Lung cancer early screening model and its application

Abstract

The present application relates to the field of bioinformatics, and specifically provides a lung cancer early screening model construction and application based on FDEM / FDBM feature extraction from LP-WGS sequencing data and PCC system noise reduction. By performing low-depth whole-genome sequencing of plasma sample cfDNA, high-throughput sequencing of lung cancer and healthy human cfDNA, extracting new features from the sequencing results, and using the PCC (Principal Component Correction) method to reduce background noise of the new features to improve the signal-to-noise ratio of cancer signals, a highly sensitive machine learning model for distinguishing lung cancer from healthy people is constructed, and ultimately achieving high-precision identification of lung cancer patients.

Description

Technical Field

[0001] The present application relates to the field of bioinformatics analysis, especially DNA identification technology, and specifically to the construction of an early screening model based on LP-WGS sequencing data to extract FDEM / FDBM features and perform PCC system noise reduction, and its application. Background Art

[0002] Currently, the main screening method for lung cancer is low-dose computed tomography (LDCT), but LDCT has a high false positive rate (about 90%), and there are risks of radiation exposure and overdiagnosis, resulting in poor compliance with LDCT. Only less than 6% of high-risk people receive lung cancer screening, which has not yet met the current clinical needs. Therefore, it is urgent to develop an effective, practical and economical lung cancer screening method suitable for high-risk people.

[0003] Low-depth second-generation sequencing has the advantages of wide detection range and low cost. It is widely used in research related to early screening of lung cancer. Some fragment group-related features have also been proposed for screening of early lung cancer patients, including the ratio of short fragments to long fragments (Ratio), fragment motifs (EDM&BPM), open chromatin state (Chromatin State), etc.; in the research related to the fragment group, it was found that the ratio of short fragments to long fragments (Ratio) feature in stage I lung cancer had an AUC of only 0.76 (95% CI: 0.64-0.88) [PMID: 34417454], and the AUC of stage I lung cancer in the EDM4 feature was 0.905 (95% CI: 0.836-.974) and the sensitivity was 0.575 (95% CI: 0.409-0.73) [PMID: 35780566], which means that the detection sensitivity in stage I lung cancer is relatively low. In addition, during the high-throughput extraction, library construction and sequencing process, factors such as experimental procedures, experimental platforms and differences in experimental personnel will introduce a series of noises, especially for low-depth sequencing data. Therefore, the current technology still needs a new method to reduce background noise from low-depth second-generation sequencing results and improve the signal-to-noise ratio of cancer signals, thereby providing a more sensitive and cost-effective analysis method for non-invasive diagnosis and screening of lung cancer, and improving the accuracy, compliance and accessibility of lung cancer screening.

[0004] In view of this, this application is hereby filed. Summary of the invention

[0005] To solve the above problems, this application collects whole-genome low-depth sequencing data of cfDNA from lung cancer and healthy human plasma samples, extracts new features (FDEM&FDBM) from the sequencing results, and uses the PCC method to reduce the background noise of the new features to improve the signal-to-noise ratio of cancer signals. We further construct a highly sensitive machine learning model for distinguishing between lung cancer and healthy people, and ultimately achieve high-precision identification of lung cancer patients.

[0006] Specifically, the technical solution of this application is as follows:

[0007] This application first provides a method for constructing a lung cancer early screening model based on extracting FDEM / FDBM features from low-depth sequencing data and performing PCC system noise reduction, comprising the following steps:

[0008] 1. Low-depth whole genome sequencing data acquisition and preprocessing;

[0009] 2. Extract the cfDNA fragment group features FDEM and FDBM; the FDEM is a 4 bp end motif pattern based on the fragment size, and the FDBM is a 4 bp breakpoint motif pattern based on the fragment size;

[0010] 3. Perform background noise reduction on the feature matrix based on principal component analysis (PCC) method;

[0011] 4. Construction of early screening model.

