Copy number variation detection method based on semi-supervised learning

Abstract

The invention relates to the technical field of gene variation detection, in particular to a copy number variation detection method based on semi-supervised learning. Comprising the following steps: acquiring a read depth signal and a mapping quality signal of each normal window of a reference genome sequence from comparison information of sequencing reads, and performing GC content deviation correction on the read depth signals of all the normal windows; all normal windows are segmented into segmented areas with uniform read depth signals by adopting a cyclic binary segmentation algorithm, and copy number variation breakpoint positions are identified in combination with a read splitting strategy; performing normalization processing on the mapping quality signal; performing smooth noise reduction processing on the read depth signal; labeling pseudo labels for the corresponding segmented areas; according to the method, clustering analysis is carried out on all segmented areas through an improved density clustering algorithm, integration and variation type judgment are carried out on abnormal segmented areas, and a copy number variation detection result is output, so that efficient detection of copy number variation is realized, and the accuracy and reliability of the detection result are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gene variation detection, and particularly relates to a copy number variation detection method based on semi-supervised learning. BACKGROUND

[0002] Genomic structural variation refers to significant changes in DNA sequence in length or position, including insertion, deletion, inversion, translocation and copy number variation (CNV) and the like. Among them, copy number variation is the most common and highest research value variation type, which refers to the duplication or deletion of genomic fragments in the range of thousands to millions of base pairs. Research shows that copy number variation plays an important role in gene expression regulation, phenotype difference and disease occurrence, and is one of the main sources of human genome diversity, and is closely related to many tumors and nervous system diseases.

[0003] In recent years, with the rapid development of high-throughput sequencing technology and computing methods, copy number variation detection has gradually developed from the initial method based on simple statistical features to the detection framework combining machine learning, deep learning and graph modeling and other advanced algorithms. According to whether relying on artificial labeled data and its utilization mode, the existing copy number variation detection methods are mainly divided into unsupervised detection methods and fully supervised detection methods.

[0004] The unsupervised detection method mainly relies on the statistical characteristics of the sequencing data itself, including sequencing depth distribution, GC content bias and paired-end distance. Among the detection tools of the unsupervised detection method, the representative detection tools are CNVkit, GROM-RD, Matchclips2 and OTSUCNV. CNVkit realizes the detection of copy number variation in the whole genome range by integrating targeted sequencing data and off-target sequencing data. CNVkit shows relatively balanced detection effect under different sequencing depths, but due to the adoption of a relatively conservative variation determination strategy, the number of detected copy number variations is relatively small, which affects the performance of CNVkit in the comprehensive evaluation index. GROM-RD detects copy number variation by analyzing the sequencing depth bias in genomic data, and improves the resolution of breakpoint detection by using windows of different sizes and overlaps for copy number variation search. However, when the tumor purity is low, the detection effect of GROM-RD decreases. Matchclips2 uses read position information and alignment features across structural variation regions to identify breakpoints, and has high breakpoint resolution. However, in the data sample with low sequencing depth, the detection performance is obviously limited due to insufficient read position information. OTSUCNV adopts an adaptive sequence segmentation method, which does not rely on prior assumptions about data distribution, and constructs a variation recognition framework combining image threshold segmentation principles. However, since the variation boundary is not finely positioned, OTSUCNV still has room for improvement in breakpoint accuracy.

[0005] The full-supervised detection method uses a large number of known true positive copy number variation results as a training set, so that the model learns the mapping relationship between the sequencing signal and the copy number variation, and realizes the automatic identification of the copy number variation region by combining multi-modal deep learning, graph modeling and neural network strategies. The full-supervised detection method usually performs better in terms of variation detection accuracy and boundary positioning accuracy, but requires higher computing resources and more complex model training process, and is highly dependent on large-scale and high-quality labeled data. However, it is often difficult to obtain enough true positive labels in real detection scenarios, which limits the generalization ability of the model to a certain extent. SUMMARY

[0006] The application provides a copy number variation detection method based on semi-supervised learning, which effectively solves the problems of insufficient detection accuracy, limited boundary recognition accuracy and excessive dependence on labeled data in complex sequencing environment. In the sequencing data sample to be detected, a small amount of labeled feature information is used to fully exploit the potential feature information in the unlabeled data, so as to achieve efficient detection of copy number variation and significantly improve the accuracy and reliability of the detection results.

