WO2023197825A1 - Multi-cancer early screening model construction method and detection device - Google Patents

Abstract

The present invention relates to a multi-cancer (lung cancer, intestinal cancer and liver cancer) early detection and cancer prediction method, a detection device, and a computer readable medium. The present invention provides: performing WGS lowpass sequencing on plasma samples cfDNA; using a high-throughput sequencing result to analyze five discriminative features of cfDNA fragments of cancers, which comprise genome-wide fragment length coverage distribution, fragment length distribution on long and short arms of chromosomes, a fragment breakpoint sequence, a fragment 5'end sequence, and 1 MB window fragment copy number variation; then using a generalized linear model, a gradient boosting machine, a random forest, a deep learning algorithm, and an extreme gradient boosting algorithm to respectively perform training modeling; and next using the generalized linear model to perform secondary set training to construct a multi-feature multi-algorithm integration model. The purpose of low-depth, high-specificity and high-sensitivity non-invasive precise early detection and origin of tissue detection for multiple cancers is realized.

Description

Multi-cancer early screening model construction method and detection device Technical field

The invention relates to a detection of tissue origins of multiple cancer types including lung cancer (Lung Adenocarcinoma, LUAD), colorectal cancer (CRC), and liver cancer (Primary Liver Cancer, PLC), and belongs to the field of molecular biomedicine technology.

Background technique

Lung cancer, colorectal cancer, and liver cancer are the three malignant tumors with the highest mortality rates worldwide.

Lung cancer, liver cancer and colorectal cancer have low early diagnosis rates due to lack of obvious symptoms or difficulty in detection. However, most of the cancer early screening products currently on the market are aimed at predicting a single cancer type. If patients need to undergo multiple early screening projects for different single cancer types, due to time-consuming, laborious and high costs, it may reduce the effectiveness of early screening for various cancer types in a wide range of people. implementation and promotion. Early screening for multiple cancer types not only covers the early screening of various cancer types, but also accurately detects their tissue origin, preventing unknown primary cancers that may appear during the development of cancer, which may complicate the condition and delay diagnosis and treatment. Therefore, our country urgently needs an early screening product that simultaneously covers the above three malignant tumors with the highest mortality rate, so as to be more efficient, economical and practical and applicable to a wider range of people.

Contents of the invention

The present invention provides a method for WGS low-depth sequencing of plasma sample cfDNA, and uses high-throughput sequencing results to analyze five differential characteristics of cfDNA fragments of various cancer types, including genome-wide fragment length distribution, fragment length distribution on each long and short arm of the chromosome, The sequence at the fragment breakpoint (8-mer Breakpoint Motif), the 5' end sequence of the fragment (8-mer End Motif) and the fragment copy number changes in the 1MB window were analyzed using generalized linear model (GLM), gradient boosting machine (GBM), and random Four algorithms, namely Random Forest, Deep Learning and XGBoost, are used for training and modeling respectively. Finally, a multi-feature multi-algorithm integrated model is built through the generalized linear model (GLM) to achieve multi-cancer diagnosis. The purpose of a low-depth, high-specificity and high-sensitivity non-invasive and precise Tissue of Origin (TOO) detection.

The first purpose of the present invention:

A method for constructing a multi-cancer early screening model. The model is used to classify whether a sample has intestinal cancer, lung cancer or liver cancer, including the following steps:

Step 1: Extract and sequence cfDNA from the samples of the positive group and the control group to obtain read data;

Step 2: Compare the read data results to the reference genome, divide the reference genome into multiple windows, and obtain the number of all reads, the number of short reads, and the number of ultra-long reads within each window, as The first feature set;

Step 3: Align the read data results to the reference genome, using the long arm and short arm of each chromosome as the region range, and obtain the number of reads in gradient intervals of different lengths within each range as the second feature set;

Step 4: Use the m base data at the 5' end in the read data as a base fragment set, and obtain the proportion of various base fragments in all fragments as the third feature set;

Step 5: Compare the read data results to the reference genome to obtain the position of the 5' end of the read on the reference genome. Position; obtain the sequence data of n bp bases upstream and downstream of the position as a set of base fragments; use the proportion of the obtained various base fragments in all fragments as a fourth feature set;

Step 6: Divide the reference genome into multiple windows and obtain copy number data within each window as the fifth feature set;

Step 7: Use the first, second, third, fourth and fifth feature sets together as the initial feature value and input it into the classification model as the model feature vector, and use whether you have cancer as the output value to train the model. Get early screening models.

In step 6, suffering from cancer refers to suffering from any one of intestinal cancer, lung cancer or liver cancer.

In the described step 6, it is also necessary to simplify the initial feature values and then use them as model feature vectors. The simplification means to respectively screen out the first, second, third, fourth and fifth feature sets in the positive group. There is a significant difference between the sample and the control group.

The screening process described was through the analysis of variance method.

The short reads are 40-80 bp in length, the number of ultra-long reads is 200-300 bp, and all reads are in the range of 40-300 bp in length.

The window size range in step 2 is 2-7Mb.

The gradient intervals of different lengths in step 3 refer to the gradient ranges of different lengths obtained by increasing in steps of 8-12 bp in the range of 40-300 bp.