[0012] Furthermore, the acquisition in step 1 is to download data files containing fragment information of lung cancer and healthy samples from the public database FinaleDB; preferably, the fragment information includes chromosome number, starting position, ending position, comparison quality value and positive and negative chains.

[0013] Furthermore, the preprocessing in step 1 includes: for each sample, fragment filtering is performed based on the alignment quality value and fragment length, and fragments with an alignment quality value less than 30 and a fragment length greater than 400 bp are removed; reads located in the hg19 genome blacklist region, spacer region, patch sequence, highly variable region and centromere-telomeric region in the fragment are removed using public database information; test samples with GC bias greater than 0.3 are filtered; and the obtained fragment data are GC corrected.

[0014] Further preferably, the detailed steps of the pretreatment in step 1 include:

[0015] 1) For each sample, firstly, the fragments were filtered based on the alignment quality value and fragment length, and the fragments with alignment quality value less than 30 and fragment length greater than 400 bp were removed;

[0016] 2) Using public database information, remove the fragments obtained in step 1 that are located in the blacklist region, spacer region, patch sequence, highly variable region, and centromere-telomeric region of the hg19 genome;

[0017] 3) Filter the test samples with GC bias greater than 0.3. The GC bias calculation steps are as follows:

[0018] (1) Split the reference genome hg19 into 100 kb windows and count the GC content in each window;

[0019] (2) For each test sample, the sequencing depth R of each window is counted separately;

[0020] (3) Calculate the sample level GC bias value S GC , the calculation formula is as follows:

[0021]

[0022]

[0023] Where N is the total number of windows, R i represents the sequencing depth of the i-th window, M represents the average sequencing depth of all windows, and R loess Indicates that the depth value is predicted by the loess model.

[0024] 4) performing GC correction on the fragment data obtained in step 2), wherein the GC correction specifically comprises the following steps:

[0025] (1) Construct a baseline, select multiple healthy individuals, and for each healthy sample, select fragments with a length between 50 and 400 bp in the files obtained in the above steps 1) to 3) and calculate the GC content of each fragment; stratify the GC content according to the interval of 1%, and count the number of fragments N corresponding to the GC content of each stratification under each chromosome. gc ;

[0026] (2) After obtaining the fragment counts of different chromosomes with different GC contents of multiple healthy samples through (1) above, the mean of each distribution is further calculated, i.e., the GC bias-corrected reference distribution M gc .

[0027]

[0028] Where n represents the number of healthy samples selected to construct the reference distribution, i represents the i-th healthy sample, and M gc Indicates the reference distribution value corresponding to a certain gc content, N gc,i It represents the number of fragments corresponding to the i-th healthy sample at a certain gc content;

[0029] (3) For the actual test sample, the file obtained after processing in steps 1)-3) contains the chromosome number, start position, end position, fragment length, and GC content information of each fragment in the hg19 human reference genome. The number of fragments N corresponding to each GC content of each test sample is counted according to method (1). gc,i , and the reference distribution M is corrected according to the GC bias of the healthy samples obtained in (2) above gc The weight value under each GC content is calculated; the weight value is used to weight each fragment according to its GC content, thus completing GC correction.

[0030] Furthermore, the extraction in step 2 comprises the following steps:

[0031] 1) For each fragment of each test sample, only the fragments with a length in the range of [100, 250] bp are retained, and the fragments are layered to obtain the fragment clusters of each layer;

[0032] Preferably, the stratification is to stratify the fragments with a span of 10 bp: 100-109 bp, 110-119 bp, ..., 240-249 bp, to obtain the fragment clusters of each layer:

[0033] F={i|100<=f i <=250},

[0034] F k ={i|i∈F,100+10k<=f i <100+10(k+1)},k∈{0,1,2,...,14}

[0035] Where i represents the i-th fragment, f i represents the length of the i-th fragment, F represents the set of all fragments whose lengths are between 100 and 250 bp, and F k Represents the set of fragments under each fragment length layer;