[0007] The application provides a copy number variation detection method based on semi-supervised learning, which effectively solves the problems of insufficient detection accuracy, limited boundary recognition accuracy and excessive dependence on labeled data in complex sequencing environment. In the sequencing data sample to be detected, a small amount of labeled feature information is used to fully exploit the potential feature information in the unlabeled data, so as to achieve efficient detection of copy number variation and significantly improve the accuracy and reliability of the detection results. Step one, input the reference genome sequence, the sequencing data sample to be detected and the seed file, align and sort the sequencing data sample and the reference genome sequence, and obtain the sorted BAM file; Step two, in the reference genome coordinate system, divide the reference genome sequence into fixed-size and non-overlapping windows, obtain the sequencing reads in the sorted BAM file, and according to the alignment information of the sequencing reads, count the read counts of all windows of the reference genome sequence, identify and remove the abnormal windows containing N bases in the reference genome sequence, and retain the normal windows not containing N bases; wherein the missing values of the alignment information of the sequencing reads are replaced by zero padding; Step three, obtain the read depth signal and mapping quality signal of each normal window of the reference genome sequence from the alignment information of the sequencing reads, and save the features of all normal windows of the reference genome sequence; wherein the read depth signal is the RD signal, which is used to represent the sequencing depth of the base in the sequencing data sample; the mapping quality signal is the MQ signal, which is used to represent the quality of the alignment of the sequencing data sample and the reference genome sequence; Step four, GC content bias correction is performed on the read depth signal of all normal windows; Step 5: Based on the corrected read depth signal, a cyclic binary segmentation algorithm is used to segment all normal windows into segmented regions with uniform read depth signals. Combined with the split read strategy, potential copy number variation breakpoint positions are identified. Step 6: normalize the mapping quality signals of all segmented regions; Step 7: Smoothing and noise reduction are performed on the read depth signals of all segmented areas; Step 8: Based on the start and end positions of the copy number variation in the seed file, pseudo labels are added to the corresponding segmented regions; Step 9: The segmented regions annotated with pseudo labels are used as cluster cores in the semi-supervised learning clustering process. The read depth signals of all segmented regions are used as cluster features. All segmented regions are clustered using an improved density clustering algorithm. The cluster analysis results are marked as normal segmented regions or abnormal segmented regions using cluster labels. Step 10: Integrate the abnormal segmented regions and determine the variation type, and output the copy number variation detection results.

[0008] Furthermore, in step 1, the sequencing data sample is aligned with the reference genome sequence using BWA software to generate an initial BAM file. Based on the reference genome sequence position, the initial BAM file is sorted using SAMtools software to obtain a sorted BAM file containing specific sequencing reads.

[0009] Furthermore, in step 2, the positions containing N bases in the reference genome sequence are identified and processed, the reference genome sequence is divided into non-overlapping windows of fixed size, the windows containing N bases are detected, and the windows containing N bases are defined as abnormal windows. The read counts of the abnormal windows are marked as specific negative values ​​for distinction. The read count refers to the number of times the sequencing reads in the BAM file appear in the reference genome sequence window, which is used to reflect the sequencing coverage of the window. If the read count of the window is negative, the window is determined to be an abnormal window and deleted; otherwise, it is determined to be a normal window, and the start and end positions of the normal window are recorded. For missing values ​​in the alignment information of the sequencing reads, zero filling is used to replace the missing values.

[0010] Furthermore, in step 3, the alignment information of the sequencing reads is obtained, and the read depth signal of each base in the normal window is counted according to the start and end positions of each normal window in the reference genome sequence, and the average read depth signal of all bases in the normal window is used as the read depth signal of the normal window, that is,

[0011] in, Indicates the The read depth signal of a normal window, Indicates the length of the normal window, Indicates the In a normal window base-level read depth signal; Obtain the alignment information of the sequencing reads, count the mapping quality signals of each base in the normal window according to the start and end positions of each normal window in the reference genome sequence, and take the average mapping quality signal of all bases in the normal window as the mapping quality signal of the normal window, that is,

[0012] in, Indicates the The mapping quality signal of a normal window, Indicates the length of the normal window, Indicates the In a normal window The mapping quality signal of each base is given.

[0013] Furthermore, in step 4, the GC content deviation correction is performed on the read depth signals of all normal windows, that is,

[0014] in, Indicates the corrected The read depth signal of a normal window, Indicates the first The read depth signal of a normal window, Represents the average reading depth signal of all normal windows before correction, Indicates the difference between before and after correction The average read depth signal of normal windows with highly consistent GC content.