The number of reads stated is normalized.

In step 4, m is any integer between 6 and 10.

In step 5, n is any integer between 2 and 5.

The window in step 6 is obtained by dividing the reference gene chromosomes 1-22 with a length of 0.8-1.2Mb without overlap.

The input to the classification model in step 7 refers to inputting the first, second, third, fourth and fifth feature sets to the generalized linear model, gradient boosting algorithm model, random forest model, deep learning model and extreme model respectively. In the gradient boosting model, multiple sub-models are obtained and connected into a linear relationship model.

Second purpose of the present invention:

A multi-cancer detection device, the device is used to classify whether a sample has intestinal cancer, lung cancer or liver cancer, including:

The sequencing module is used to extract and sequence cfDNA from the samples of the positive group and the control group to obtain read data;

The first feature set acquisition module is used to compare the read data results to the reference genome, divide the reference genome into multiple windows, and obtain the number of all reads, the number of short reads, and the number of super-reads within each window. The number of long reads, as the first feature set;

The second feature set acquisition module is used to compare the read data results to the reference genome, using the long arm and short arm on each chromosome as the regional range, and obtain the reads in different length gradient intervals within each range. The number of segments, as the second feature set;

The third feature set acquisition module is used to use the m base data at the 5' end in the read data as a base fragment set, and obtain the proportion of various base fragments in all fragments as a third feature set;

The fourth feature set acquisition module is used to compare the read data results to the reference genome and obtain the position of the 5' end of the read on the reference genome; obtain the n bp bases upstream and downstream of the position. Sequence data, as base Base fragment set; the proportion of the obtained various base fragments in all fragments is used as the fourth feature set;

The fifth feature set acquisition module is used to divide the reference genome into multiple windows and obtain copy number data within each window as the fifth feature set;

The model building module is used to use the first, second, third, fourth and fifth feature sets as initial feature values and input them into the classification model as model feature vectors, and use whether there is cancer as the output value to evaluate the model. Perform training to obtain an early screening model.

The third purpose of the present invention:

A computer-readable medium records a computer program that can run a method for constructing a multi-cancer early screening model.

The fourth object of the present invention:

A method for constructing a multi-cancer early screening model, the model is used to distinguish intestinal cancer, lung cancer or liver cancer from cancer samples;

Includes the following steps:

Step 1: Extract and sequence cfDNA from intestinal cancer, lung cancer, and liver cancer samples to obtain read data;

Step 2: Compare the read data results to the reference genome, divide the reference genome into multiple windows, and obtain the number of all reads, the number of short reads, and the number of ultra-long reads within each window, as The first feature set;

Step 3: Align the read data results to the reference genome, using the long arm and short arm of each chromosome as the region range, and obtain the number of reads in gradient intervals of different lengths within each range as the second feature set;

Step 4: Use the m base data at the 5' end in the read data as a base fragment set, and obtain the proportion of various base fragments in all fragments as the third feature set;

Step 5: Compare the read data results to the reference genome to obtain the position of the 5' end of the read on the reference genome; obtain the sequence data of n bp bases upstream and downstream of the position as bases Fragment set; the proportion of the obtained various base fragments in all fragments is used as the fourth feature set;

Step 6: Divide the reference genome into multiple windows and obtain copy number data within each window as the fifth feature set;

Step 7: Establish three control experimental groups respectively. The positive samples in each group are intestinal cancer, lung cancer or liver cancer samples. The control samples in each group are the remaining two cancer samples except the positive samples. In the experimental group, the first, second, third, fourth and fifth feature sets are used together as the initial feature values to screen out the feature values with significant differences between the positive samples and the control samples, and then the three groups are compared The significantly different feature values in the experimental group are combined and input into the classification model as a model feature vector, and the probability of having intestinal cancer, lung cancer or liver cancer is used as the output value to train the model and obtain an early screening model.

In the described step 7, inputting to the classification model means inputting the first, second, third, fourth and fifth feature sets to the gradient boosting algorithm model, random forest model, deep learning model and extreme gradient boosting model respectively. , multiple sub-models are obtained, and the sub-models are combined into a linear relationship model.

The screening process described was through the analysis of variance method.

The short reads are 40-80 bp in length, the number of ultra-long reads is 200-300 bp, and all reads are in the range of 40-300 bp in length.

The window size range in step 2 is 2-7Mb.

The gradient intervals of different lengths in step 3 refer to the gradient ranges of different lengths obtained by increasing in steps of 8-12 bp in the range of 40-300 bp.

The number of reads stated is normalized.

In step 4, m is any integer between 6 and 10.

In step 5, n is any integer between 2 and 5.

The window in step 6 is obtained by dividing the reference gene chromosomes 1-22 with a length of 0.8-1.2Mb without overlap.