[0036] 2) Extract base sequences for each fragment: for FDEM features, extract the 4 bases upstream and downstream of the reference genome position corresponding to the 5' end of the two chains; for FDBM features, extract the 4 bases ±2bp of the reference genome position corresponding to the 5' end of the two chains, that is, extend 2bp upstream and downstream from the start position and the end position respectively;

[0037] Preferably, the following possible combinations are obtained:

[0038] M∈{AAAA,AAAT,...,TTTA,TTTT,...CCCA,....GGGA},len(M)=256

[0039] Among them, M represents the possible combinations of sequences, a total of 256 combinations;

[0040] 3) Count the GC-corrected frequency information of each motif in each fragment cluster, that is, the cumulative sum of the fragment weight values;

[0041] Preferably, the cumulative sum accounts for f m,k , the calculation formula is as follows;

[0042]

[0043] Among them, N m,k represents the number of fragments of a certain motif under a certain fragment length k, N k Represents the total number of fragments under a certain fragment length k.

[0044] Furthermore, the background noise reduction in step 3 includes the following steps:

[0045] 1) Randomly select multiple healthy samples from the training set samples to obtain the FDEM / FDBM feature matrix H of the multiple healthy samples;

[0046] 2) Perform singular value decomposition SVD on the matrix H, H = UDV T , where U and V are both orthogonal matrices; D is a diagonal matrix with non-negative singular values ​​on the diagonal; the column vectors of U are called left singular vectors, and the column vectors of V are called right singular vectors. T represents the transposed matrix of matrix V; the diagonal elements of D are called singular values ​​and are arranged in descending order; in SVD decomposition, singular values ​​represent the importance of the original matrix, and the features represented by the left and right singular vectors with larger singular values ​​contribute more to the data; when the cumulative sum of k singular values ​​is above 90%, V T The matrix X composed of the first k column vectors of the matrix can be expressed as the main features of the matrix H, that is, the background noise that needs to be corrected;

[0047] 3) Subtracting the orthogonal projection of each test sample in the X subspace to correct the background noise; the background noise includes healthy signals, batch signals, and GC bias noise;

[0048] Preferably, the calculation formula for the correction is as follows:

[0049]

[0050] Among them, F i ′ represents the eigenvector of sample i after principal component correction, F irepresents the original FDEM eigenvector of sample i, j represents the jth principal component, and dot represents the vector dot product.

[0051] Furthermore, the construction in step 4 is:

[0052] The feature data in the training set that has undergone PCC denoising is subjected to PCA dimensionality reduction, and the features that satisfy the explanatory power of more than 90% are retained for model construction; the model is trained on the features after dimensionality reduction, and 5-fold cross-validation is selected to perform grid parameter adjustment on the model. The parameters corresponding to the highest AUC are used as the final model parameters to obtain a single feature prediction model; the predicted values ​​obtained from the single feature training are merged as the input matrix of the integrated model, and the logistic regression model is used for training to obtain the final integrated model, and the final prediction score value of the sample is obtained.

[0053] Further preferably, the detailed steps of model construction in step 4 are as follows:

[0054] 1) Feature dimensionality reduction: The preProcess module in the caret package of the R language was used to perform PCA dimensionality reduction on the feature data in the training set, and only features with an explanatory power of more than 90% were retained for model construction.

[0055] 2) Model training: Use the glmnet model in the caret package of R language to train the features after dimensionality reduction in 1). Select the 5-fold cross-validation method (method=repeatedcv) repeated 5 times to adjust the model parameters. The parameters include alpha=seq(0,1,by=0.1) and lambda=10^seq(-5,0,length.out=100)). The parameters corresponding to the highest AUC are used as the final parameters of the glmnet model to obtain a single feature prediction model.