[0015] Furthermore, in step five, all normal windows are segmented using a cyclic binary segmentation algorithm to divide the normal windows into segmented regions with uniform read depth signals. During the segmentation process, for discontinuous windows containing gaps, the discontinuous windows are split and merged according to the fixed length of the segmented regions to restore the continuity of the reference genome sequence coordinates. For wide windows whose window length exceeds a preset length threshold, a fixed-size sliding window is used, starting from the starting point of the wide window, and sliding sequentially according to the set step size to further subdivide the wide window into segmented regions of uniform length. The split-read strategy based on CIGAR string is adopted to identify the breakpoint position of copy number variation. The breakpoint is inferred by comparing the relative position relationship between the truncated part and the matched region of the reference genome sequence in the CIGAR string. If the truncated part is located before the matched region, the breakpoint position is located at the end position of the truncated part. If the truncated part is located after the matched region, the breakpoint position is located at the start position of the truncated part. That is,

[0016] wherein, represents the continuous region matched between the sequencing read and the reference genome sequence, represents the soft truncation in the CIGAR string, represents the hard truncation in the CIGAR string, represents that there are bases in the sequencing read matched with the reference genome sequence, represents that the bases with length of in the sequencing read are not aligned to the reference genome sequence, but are still reserved in the sequencing read, represents that the bases with length of in the sequencing read are not aligned to the reference genome sequence, and are discarded in the sequencing read; and represents that the truncated part is located after the matched region, and represents that the truncated part is located before the matched region.

[0017] Further, in step six, the mapping quality signals of all segmented regions are normalized to eliminate systematic bias introduced by sequencing depth difference, so that the mapping quality signals more truly reflect the mapping quality, that is,

[0018] wherein, represents the mapping quality signal of the segmented region after normalization, represents the average value of the mapping quality signal of the segmented region before normalization, represents the average value of the read depth signal of the segmented region before normalization.

[0019] Further, in step seven, the read depth signals of all segmented regions are smoothed and denoised by using a total variation model. A fidelity term is constructed by calculating the difference between the original read depth signal and the smoothed and denoised read depth signal, which is used to constrain the smoothing and denoising process to ensure that the main features of the original read depth signal are maintained during the smoothing and denoising process. The sum of the absolute values of the differences between adjacent segmented regions in the read signal is calculated as a total variation regularization term, which quantifies the changes in read depth signals between adjacent segmented regions. Different optimization weights are assigned to different segmented regions according to the size of the difference in read depth signals between adjacent segmented regions. That is,

[0020] wherein represents the read depth signal of the segmented region after smoothing and denoising, represents the read depth signal of the segmented region before smoothing and denoising, i.e., the original read depth signal, represents a regularization parameter for balancing the fidelity and smoothness of the read depth signal, represents the number of segmented regions, is the fidelity term of the total variation model, which prevents the read depth signal after smoothing and denoising from deviating too much from the original read depth signal, represents the deviation of the read depth signal of the th segmented region from the read depth signal of the th segmented region, represents the total variation regularization term, which penalizes sharp fluctuations in the read depth signal, smooths the read depth signal, and preserves the true inflection points or mutations.

[0021] Further, in step eight, pseudo labels are annotated for the corresponding segmented regions based on the start and end positions of the copy number variations in the seed file. The seed file is constructed from the intersection region of the detection results of multiple existing copy number variation detection tools. If a segmented region with a pseudo label meets the following two screening conditions, i.e., the segmented region with a pseudo label belongs to the intersection region of the detection results of existing copy number variation detection tools, and the read depth signal of the segmented region with a pseudo label exceeds the principle that the read depth signal of the segmented region with a pseudo label exceeds the average value of the read depth signals of all segmented regions by more than three standard deviations, it is considered to be a high-confidence pseudo label region, i.e.,

[0022] wherein represents the read depth signal of the th segmented region, represents the average value of the read depth signals of all segmented regions, represents the standard deviation of the read depth signals of all segmented regions.

[0023] Furthermore, in step nine, the high-confidence pseudo-label region is used as the clustering core, and the read depth signals of all segmented regions are used as clustering features. All segmented regions are clustered based on the clustering core and the improved density clustering algorithm. The improved density clustering algorithm is based on the traditional density clustering algorithm. On the basis of the original two parameters of neighborhood radius and minimum number of neighbors, the z-score is introduced as the third constraint of the improved density clustering algorithm to quantify the degree of deviation of the read depth signal of a segment relative to the average value of the read depth signals of all segmented regions. Under the conditions of satisfying the neighborhood radius and the minimum number of neighbors, the segmented region with a z-score higher than the deviation threshold is marked as an abnormal segmented region through the clustering label, otherwise it is marked as a normal segmented region, wherein,

[0024] For the The z-score of each segmented region, Indicates the The read depth signal of each segmented region, represents the average value of the read depth signal of all segmented regions, Represents the standard deviation of the read depth signal across all segmented regions.