The fifth object of the present invention:

A multi-cancer detection device, the device is used to distinguish intestinal cancer, lung cancer or liver cancer from cancer samples, including:

The sequencing module is used to extract and sequence cfDNA from intestinal cancer, lung cancer, and liver cancer samples to obtain read data;

The first feature set acquisition module is used to compare the read data results to the reference genome, divide the reference genome into multiple windows, and obtain the number of all reads, the number of short reads, and the number of super-reads within each window. The number of long reads, as the first feature set;

The second feature set acquisition module is used to compare the read data results to the reference genome, using the long arm and short arm on each chromosome as the regional range, and obtain the reads in different length gradient intervals within each range. The number of segments, as the second feature set;

The third feature set acquisition module is used to use the m base data at the 5' end in the read data as a base fragment set, and obtain the proportion of various base fragments in all fragments as a third feature set;

The fourth feature set acquisition module is used to compare the read data results to the reference genome and obtain the position of the 5' end of the read on the reference genome; obtain the n bp bases upstream and downstream of the position. Sequence data is used as a set of base fragments; the proportion of various base fragments obtained in all fragments is used as the fourth feature set;

The fifth feature set acquisition module is used to divide the reference genome into multiple windows and obtain copy number data within each window as the fifth feature set;

The model building module is used to establish three control experimental groups respectively. The positive samples in each group are intestinal cancer, lung cancer or liver cancer samples. The control samples in each group are the remaining two cancer samples except the positive samples. The first, second, third, fourth and fifth feature sets were used as the initial feature values in the three control experimental groups respectively, and the feature values with significant differences between the positive samples and the control samples were screened out, and then the The significantly different feature values in the three control experimental groups are merged and input into the classification model as a model feature vector, and the probability of suffering from intestinal cancer, lung cancer or liver cancer is used as the output value to train the model and obtain early diagnosis. Sieve model.

The sixth object of the present invention:

A computer-readable medium records a computer program that can run a method for constructing a multi-cancer early screening model.

beneficial effects

The length distribution of low-depth WGS (~5X) cfDNA reads in 191 liver cancer patients, 149 colorectal cancer patients, and 146 lung cancer patients within the genome, length distribution within each long and short arm of the chromosome, proportion of fragment end sequences, and fragmentation The sequence proportion and regional copy number changes at the points were statistically calculated, and five different training learning algorithms were used to build models. type, and perform secondary set training on all models to improve the model's prediction performance for early detection of cancer and cancer type prediction. For the first time, this invention provides a multi-molecule feature multi-training algorithm secondary integrated diagnostic model based on high-throughput and low-depth sequencing of plasma cfDNA. This model can not only diagnose multiple early cancers and their tissue origins, but also has non-invasive detection, low throughput, and detection Characterized by high specificity and sensitivity.

Description of the drawings

Figure 1 is a schematic diagram of the model construction process;

Figure 2 is a schematic diagram of the process of building a multi-cancer early detection model;

Figure 3 is a schematic diagram of the construction process of the tissue origin model of multiple cancer types;

Figure 4 shows the distribution of the largest difference feature among the five features between the cancer group and the non-cancer group;

Figure 5 shows the AUC performance of the multi-cancer early detection model in the training set;

Figure 6 shows the AUC performance of the multi-cancer early detection model in the test set;

Figure 7 shows the distribution of the unique largest difference among the five characteristics of liver cancer between liver cancer and other cancer types;

Figure 8 shows the distribution of the unique largest differential features among the five features of colorectal cancer between colorectal cancer and other cancer types;

Figure 9 shows the distribution of the unique largest difference feature among the five features of lung cancer between lung cancer and other cancer types;

Detailed ways

The present invention relates to early detection of multiple cancer types (lung cancer, intestinal cancer and liver cancer) and cancer type prediction markers, detection methods, detection devices and computer-readable media. The present invention provides a method for WGS low-depth sequencing of plasma sample cfDNA, and uses high-throughput sequencing results to analyze five differential characteristics of cfDNA fragments of various cancer types, including genome-wide fragment length coverage distribution, fragment length distribution on each long and short arm of the chromosome, The sequence at the fragment breakpoint (8-mer Breakpoint Motif), the 5' end sequence of the fragment (8-mer End Motifs) and the 1MB window fragment copy number changes are used to reuse the generalized linear model (Generalized Linear Mode, GLM), gradient boosting machine Five algorithms (Gradient Boosting Machine, GBM), Random Forest (RF), Deep Learning (DL) and Extreme Gradient Boosting (XGBoost) are trained and modeled respectively, and then the generalized linear model (GLM) is used for training Secondary set training builds a multi-feature and multi-algorithm integrated model to achieve the purpose of low-depth, high-specificity and high-sensitivity non-invasive early detection and tissue origin detection of multiple cancer types.

The calculation method in the present invention is detailed as follows:

The present invention first requires steps such as extraction of cfDNA from blood samples, library construction, and sequencing. The extraction and database construction methods here are not particularly limited and can be adjusted from extraction methods in the prior art. During the sequencing process here, existing sequencing technology can be used to obtain the base information of cfDNA.