[0056] 3) Model integration: The predicted values ​​obtained from each single feature training in step 2) are merged as the input matrix of the integrated model, and then trained through the logistic regression model (family = binomial) in the glmnet package of the R language to obtain the final integrated model and the final predicted score value of the sample.

[0057] This application also provides a lung cancer early screening model, including the following modules:

[0058] 1) Data acquisition module: used for acquisition and preprocessing of low-depth whole genome sequencing data;

[0059] 2) Feature extraction module, used to extract cfDNA fragment group features FDEM and FDBM, wherein the FDEM is a 4 bp end motif pattern based on fragment size, and the FDBM is a 4 bp breakpoint motif pattern based on fragment size;

[0060] 3) Background noise reduction module: used to perform background noise reduction on the feature matrix based on the principal component analysis (PCC) method;

[0061] 4) Model building module: used for building early screening models.

[0062] Furthermore, the features in each module are defined with reference to the features in the above-mentioned model building method.

[0063] The present application also provides the application of the above-mentioned early screening model in early screening of lung cancer.

[0064] The present application also provides a computer storage medium, wherein the computer storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, any of the above methods is implemented.

[0065] The present application also provides an electronic device, which includes a memory and a processor, wherein the memory and the processor are connected, the memory is used to store a computer program, and the processor is used to call the computer program to implement any of the above methods.

[0066] This application includes two methods for obtaining cfDNA fragment features in the extraction of cfDNA fragment features, and obtains principal components with an explanatory degree of more than 90% through randomly selected healthy samples, and then subtracts the orthogonal projections in the direction of these principal components from the test samples in turn to correct the feature data of the test samples, thereby improving the signal-to-noise ratio of the cancer signal, and accurately classifying lung cancer samples and healthy samples by constructing a machine learning model. The model of this application can distinguish lung cancer from healthy people with high sensitivity, and ultimately achieve high-precision identification of lung cancer patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the specific implementation methods of the present application or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0068] Figure 1 , the technical roadmap of this application;

[0069] Figure 2 , schematic diagram of new feature extraction;

[0070] Figure 3 , ROC curve of the model in the training set;

[0071] Figure 4 , ROC curve of the model on the test set. DETAILED DESCRIPTION

[0072] The embodiments of the present application will be described in detail below in conjunction with the examples, but those skilled in the art will understand that the following examples are only used to illustrate the present application and should not be regarded as limiting the scope of the present application, and the examples are part of the present application, rather than all of the examples. Based on the examples in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0073] Definition of some terms

[0074] Unless otherwise defined below, the meaning of all technical terms and scientific terms used in the specific embodiments of the present application is intended to be the same as those generally understood by those skilled in the art. Although it is believed that the following terms are well understood by those skilled in the art, the following definitions are still set forth to better explain the present application.

[0075] As used in this application, the terms "comprises", "comprising", "having", "containing" or "involving" are inclusive or open-ended and do not exclude other unrecited elements or method steps. The term "consisting of" is considered a preferred embodiment of the term "comprising". If a group is defined below as comprising at least a certain number of embodiments, this should also be understood to disclose a group that preferably consists of only these embodiments.

[0076] The terms "approximately" and "substantially" in this application represent the accuracy range that can be understood by those skilled in the art to still ensure the technical effect of the feature in question. The term usually represents ±10% deviation from the indicated value, preferably ±5%.

[0077] When referring to a singular noun an indefinite or definite article e.g. "a" or "an", "the" is used, this includes a plural of that noun.

[0078] In addition, the terms first, second, third, (a), (b), (c), and the like in the specification and claims are used to distinguish similar elements and are not necessarily required to describe a sequential or chronological order. It should be understood that the terms so used are interchangeable under appropriate circumstances, and the embodiments described in this application can be implemented in other sequences than those described or illustrated in this application.

[0079] The beneficial effects of the present application will be further described below in conjunction with specific examples. It should be noted that the following examples are only exemplary descriptions and do not limit the method of the present application to only the following methods.

