of positive samples. After the model is trained, the validation samples are used to predict the classification results to obtain the traceability probabilities of different digestive tract cancer types. The cancer types ranked first and second in probability (first / second classification) are selected to verify the accuracy. Its performance is shown in Tables 13 and 14. Table 13 Performance of the digestive tract multi-cancer tissue tracing model (GI TOO Model)

[0063] More specific test data is shown in Table 14: Table 14 Performance data of the traceability model for different digestive tract cancers

[0064] The method of the present invention can classify different types of cancer, and can be combined with relevant clinical actual indicators to make reasonable selections in clinical practice.

Claims

1. The use of gene markers in the preparation of early screening reagents for multiple digestive tract cancers, characterized in that: The digestive tract multiple cancer early screening reagent is used to distinguish between patients with digestive tract multiple cancers and healthy people; or, the digestive tract multiple cancer early screening reagent is used to trace and distinguish the cancer types in patients with digestive tract multiple cancers; The digestive tract cancers include liver cancer, intestinal cancer, and gastric cancer. The gene markers include: First marker: copy number in different windows on chromosomes in WGS data; Second marker: the proportion of short reads and long reads in different windows of cfDNA fragments aligned to the reference genome; The third marker: nucleosome coverage pattern; The third marker is obtained by the following steps: obtaining transcription factors from the GTRD database, and excluding transcription factors that do not have known transcription sites in the CIS-BP database; taking the range of -5kb to +5kb near the transcription site of the obtained transcription factor as a window, obtaining fragments with a length of 100-220bp that can be aligned to these windows, and performing GC correction and sequencing depth smoothing on the read data in the window in turn to obtain a coverage pattern curve for each transcription factor; for each transcription factor, obtaining the following three features, which are collectively used as the nucleosome coverage pattern: 1) for all transcription sites of the transcription factor, calculating the average depth from the upper 1kb to the lower 1kb of the transcription site; 2) for the obtained coverage pattern curve, obtaining the amplitude value of the curve trough as the center depth of the transcription factor; 3) performing fast Fourier transform on the obtained coverage pattern curve to obtain the amplitude value of the highest point of the nucleosome amplitude signal.

2. The use according to claim 1, characterized in that: The first marker is obtained by the following steps: dividing the reference genome into multiple windows, and respectively obtaining copy number data in different windows on chromosomes 1-22 in the WGS data.

3. The use according to claim 1, characterized in that: The second marker is obtained by the following steps: aligning the read data results to the reference genome, dividing the reference genome into multiple windows, and obtaining the proportion of short reads and ultra-long reads in each window respectively; the short reads refer to reads with a length of 60-150bp; the long reads refer to reads with a length of 151-220bp.

4. A method for constructing a digestive tract multi-cancer early screening model, characterized in that: The multi-cancer early screening reagent is used to distinguish between patients with multiple digestive tract cancers and healthy people; it includes the following steps: S1: Extraction and sequencing of cfDNA samples: Extract circulating free DNA from samples of the positive group and the control group, and sequence them to obtain read data; S2: Collect chromosome copy number data: divide the reference genome into multiple windows, and collect the copy number information of the whole genome sequencing data in these windows to form the first set of features; S3: Analyze read data and collect window features: align the read data obtained by sequencing with the reference genome, divide the genome into multiple windows, and count the proportion of short reads and ultra-long reads in each window to form the second set of features; S4: Screening and analyzing transcription factors: Transcription factor information was obtained from the GTRD database, and those transcription factors without known transcription sites in the CIS-BP database were excluded. The range of -5kb to +5kb near the transcription sites of the obtained transcription factors was used as a window to obtain fragments of 100-220bp in length that could be aligned to these windows. The read data in the windows were subjected to GC correction and sequencing depth smoothing in turn to obtain the coverage pattern curve of each transcription factor. For each transcription factor, the following three features were further extracted to form the third set of features: 1) Calculate the average depth of all transcription factor transcription sites within the range of 1 kb above and 1 kb below. 2) Extract the amplitude value of the trough from the coverage pattern curve as the central depth of the transcription factor; 3) Perform fast Fourier transform on the coverage pattern curve to extract the amplitude value of the peak value of the nucleosome amplitude signal; S5: Model training and construction of an early screening model for multiple gastrointestinal cancers: The above three sets of feature sets are used as initial feature values ​​to form the feature vectors of the model, and the probability of cancer is used as the output value. These feature vectors are input into the convolutional neural network model for training to obtain an early screening model for multiple gastrointestinal cancers.

5. The method for constructing a multi-cancer early screening model according to claim 4, characterized in that: The window size in S2 is 1Mb; S3 includes: S3-1, dividing the reference genome into multiple windows, and obtaining the number of long reads and the number of short reads in each window respectively; S3-2, standardizing the number of short reads and the number of long reads in all windows of S3-1, and taking the ratio of the normalized number of short reads to the number of long reads as the first eigenvalue; the window size in S3-1 is 5Mb; the short reads are 100-150bp in length, and the long reads are 151-220bp in length.

6. The method for constructing a multi-cancer early screening model according to claim 4, characterized in that: The multi-cancer early screening model is also used to trace the origin of cancer in patients with multiple cancers; it also includes step S6, taking the first, second and third feature sets as initial feature values, and inputting them into different classifier models as model feature vectors to train the model, and obtain multiple sub-classifier models for the first, second and third feature sets respectively; then the sub-classifier models are subjected to secondary set training to obtain a tracing and differentiation model; the different classifier models are one or a combination of several of the generalized linear regression model, gradient boosting model, random forest, deep learning neural network model, and extreme gradient boosting model; the generalized linear model is used in the secondary set training.

7. A multi-cancer early screening detection device, characterized in that: The device is obtained by the construction method described in any one of claims 4 to 6; the device consists of the following parts: Sequencing unit: responsible for extracting circulating free DNA from positive group and control group samples and sequencing them to obtain read segment data; Feature extraction unit 1: divides the reference genome into multiple windows, and collects the copy numbers of WGS data on chromosomes in different windows to form the first set of features; Feature extraction unit 2: align the read data obtained by sequencing with the reference genome, divide the genome into multiple windows, and count the proportion of short reads and ultra-long reads in each window to form the second set of features; Feature extraction unit 3: used to obtain transcription factors from the GTRD database and exclude transcription factors that do not have known transcription sites in the CIS-BP database, with a total of 854 transcription factors counted; the range of -5kb to +5kb near the transcription sites of the obtained transcription factors is used as a window to obtain fragments of 100-220bp in length that can be aligned to these windows, and the read data in the windows are sequentially GC corrected and sequence depth smoothed to obtain the coverage pattern curve of each transcription factor; for each transcription factor, the following three feature sets are obtained to form the third set of features: 1) For all transcription sites of the transcription factor, the average depth from the upper 1kb to the lower 1kb of the transcription site is calculated; 2) For the obtained coverage pattern curve, the amplitude value of the curve trough is obtained as the central depth of the transcription factor; 3) For the obtained coverage pattern curve, a fast Fourier transform is performed to obtain the amplitude value of the highest point of the nucleosome amplitude signal; Sample classification unit: The first, second and third groups of features are used as initial feature values ​​and input into the convolutional neural network model for training, and the probability of cancer is used as the output value to construct an early screening model for multiple gastrointestinal cancers.

8. The device according to claim 7, characterized in that In feature extraction unit 1, the window size is 1Mb.

9. The device according to claim 7, characterized in that In feature extraction unit 2, the window size is 5Mb.

Images