[0025] Furthermore, in step 10, the abnormal segmentation regions screened out from the cluster analysis results are sorted by starting position. If the interval between two adjacent abnormal segmentation regions is very small, the two abnormal segmentation regions are merged, and the average values ​​of the read depth signal, GC content and mapping quality signal are taken as new features of the merged segmentation region. If the interval is large, the abnormal segmentation region is recorded as an independent variation region. For the merged segmentation region, if the read depth signal of the abnormal segmentation region is higher than the average value of the read depth signals of all merged segmentation regions, the variation type of the abnormal segmentation region is defined as duplication, otherwise the variation type is defined as deletion. After the starting position, end position, read depth signal and variation type of the merged abnormal segmentation region are integrated, they are saved in a TXT file as the detection result and output.

[0026] The present invention provides a copy number variation detection method based on semi-supervised learning, suitable for identifying and localizing copy number variations in the human genome. For sequencing data samples generated by second-generation sequencing technology, this method can utilize a small amount of high-confidence annotated data to detect a large amount of unannotated data, effectively identifying potential copy number variation regions. It has the following technical advantages: (1) This invention innovatively uses the results of existing copy number variation detection tools for semi-supervised learning methods, breaking through the previous limitation of being used only for benchmark verification and significantly improving the efficiency of using detection results.

[0027] (2) On the basis of the traditional density clustering algorithm framework, the z-score is introduced as the third constraint condition of the improved density clustering algorithm, which is used to quantify the deviation degree of the read depth signal of a segmented region relative to the average value of the read depth signals of all segmented regions, so as to improve the identification accuracy of abnormal segmented regions in the clustering process.

[0028] (3) The high-confidence pseudo-label region is embedded into the clustering process as a clustering core, which effectively guides the determination of the clustering core and the clustering expansion process, and realizes more targeted detection of copy number variation regions. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 is the implementation flowchart of the present application; Figure 2 is the contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV under the condition that the sequencing depth of the simulation sequencing data sample is 10X and the tumor purity is 0.4; Figure 3 is the contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV under the condition that the sequencing depth of the simulation sequencing data sample is 10X and the tumor purity is 0.6; Figure 4 is the contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV under the condition that the sequencing depth of the simulation sequencing data sample is 10X and the tumor purity is 0.8; Figure 5 is the contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV under the condition that the sequencing depth of the simulation sequencing data sample is 20X and the tumor purity is 0.4; Figure 6 is the contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV under the condition that the sequencing depth of the simulation sequencing data sample is 20X and the tumor purity is 0.6; Figure 7 is the contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV under the condition that the sequencing depth of the simulation sequencing data sample is 20X and the tumor purity is 0.8; Figure 8A contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the case of a simulated sequencing data sample with a sequencing depth of 30X and a tumor purity of 0.4; Figure 9 A contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the case of a simulated sequencing data sample with a sequencing depth of 30X and a tumor purity of 0.6; Figure 10 A contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the case of a simulated sequencing data sample with a sequencing depth of 30X and a tumor purity of 0.8; Figure 11 A contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the case of a simulated sequencing data sample with a sequencing depth of 50X and a tumor purity of 0.4; Figure 12 A contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the case of a simulated sequencing data sample with a sequencing depth of 50X and a tumor purity of 0.6; Figure 13 A contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the case of a simulated sequencing data sample with a sequencing depth of 50X and a tumor purity of 0.8; Figure 14 A bar chart of the F1 scores of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in four real sequencing data samples; Figure 15 A contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the real sequencing data sample NA19238; Figure 16 A contour plot of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the real sequencing data sample NA19239; Figure 17A contour plot of the comparison of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the real sequencing data sample NA19240; Figure 18 A contour plot of the comparison of the detection results of the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the real sequencing data sample HG002; Figure 19 An UpSet plot of the number of true positive CNVs detected by the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the real sequencing data sample NA19238; Figure 20 An UpSet plot of the number of true positive CNVs detected by the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the real sequencing data sample NA19239; Figure 21 An UpSet plot of the number of true positive CNVs detected by the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the real sequencing data sample NA19240; Figure 22 An UpSet plot of the number of true positive CNVs detected by the present application and the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in the real sequencing data sample HG002. DETAILED DESCRIPTION

[0030] As shown in Figure 1 , the present application provides a copy number variation detection method based on semi-supervised learning, which is mainly realized through the following steps.