[0080] Example 1: Construction of the method system of this application

[0081] 1. Obtaining public data of low-depth whole genomes

[0082] Data files containing fragment information (including chromosome number, start position, end position, alignment quality value, and positive and negative chains) of lung cancer and healthy samples were downloaded from the public database FinaleDB (FinaleDB: comprehensive database of cell-free DNA fragmentation patterns (cchmc.org)). The lung cancer and healthy samples included in this database are mainly from the studies of Jiang et al., 2015 and Cristiano et al., 2019, including 268 healthy samples and 79 lung cancer patients.

[0083] 2. Low-depth whole genome sequencing data preprocessing

[0084] 1. For each sample, firstly, the fragments were filtered based on the alignment quality value and the fragment length, and the fragments with alignment quality value less than 30 and fragment length (end position - start position) greater than 400 bp were removed;

[0085] 2. Using public database information (database source: https: / / genome.ucsc.edu / cgi-bin / hgTables), remove the reads located in the blacklist regions, spacer regions, patch sequences, highly variable regions, and centromere-telomeric regions of the hg19 genome from the fragments obtained in 1).

[0086] 3. Sample GC bias calculation, filter out test samples with GC bias greater than 0.3;

[0087] The steps for GC bias value are as follows:

[0088] (1) Use bedtools makewindows to split the reference genome hg19 into 100 kb windows and count the GC content in each window;

[0089] (2) For each test sample, the sequencing depth R of each window is counted separately;

[0090] (3) Calculate the GC bias value S at the sample level GC, represents the deviation level of sequencing depth caused by GC distribution, and the calculation formula is as follows:

[0091]

[0092]

[0093] Where N is the total number of windows, R i represents the sequencing depth of the i-th window, M represents the average sequencing depth of all windows, and R loess Indicates the depth value (expected depth value) predicted by the loess model.

[0094] 4. GC correction: Since different GC contents of fragments will lead to coverage deviation in the experimental PCR amplification process, it is necessary to perform GC correction on the fragment data obtained in 2).

[0095] (1) Construct a baseline and select 29 healthy individuals. For each healthy sample, select fragments with a length of 50-400 bp in the files obtained in step 1-3) and calculate the GC content of each fragment. The GC content is stratified according to the 1% interval, and the number of fragments N corresponding to each GC content in each stratification under each chromosome (excluding chromosomes X and Y) is counted. gc ;

[0096] (2) After obtaining the fragment counts of different chromosomes with different GC contents of 29 healthy samples through the above (1), the means of the 29 distributions are further calculated, namely the GC bias-correction reference distribution.

[0097]

[0098] Where n represents the number of healthy samples selected to construct the reference distribution, i represents the i-th healthy sample, and M gc Indicates the reference distribution value corresponding to a certain gc content, N gc,i It indicates the number of fragments corresponding to the i-th healthy sample at a certain GC content.

[0099] (3) For the actual test sample, the file obtained after processing in steps 1-3) contains the chromosome number, starting position, ending position, fragment length, and GC content information of each fragment in the hg19 human reference genome. The number of fragments N corresponding to each GC content of each test sample is counted according to the method (1) above. gc,i , and the reference distribution M is corrected according to the GC bias of the healthy samples obtained in (2) above gc Calculate the weight value for each GC content Therefore, corresponding weight values ​​can be assigned to fragments with different GC contents, and the weight values ​​can be used to weight each fragment according to its GC content to complete GC correction.