The data sets used in the model building process in this invention are as follows:

Extraction and sequencing methods for plasma cfDNA samples:

Use purple blood collection tubes (EDTA anticoagulant tubes) to collect 8 ml of whole blood samples from the patient, and centrifuge the plasma promptly (within 2 hours). After being transported to the laboratory, the plasma samples are extracted with cfDNA using the QIAGEN Plasma DNA Extraction Kit according to the instructions. After building a library of the collected cfDNA samples, WGS ~ 5-fold sequencing was performed. After obtaining the offline data, compare the data to the human reference genome to obtain the base data information of the corresponding read segments.

data processing

The marker data in the present invention mainly utilizes five molecular characteristics:

1. DNA fragment size ratio (Fragmentation Size Coverage, FSC)

As for the DNA fragment size ratio, it reflects the length and size ratio characteristics of cfDNA reads. Use DNA fragment size coverage depth (fragmentation size ratio) for machine learning to establish a prediction model to distinguish patients with lung cancer, bowel cancer and liver cancer. By comparing the lengths of cfDNA reads in 486 patients with lung cancer, intestinal cancer or liver cancer, it was found that the number of fragments between 40-80bp, 81-300bp and 40-300bp was distributed differently on the chromosome, which can be used as a distinguishing feature.

The cfDNA read length data is obtained through the following method: in the aligned BAM, the quality, length and alignment position information of each read are recorded. The human reference genome is selected from the University of California, Cruz. , Santa Cruz, UCSC) provided the hg19 sequence. The human reference genome was cut into 572 windows according to the length of 5Mb, and the number of all reads (40-300bp), the number of short reads (40-80bp) and the number of long reads (81-300bp) in each window were counted. . According to the statistical results of the number of various reads in all windows, the number of each read segment is standardized and converted separately, that is, the standardized value = (original value – average) / standard deviation. This resulted in 572 sets of numbers with different lengths of reads.

2. DNA fragment size distribution (Fragmentation Size Distribution, FSD)

On the basis of obtaining the size ratio of DNA fragments, in order to obtain high-resolution reading results, 41 regions on the long and short arms of each chromosome of the human reference genome were used as windows, as shown below:

Divide 40-300bp fragments into 27 length gradients in 10bp increments (for example, 40-49bp, 50-59bp... on the 1q arm of chr1), and count the number of fragments in each long and short arm window of each length gradient. , and perform standardized conversion to obtain high-resolution DNA fragment size distribution results with a total of 1107 feature results (2823=41*27 length gradient normalized results).

3. Proportion of sequence at the 5’ end of the fragment (8-mer End Motif, EDM)

The human reference genome is a DNA double helix structure, linked by hydrogen bonding based on complementary base pairing; during normal aging and cancer progression, the pH of the environment around cells changes, thereby destroying base complementary hydrogen bonds and causing rupture; due to rupture The final DNA fragments have different terminal base sequences, and the proportions of different terminal sequences will also be different. Collection method: After comparison, obtain the 5' end 8bp sequence of each read segment, count the number of reads for each end sequence (a total of 4**8 = 65536 types), and thereby calculate the proportion of 65536 end sequence reads. For example, the proportion of AAAAAAAA sequences = the number of AAAAAAAA reads/the total number of all terminal sequence reads.

4. Proportion of the number of sequence reads at the 5’ end of the fragment (8-mer Breakpoint Motif, BKM)

Similar to the proportion of terminal sequences, due to the different base sequences at the break, the proportion of sequences containing different breakpoint sequences will also be different. Collection method: In the aligned BAM, the basic information of each read segment and the aligned position are recorded. Confirm that the 5' end of each read segment is located at the left and right 4bp sequences of the human reference genome sequence coordinates, and count each type. The number of reads of sequences at breakpoints (a total of 4**8 = 65536 types) is used to calculate the proportion of reads of 65536 types of sequences at breakpoints, for example, the proportion of AAAAAAAA sequences = the number of reads of AAAAAAAA/all breakpoint sequence reads total.

5.1 Mb window copy number variation (1Mb-Bin Copy Number Variation, CNV)

Copy number changes are highly correlated with individual cancers. Although differentiation can be made by detecting copy number changes in some cancer-related genes or specific genomic regions, there are other rare or unknown genes or regions that can provide potential copy number changes. information. Collection method: For the WGS data of each sample to be tested, divide the reference gene chromosomes 1-22 into windows with a length of 1Mb without overlap, use bedtools coverage to calculate the read depth in each window for each sample, and calculate the read depth in each window according to the length of each window. GC content and average comparison ability records (UCSC BigWig files) were corrected to obtain 2475 window individual read depth information. Hidden Markov Model (HMM) was used to compare the baseline depth with each window group to construct each The logarithm of the copy number change in each window, that is, log2 (the depth after correction and normalization of the sample to be tested/the depth after normalization of the population baseline correction), is used to obtain the copy number change information of each sample to be tested.

Through the above data acquisition, the initial data vectors of these five types of data can be obtained respectively. Next, the corresponding calculation method is designed: the landmark data in this invention mainly uses five single-feature machine learning algorithms:

1. Generalize Linear Model (GLM)

The generalized linear model is an extension of the linear model that establishes the relationship between the mathematical expectation value of the response variable and the linear combination of predictor variables through a link function. The main feature is that it does not forcibly change the natural measurement of the data, and it is a commonly used two-class classification strategy.