[0100] 3. Extraction of new features of cfDNA fragment groups (Fragment Distribution End / Break Motifpattern (FDEM / FDBM))

[0101] For each measured sample, after the above steps, high-quality cfDNA fragment information extracted from plasma can be obtained, including the chromosome number of each fragment in the hg19 human reference genome, the starting position, the ending position, the fragment length, the GC content, etc. Based on this information, two new features, FDEM and FDBM, can be extracted. The specific steps are as follows:

[0102] 1. Fragment sorting

[0103] For each fragment of each test sample, only fragments with a length of 100-250 bp were retained, and the fragments were stratified with a span of 10 bp (100-109, 110-119, ..., 240-249) to obtain the fragment clusters of each layer:

[0104] F={i|100<=f i <=250},

[0105] F k ={i|i∈F,100+10k<=f i <100+10(k+1)},k∈{0,1,2,...,14}

[0106] Where i represents the i-th fragment, f i represents the length of the i-th fragment, F represents the set of all fragments whose lengths are between 100 and 250 bp, and F k It represents the fragment collection under each fragment length layer (15 types in total);

[0107] 2. Extract base sequence

[0108] The method of the present application includes the acquisition of two types of cfDNA fragment features: the 4bp terminal motif pattern (FDEM) based on fragment size and the 4bp breakpoint motif pattern (FDBM) based on fragment size, and can accurately classify early lung cancer samples and healthy samples.

[0109] Specifically, each fragment can extract a 4bp base sequence from the end of the fragment, including two extraction methods that constitute the following two different characteristics:

[0110] For the FDEM feature, extract the 4 bases upstream and downstream of the reference genome position corresponding to the 5' end of the two chains. It will start from the starting position and extend 4bp downstream (range: start, start+4), and only need to extract the original base sequence of the reference genome at this position, and the end position will extend 4bp upstream (range: end-4, end) to extract the complementary base sequence of the reference genome at this position;

[0111] For the FDBM feature, the four bases ±2bp corresponding to the reference genome position at the 5' end of each chain are extracted. It will extend 2bp upstream and downstream from the start position and the end position respectively (range: start / end-2, start / end+2), and the other steps are consistent with FDEM;

[0112] M∈{AAAA,AAAT,...,TTTA,TTTT,...CCCA,....GGGA},len(M)=256

[0113] Among them, M represents the possible combinations of sequences, a total of 256 combinations;

[0114] 3. Calculate eigenvalues

[0115] Finally, the proportion of each motif (256 types) under each type of fragment cluster after GC correction (accumulating and summing the fragment weight values) is counted. m,k ; Therefore, for each detection sample, 15*256=3840 features can be obtained.

[0116]

[0117] Among them, N m,k represents the number of fragments of a certain motif under a certain fragment length k, N k represents the total number of fragments under a certain fragment length k;

[0118] 4. Background noise reduction of feature matrix based on principal component analysis (Principal Component Correction, PCC)

[0119] The basic principle of PCC is to derive the relevant principal components (PCs) from the feature matrix of the healthy sample set, and then correct the feature data of the test sample by subtracting its orthogonal projection on the hyperplane determined by the principal components (PCs) from the feature data of each test sample.

[0120] 1) Randomly select 24 healthy samples from the samples in the training set and obtain the FDEM / FDBM feature matrix H (dimension is 24×3840) of the 24 healthy samples;

[0121] 2) Perform singular value decomposition (SVD) on the matrix H, H = UDV T , where U and V are both orthogonal matrices; D is a diagonal matrix with non-negative singular values ​​on the diagonal; the column vectors of U are called left singular vectors, and the column vectors of V are called right singular vectors. T Represents the transposed matrix of matrix V. The diagonal elements of D are called singular values ​​and are usually arranged in descending order. In SVD decomposition, singular values ​​represent the importance of the original matrix. The features represented by the left and right singular vectors with larger singular values ​​contribute more to the data. When the cumulative sum of k singular values ​​is above 90%, V T The matrix X (dimension is 3840×k) composed of the first k column vectors of the matrix can be expressed as the main features of the matrix H, that is, the background noise that needs to be corrected;

[0122] 3) In order to correct the background noise (including health signals, batch signals, and GC bias noise), we subtract the orthogonal projection of each detection sample in the X subspace from its orthogonal projection in the X subspace. The formula is as follows: Among them, F i ′ represents the eigenvector of sample i after principal component correction, F i represents the original FDEM / FDBM eigenvector of sample i, j represents the jth principal component, and dot represents the vector dot product;

[0123] 5. Model Construction

[0124] 1. Feature Dimensionality Reduction

[0125] The preProcess module in the caret package of R language was used to perform PCA dimensionality reduction on the feature data in the training set, and only features with an explanatory power of more than 90% were retained for model construction.