2. Gradient Boosting Machine (GBM)

The gradient boosting algorithm is a common type of algorithm in machine learning. Its basic principle is to train a newly added weak classifier based on the negative gradient information of the current model loss function, and then combine the trained weak classifiers in an additive form to the current classifier. In some models, the optimal model is obtained. This model has the advantages of good training effect and is not easy to overfit. In order to prevent GBM from over- or under-fitting during the learning process, set the GBM parameters as follows: ntrees=300, max_depth=9, learning_rate=0.01, sample=0.8, cross_validation=10.

3. Random Forest (RF)

Random forest is a powerful classification and regression tool. When a set of data is provided, the random forest can randomly extract part of the information to generate a set of decision trees to help classification or regression, make node split attributes, and repeat the random extraction until it can no longer be split; finally, combine all split attribute results to obtain the final prediction result. In order to prevent RF from over- or under-fitting during the learning process, set the RF parameters as follows: ntrees=300, max_depth=9, cross_validation=10.

4. Deep Learning Algorithm (Deep Learning, DL)

Deep learning is based on multi-layer feedforward artificial neural networks trained with stochastic gradient descent using backpropagation. The network can contain a large number of hidden layers consisting of neurons with hyperbolic tangent, rectification and maximum power activation functions. Advanced features such as adaptive learning rate, rate annealing, momentum training, dropout, L1 or L2 regularization, checkpointing and grid search enable high prediction accuracy. During learning training, each compute node uses multi-threading (asynchronously) to train a copy of the global model parameters on its local data and regularly contributes to the global model through model averaging across the network. Feedforward artificial neural network (ANN) models, also known as deep neural networks (DNN) or multilayer perceptrons (MLP), are the most common type of deep neural networks. The main principle is to design multiple perceptrons with multiple inputs and multiple outputs and establish an appropriate number of neuron computing nodes and multi-layer computing hierarchies, select appropriate input layers and output layers, and establish a network through network learning and tuning. The functional relationship from input to output can approximate the realistic correlation relationship as much as possible. In order to prevent over- or under-fitting during the DL learning process, the DL parameters are set as follows: epoch=300, hidden={100,100,100}, input_dropout_ratios=0.05, rho=0.95, mini_batch_size=10, cross_validation=10.

5. Extreme Gradient Boosting algorithm (Extreme Gradient Boosting, XGBoost)

Extreme Gradient Boosting is an efficient open source implementation of the gradient boosting algorithm. Compared with traditional GMB, XGBoost introduces parallelization, so it is faster; XGBoost introduces a second-order approximation to the objective function, obtains an analytical solution, and uses the analytical solution to build a decision tree to optimize the objective function; The term part can control the complexity of the model and prevent overfitting; Xgboost introduces feature subsampling, which is similar to random forest, which can reduce overfitting and calculation. In order to prevent XGBoost from over- or under-fitting during the learning process, set the XGBoost parameters as follows: ntrees=300, max_depth=9, cross_validation=10.

In order to improve the efficiency of machine learning and reduce the interference of useless features, analysis of variance (ANOVA) is performed on the feature values between different groups to filter out the feature values with large differences between the groups. This step is done through the aov() function of the R package stats The F value (Fvalue) result and the corrected pvalue implementation of the p.adjust() function.

The establishment process of multi-cancer early detection model

In order to establish a multi-cancer early detection model, the training set was divided into cancer groups and non-cancer groups, variance analysis was performed on the five features respectively, and sorted in descending order according to the Fvalue results, and the top 200 features were retained as prediction input values.

The establishment process of cancer tissue origin model

In order to establish a multi-cancer tissue origin model, 191 cases of liver cancer, 149 cases of intestinal cancer and 146 cases of lung cancer in the training set were divided into three groups, and each cancer group was analyzed in the "single cancer type vs. other cancer types" mode. Conduct feature variance analysis on each of the five features of a single cancer type in descending order according to the Fvalue results, and retain the top 100 features as prediction input values.

Data results from the eigenvalue filtering process

The top 200 significantly different feature values of FSC between the cancer group and the non-cancer group are as follows: long/short/total represent long reads, short reads and all reads respectively, and the numbers represent the window position number;

The first 200 significantly different eigenvalues of FSD between the cancer group and the non-cancer group are as follows: chrx, xp/q among them represent the long arm or short arm on chromosome x, and the number part refers to the number of the gradient position;

The top 200 significantly different eigenvalues of EDM between the cancer group and the non-cancer group are as follows: the numbers consisting of 8-digit ATCG represent the base sequences of different eigenvalues;

The top 200 significantly different eigenvalues of BKM between the cancer group and the non-cancer group are as follows: the numbers consisting of 8-bit ATCG represent the base sequences of different eigenvalues;

The top 200 significantly different eigenvalues of CNV between the cancer group and the non-cancer group are as follows: chrx represents chromosome x, and the number part refers to the range of positions on the chromosome;

The top 100 characteristic values of liver cancer FSC that are significantly different from other cancer types are as follows:

The top 100 characteristic values of liver cancer FSD that are significantly different from other cancer types are as follows:

The top 100 characteristic values of liver cancer EDM that are significantly different from other cancer types are as follows:

The top 100 characteristic values of liver cancer BPM that are significantly different from other cancer types are as follows:

The top 100 characteristic values of liver cancer CNVs that are significantly different from other cancer types are as follows:

The top 100 characteristic values of intestinal cancer FSC that are significantly different from other cancer types are as follows:

The top 100 characteristic values of bowel cancer FSD that are significantly different from other cancer types are as follows:

The top 100 characteristic values of bowel cancer EDM that are significantly different from other cancer types are as follows:

The top 100 characteristic values of bowel cancer BPM that are significantly different from other cancer types are as follows:

The top 100 characteristic values of intestinal cancer CNVs that are significantly different from other cancer types are as follows:

The top 100 characteristic values of lung cancer FSC that are significantly different from other cancer types are as follows:

The top 100 characteristic values of lung cancer FSD that are significantly different from other cancer types are as follows:

The top 100 characteristic values of lung cancer EDM that are significantly different from other cancer types are as follows:

The top 100 characteristic values of lung cancer BPM that are significantly different from other cancer types are as follows:

The top 100 characteristic values of lung cancer CNVs that are significantly different from other cancer types are as follows:

After screening the different significant features, 200 features of each of the five types were obtained for the multi-cancer early detection model. Each feature of the samples in all training sets was used as the input value, and the prediction of "cancer/health" was used as the feedback result. The generalized model was used respectively. Linear models, gradient boosting algorithm models, random forest models, deep learning models and extreme gradient boosting models were trained and modeled, and 25 two-class basic models were obtained;

In order to further improve the prediction performance of the classifier, secondary set training (stacking) is performed on the above multiple training basic model results. Stacking is an integrated learning technology that performs meta-learning ( ^2nd -level meta-learning) on multiple underlying weak classifiers ( ^1st -level base model) to collect the characteristics of each underlying classifier and find the best Optimize the integration method to improve the model prediction performance. The training algorithm used in Stacking in this patent is the Generalized Linear Model (GLM). It establishes the relationship between the mathematical expectation value of the response variable and the linear combination of predictor variables through the link function, and converts various training basic models into the final linear equation. :

ALLStacked＝Intercept+A1*FSC_GLM+A2*FSC_GBM+A3*FSC_RF+A4*FSC_DL+A5*FSC_XGBoost+B1*F SD_GLM+B2*FSD_GBM+B3*FSD_RF+B4*FSD_DL+B5*FSD_XGBoost+C1*EDM_GLM+C2 *EDM_GBM+C3*EDM_RF+C4*EDM_DL+C5*EDM_XGBoost+D1*BPM_GLM+D2*BPM_GBM+D3*BPM_RF+D4*BPM_DL+D5*BPM_XGBoost+E1*CNV_GLM+E2*CNV_GBM+E3*CNV_RF+E4*CNV_DL +E5*CNV_XGBoost

Among them, Intercept and A1-E5 are all linear equation parameters. FSC_GLM, etc. all refer to the output value obtained by the model after obtaining the input data. The characters before the symbol "_" represent the type of feature set, and the characters after the symbol "_" represent the algorithm type. The output value of the multi-cancer early screening model is Cancer probability.

The model of tissue origin of multiple cancer types is mainly used to further confirm the specific cancer type for samples that have been confirmed to have one of the above three cancers. Therefore, when classifying samples, three groups of training samples are established:

The first set of training samples: positive is intestinal cancer, and the control is lung cancer and liver cancer; the judgment is divided into two categories: intestinal cancer and other two cancers.

The second set of training samples: positive is lung cancer, and the control is intestinal cancer and liver cancer; the judgment is divided into two categories: lung cancer and other two cancers.

The third group of training samples: positive is liver cancer, and the control is liver cancer and intestinal cancer; the judgment is divided into two categories: liver cancer and other two types of cancer.

In each group of samples, variance analysis is performed separately, and the characteristic values with significant differences in each feature set can be found in each group; and after the analysis of all three groups is completed, corresponding values can be obtained between each group. There will be overlap among the eigenvalues with significant differences. Therefore, each group of filtered eigenvalues will be merged and repeated to obtain the eigenvalues required in the final model.

Finally, 180 FSC features, 205 FSD features, 295 EDM features, 297 BKM features, and 204 CNV features were obtained for the multi-cancer tissue origin model. Each set of features of the cancer samples in the training connection set is used as the input value, and the prediction of "intestinal cancer/liver cancer/lung cancer" is used as the feedback result. The gradient boosting algorithm model, random forest model, deep learning model and The extreme gradient boosting model is trained and modeled, and 20 multi-classification basic models are obtained.