[0126] 2. Model training

[0127] The glmnet model in the caret package of R language is used to train the model for 1) the features after dimensionality reduction. The 5-fold cross-validation method (method=repeatedcv) repeated 5 times is selected to adjust the model parameters in the grid. The parameters include alpha=seq(0,1,by=0.1) and lambda=10^seq(-5,0,length.out=100)). The parameters corresponding to the highest AUC are used as the final parameters of the glmnet model, and a single feature prediction model can be obtained.

[0128] 3. Model Integration

[0129] The predicted values ​​obtained from each single feature training in step 2) are combined as the input matrix of the integrated model, and then trained through the logistic regression model (family = binomial) in the glmnet package of the R language to obtain the final integrated model and the final predicted score value of the sample.

[0130] Example 2: Performance evaluation of early screening model

[0131] In this example, 268 healthy samples and 79 lung cancer patients were randomly divided into training set and test set in a ratio of 6:4. After extracting the FDEM / FDBM features of cfDNA in the cancer group and the non-cancer group, the model was trained and predicted to obtain AUC, sensitivity and specificity.

[0132] In the training set, the AUCs of the FDEM and FDBM single-feature models were 0.9739 (95% CI: 0.9418-1) and 0.9785 (95% CI: 0.9492-1), respectively. At 95% specificity, the cutoffs were 0.611 and 0.361, respectively. In the test set, the AUCs of the FDEM and FDBM single-feature models were 0.9947 (95% CI: 0.9764-1) and 0.994 (95% CI: 0.9746-1), respectively. At the above corresponding cutoffs, the sensitivities of the test sets were 1.0 (95% CI: 1.0-1.0) and 0.9032 (95% CI: 0.7991-1), respectively. The specificities were 0.9655 (95% CI: 0.9358-0.9952) and 0.9862 (95% CI: 0.9672-1), respectively.

[0133] In the integrated model, the AUC of the training set was 0.9804 (95% CI: 0.9525-1), and the cutoff was 0.296 under the condition of 95% specificity. The AUC of the test set was 0.9963 (95% CI: 0.9812-1), and the sensitivity of the test set under the corresponding cutoff was 0.9677 (95% CI: 0.9055-1), and the specificity was 0.9862 (95% CI: 0.9672-1).