In order to improve the prediction performance, secondary set training is also used. The method is basically the same as the above process. The difference is that the linear equation used is:

ALLStacked＝Intercept+A2*FSC_GBM+A3*FSC_RF+A4*FSC_DL+A5*FSC_XGBoost++B2*FSD_GBM+B3*FSD_RF+B4*FSD_DL+B5*FSD_XGBoost+C2*EDM_GBM+C3*EDM_RF+C4*EDM_DL+C5 *EDM_XGBoost++D2*BPM_GBM+D3*BPM_RF+D4*BPM_DL+D5*BPM_XGBoost++E2*CNV_GBM+E3*CNV_RF+E4*CNV_DL+E5*CNV_XGBoost

Among them, Intercept and A2-E5 are all linear equation parameters. FSC_GBM, etc. all refer to the output value obtained by the model after obtaining the input data. The characters before the symbol "_" represent the type of feature set, and the characters after the symbol "_" represent the algorithm type. The output value of the multi-cancer early screening model is Cancer probability, the multi-cancer tissue origin model is the cancer type probability (the multi-cancer tissue origin integrated model will predict the possibility of liver cancer, bowel cancer and lung cancer for the samples to be predicted, and use the three prediction results as the The maximum value is taken as the final judgment result).

The integrated model for early detection of multiple cancer types can effectively distinguish cancer from healthy people. The sensitivity and specificity in the training set both reached 94%. At the same time, the test set verified the integrated model, and the sensitivity and specificity reached 95%. There were no inter-set results. Differences, the specific results are shown in the table below:

The multi-cancer tissue origin set model can effectively distinguish the tissue origins of lung cancer, liver cancer and intestinal cancer. The overall accuracy rate in the training set reached 95.1%, and the overall accuracy of the samples that successfully predicted cancer in the test set was 93.1%. Specifically The results are shown in the table below:

Control experiment 1:

Among the eigenvalues used in the model, the 5' end sequence proportion of the fragment (EDM) is not included, and only the other four are used. The model establishment process is the same as above. The cancer species origin model is established, and the final calculation results of the test set samples are obtained. as follows:

GLM is a two-classification algorithm. Its advantages are not obvious when performing multiple classifications. It cannot show good classification performance in the process of cancer classification. Therefore, the basic model of glm is not used in the classification model of this part. It is only used in cancer classification. /Used in the process of classifying healthy samples.

The technical solution of this patent is explained and described through the above embodiments, but does not constitute a limitation on the protection scope of this patent.

Claims (10)

1. A method for constructing a multi-cancer early screening model. The model is used to classify whether a sample has intestinal cancer, lung cancer or liver cancer. It is characterized by including the following steps:

   Step 1: Extract and sequence cfDNA from the samples of the positive group and the control group to obtain read data;

   Step 2: Compare the read data results to the reference genome, divide the reference genome into multiple windows, and obtain the number of all reads, the number of short reads, and the number of ultra-long reads within each window, as The first feature set;

   Step 3: Align the read data results to the reference genome, using the long arm and short arm of each chromosome as the region range, and obtain the number of reads in gradient intervals of different lengths within each range as the second feature set;

   Step 4: Use the m base data at the 5' end in the read data as a base fragment set, and obtain the proportion of various base fragments in all fragments as the third feature set;

   Step 5: Compare the read data results to the reference genome to obtain the position of the 5' end of the read on the reference genome; obtain the sequence data of n bp bases upstream and downstream of the position as bases Fragment set; the proportion of the obtained various base fragments in all fragments is used as the fourth feature set;

   Step 6: Divide the reference genome into multiple windows and obtain copy number data within each window as the fifth feature set;

   Step 7: Use the first, second, third, fourth and fifth feature sets together as the initial feature value and input it into the classification model as the model feature vector, and use whether you have cancer as the output value to train the model. Get early screening models.
2. The method for constructing a multi-cancer early screening model according to claim 1, characterized in that in the step 6, suffering from cancer refers to suffering from any one of intestinal cancer, lung cancer or liver cancer; the step In 6, it is also necessary to simplify the initial feature values and then use them as model feature vectors. The simplification means filtering out the first, second, third, fourth and fifth feature sets in the positive group and the control group respectively. There are eigenvalues with significant differences between samples; the screening process is through the analysis of variance method
3. The method for constructing a multi-cancer early screening model according to claim 1, wherein the short reads are 40-80 bp in length, and the number of ultra-long reads is 200-300 bp; all reads It means that the length is in the range of 40-300bp; the size range of the window in step 2 is 2-7Mb.
4. The method for constructing a multi-cancer early screening model according to claim 1, wherein the different length gradient intervals in step 3 refer to different gradient intervals obtained by increasing in steps of 8-12 bp in the range of 40-300 bp. Length gradient range; number of reads stated is normalized.
5. The method for constructing a multi-cancer early screening model according to claim 1, characterized in that, in the step 4, m is any integer between 6 and 10; and in the step 5, n is 2-10. Any integer between 5;
6. The method for constructing a multi-cancer early screening model according to claim 1, characterized in that the window in step 6 is obtained by dividing the reference gene chromosomes 1-22 with a length of 0.8-1.2Mb without overlap. ; The input to the classification model in step 7 refers to inputting the first, second, third, fourth and fifth feature sets to the generalized linear model, gradient boosting algorithm model, random forest model, deep learning model and In the extreme gradient boosting model, multiple sub-models are obtained and connected into a linear relationship model.
7. A multi-cancer detection device is characterized in that the device is used to classify whether a sample has intestinal cancer, lung cancer or liver cancer, including:

   The sequencing module is used to extract and sequence cfDNA from the samples of the positive group and the control group to obtain read data;

   The first feature set acquisition module is used to compare the read data results to the reference genome, divide the reference genome into multiple windows, and obtain the number of all reads, the number of short reads, and the number of super-reads within each window. Number of long reads, As the first feature set;

   The second feature set acquisition module is used to compare the read data results to the reference genome, using the long arm and short arm on each chromosome as the regional range, and obtain the reads in different length gradient intervals within each range. The number of segments, as the second feature set;

   The third feature set acquisition module is used to use the m base data at the 5' end in the read data as a base fragment set, and obtain the proportion of various base fragments in all fragments as a third feature set;

   The fourth feature set acquisition module is used to compare the read data results to the reference genome and obtain the position of the 5' end of the read on the reference genome; obtain the n bp bases upstream and downstream of the position. Sequence data is used as a set of base fragments; the proportion of various base fragments obtained in all fragments is used as the fourth feature set;

   The fifth feature set acquisition module is used to divide the reference genome into multiple windows and obtain copy number data within each window as the fifth feature set;

   The model building module is used to use the first, second, third, fourth and fifth feature sets as initial feature values and input them into the classification model as model feature vectors, and use whether there is cancer as the output value to evaluate the model. Perform training to obtain an early screening model.
8. A method for constructing a multi-cancer early screening model, characterized in that the model is used to distinguish intestinal cancer, lung cancer or liver cancer from cancer samples;

   Step 1: Extract and sequence cfDNA from intestinal cancer, lung cancer, and liver cancer samples to obtain read data;

   Step 2: Compare the read data results to the reference genome, divide the reference genome into multiple windows, and obtain the number of all reads, the number of short reads, and the number of ultra-long reads within each window, as The first feature set;

   Step 3: Align the read data results to the reference genome, using the long arm and short arm of each chromosome as the region range, and obtain the number of reads in gradient intervals of different lengths within each range as the second feature set;

   Step 4: Use the m base data at the 5' end in the read data as a base fragment set, and obtain the proportion of various base fragments in all fragments as the third feature set;

   Step 5: Compare the read data results to the reference genome to obtain the position of the 5' end of the read on the reference genome; obtain the sequence data of n bp bases upstream and downstream of the position as bases Fragment set; the proportion of the obtained various base fragments in all fragments is used as the fourth feature set;

   Step 6: Divide the reference genome into multiple windows and obtain copy number data within each window as the fifth feature set;

   Step 7: Establish three control experimental groups respectively. The positive samples in each group are intestinal cancer, lung cancer or liver cancer samples. The control samples in each group are the remaining two cancer samples except the positive samples. In the experimental group, the first, second, third, fourth and fifth feature sets are used together as the initial feature values to screen out the feature values with significant differences between the positive samples and the control samples, and then the three groups are compared The significantly different feature values in the experimental group are combined and input into the classification model as a model feature vector, and the probability of having intestinal cancer, lung cancer or liver cancer is used as the output value to train the model and obtain an early screening model.
9. The method for constructing a multi-cancer early screening model according to claim 8, wherein in step 7, inputting into the classification model means to respectively add the first, second, third, fourth and fifth The feature set is input into the gradient boosting algorithm model, random forest model, deep learning model and extreme gradient boosting model, multiple sub-models are obtained, and the sub-models are connected into a linear relationship model.
10. A multi-cancer detection device is characterized in that the device is used to detect intestinal cancer, lung cancer or liver cancer samples on cancer samples. The distinction between cancers includes:

    The sequencing module is used to extract and sequence cfDNA from intestinal cancer, lung cancer, and liver cancer samples to obtain read data;

    The first feature set acquisition module is used to compare the read data results to the reference genome, divide the reference genome into multiple windows, and obtain the number of all reads, the number of short reads, and the number of super-reads within each window. The number of long reads, as the first feature set;

    The second feature set acquisition module is used to compare the read data results to the reference genome, using the long arm and short arm on each chromosome as the regional range, and obtain the reads in different length gradient intervals within each range. The number of segments, as the second feature set;

    The third feature set acquisition module is used to use the m base data at the 5' end in the read data as a base fragment set, and obtain the proportion of various base fragments in all fragments as a third feature set;

    The fourth feature set acquisition module is used to compare the read data results to the reference genome and obtain the position of the 5' end of the read on the reference genome; obtain the n bp bases upstream and downstream of the position. Sequence data is used as a set of base fragments; the proportion of various base fragments obtained in all fragments is used as the fourth feature set;

    The fifth feature set acquisition module is used to divide the reference genome into multiple windows and obtain copy number data within each window as the fifth feature set;

    The model building module is used to establish three control experimental groups respectively. The positive samples in each group are intestinal cancer, lung cancer or liver cancer samples. The control samples in each group are the remaining two cancer samples except the positive samples. The first, second, third, fourth and fifth feature sets were used as the initial feature values in the three control experimental groups respectively, and the feature values with significant differences between the positive samples and the control samples were screened out, and then the The significantly different feature values in the three control experimental groups are merged and input into the classification model as a model feature vector, and the probability of suffering from intestinal cancer, lung cancer or liver cancer is used as the output value to train the model and obtain early diagnosis. Sieve model.