[0134] It can be seen that the model constructed by this application has a very significant effect advantage in the prediction of lung cancer screening.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some or all of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for constructing a lung cancer early screening model based on extracting FDEM / FDBM features from low-depth sequencing data and performing PCC system noise reduction, characterized in that: The steps include: 1) Low-depth whole genome sequencing data acquisition and preprocessing; 2) Extracting cfDNA fragment group features FDEM and FDBM; the FDEM is a 4 bp end motif pattern based on fragment size, and the FDBM is a 4 bp breakpoint motif pattern based on fragment size; 3) Perform background noise reduction on the feature matrix based on principal component analysis (PCC) method; 4) Construction of early screening model; The preprocessing in step 1) includes: for each sample, performing fragment filtering based on the alignment quality value and fragment length, removing fragments with an alignment quality value less than 30 and a fragment length greater than 400 bp; using public database information to remove reads located in the hg19 genome blacklist region, spacer region, patch sequence, highly variable region and centromere-telomeric region; filtering test samples with GC bias greater than 0.3; and performing GC correction on the obtained fragment data; The extraction in step 2) comprises the following steps: 1) For each fragment of each test sample, only the fragments with a length of 100-250 bp are retained, and the fragments are layered to obtain the fragment clusters of each layer; The stratification is to stratify the fragments with a span of 10 bp: 100-109 bp, 110-119 bp, ..., 240-249 bp, to obtain the fragment clusters of each layer: F={i|100<=f i <=250}, F k ={i|i∈F,100+10k<=f i <100+10(k+1)},k∈{0,1,2,...,14} Where i represents the i-th fragment, f i represents the length of the i-th fragment, F represents the set of all fragments whose lengths are between 100 and 250 bp, and F k Represents the set of fragments under each fragment length layer; 2) Extract base sequences for each fragment: for FDEM features, extract the 4 bases upstream and downstream of the reference genome position corresponding to the 5' end of the two chains; for FDBM features, extract the 4 bases ±2bp of the reference genome position corresponding to the 5' end of the two chains, that is, extend 2bp upstream and downstream from the start position and the end position respectively; The extraction yielded the following possible combinations: M∈{AAAA,AAAT,...,TTTA,TTTT,...CCCA,....GGGA},len(M)=256 Among them, M represents the possible combinations of sequences, a total of 256 combinations; 3) The frequency information of each motif after GC correction under each fragment cluster is counted respectively, that is, the proportion of the cumulative sum of the fragment weight values; the proportion of the cumulative sum is f m,k , the calculation formula is as follows; Among them, N m,k represents the number of fragments of a certain motif under a certain fragment length k, N k represents the total number of fragments under a certain fragment length k; The background noise reduction in step 3) comprises the following steps: 1) Randomly select multiple healthy samples from the training set samples to obtain the FDEM / FDBM feature matrix H of the multiple healthy samples; 2) Perform singular value decomposition SVD on the matrix H, H = UDV T , where U and V are both orthogonal matrices; D is a diagonal matrix with non-negative singular values ​​on the diagonal; the column vectors of U are called left singular vectors, and the column vectors of V are called right singular vectors. T represents the transposed matrix of matrix V; the diagonal elements of D are called singular values, which are arranged in descending order. In SVD decomposition, singular values ​​represent the importance of the original matrix. The features represented by the left and right singular vectors with larger singular values ​​contribute more to the data; when the cumulative sum of k singular values ​​is above 90%, V T The matrix X composed of the first k column vectors of the matrix can be expressed as the main features of the matrix H, that is, the background noise that needs to be corrected; 3) Subtracting the orthogonal projection of each test sample in the X subspace to correct the background noise; the background noise includes healthy signals, batch signals and GC bias noise; The correction calculation formula is as follows: Among them, F i ′ represents the eigenvector of sample i after principal component correction, F i represents the original FDEM eigenvector of sample i, j represents the jth principal component, and dot represents the vector dot product.

2. The construction method according to claim 1, characterized in that: The acquisition in step 1) is to download data files containing fragment information of lung cancer and healthy samples from the public database FinaleDB; the fragment information includes chromosome number, starting position, ending position, comparison quality value and positive and negative chains.

3. The construction method according to any one of claims 1-2, characterized in that: The construction of the early screening model in step 4) is specifically as follows: The feature data in the training set that has undergone PCC denoising is subjected to PCA dimensionality reduction, and the features that satisfy the explanatory power of more than 90% are retained for model construction; the model is trained on the features after dimensionality reduction, and 5-fold cross-validation is selected to perform grid parameter adjustment on the model. The parameters corresponding to the highest AUC are used as the final model parameters to obtain a single feature prediction model; the predicted values ​​obtained from the single feature training are merged as the input matrix of the integrated model, and the logistic regression model is used for training to obtain the final integrated model, and the final prediction score value of the sample is obtained.

4. A computer storage medium, characterized in that: The computer storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the method according to any one of claims 1 to 3 is implemented.

5. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory and the processor are connected, the memory is used to store a computer program, and the processor is used to call the computer program to implement the method according to any one of claims 1 to 3.

Images