[0031] Step one, use the BWA-MEM function in the BWA software to align the sequencing data sample with the reference genome sequence to generate an initial BAM file. Use the SAMtools tool to sort the initial BAM file according to the position of the reference genome sequence to obtain a sorted BAM file containing specific sequencing reads.

[0032] In step two, in order to ensure the accuracy and reliability of the analysis results, the N base in the reference genome sequence and the missing value of the alignment information of the sequencing reads need to be identified and processed. In the reference genome sequence, the base types mainly include A, G, C and T four types, and there is a special N base, which is used to represent the undetermined base, and the N base can be any base. Because the position corresponding to the N base lacks exact base information, if it is directly involved in analysis, it will introduce bias and affect the reliability of the results. Therefore, the region containing the N base in the reference genome sequence needs to be identified and deleted. The reference genome sequence is divided into fixed-size non-overlapping windows, the window containing the N base is marked as an abnormal window, and the read count of the abnormal window is set to a specific negative number (such as: -9999) for marking. If it is detected that the read count of a window is less than 0, it is determined that the window contains the N base, which is considered as an abnormal window and is deleted from the analysis; otherwise, if it is detected that the read count of a window is a non-negative number, it is determined that the window does not contain the N base, which is considered as a normal window, and the start and end positions of the normal window are retained. For the missing value of the alignment information generated by the failure of the sequencing data sample to align with the reference genome sequence or the failure to record the main alignment information, the missing value is replaced by zero padding at the position of the missing value to ensure the consistency of subsequent statistical analysis.

[0033]

[0034] wherein, represents the read depth signal of the i-th normal window, represents the length of the normal window, represents the read depth signal of the j-th base in the i-th normal window; Obtain the alignment information of the sequencing reads, and according to the start and end positions of each normal window in the reference genome sequence, count the mapping quality signal of each base in the normal window, and take the average value of the mapping quality signals of all bases in the normal window as the mapping quality signal of the normal window, that is,

[0035] wherein, represents the mapping quality signal of the i-th normal window, represents the length of the normal window, represents the mapping quality signal of the j-th base in the i-th normal window.

[0036] ​​​​​​Step 4: GC content bias refers to the phenomenon that the GC content of sequencing data samples is unevenly distributed in different regions. In high-throughput sequencing, GC content will produce systematic bias in sequencing: regions with higher GC content are prone to enrichment during library construction and amplification, resulting in increased sequencing depth; while regions with lower GC content may have insufficient sequencing coverage, resulting in reduced sequencing depth. In order to reduce the impact of GC content bias on the analysis results, it is necessary to perform GC content correction on the extracted read depth signal, that is,

[0037] in, Indicates the corrected The read depth signal of a normal window, Indicates the first The read depth signal of a normal window, Represents the average reading depth signal of all normal windows before correction, Indicates the difference between before and after correction The average read depth signal of normal windows with highly consistent GC content.

[0038] Step 5: Segment all normal windows using a cyclic binary segmentation algorithm. This algorithm continuously segments normal windows into segmented regions with consistent read depth signals. Each segmented region typically corresponds to a specific copy number state. For segmented regions containing gaps, the segmented regions are fixed in length, split at the window gaps, and merged into small windows to restore the continuity of the reference genome sequence coordinates. Secondly, for wide windows exceeding a preset length threshold (e.g., 5000bp), a fixed-size sliding window is used, starting from the starting point of the wide window and sliding sequentially at a set step size, further subdividing the wide window into segmented regions of consistent length.

[0039] In order to further improve the accuracy of copy number variation breakpoint identification, a split read strategy based on CIGAR strings is adopted. For the alignment information of the sequencing reads, the alignment information containing soft truncation or hard truncation operations is preferentially screened. The truncation operation indicates that part of the sequence of the sequencing read fails to match the reference genome sequence during the alignment process, and the truncated position may be the copy number variation breakpoint position. By comparing the relative position relationship between the truncated part in the CIGAR string and the matching region of the reference genome sequence, the possible breakpoint position is inferred. If the truncated part is located before the matching region, the breakpoint position is located at the end position of the truncated part; if the truncated part is located after the matching region, the breakpoint position is located at the beginning position of the truncated part, that is,

[0040] in, denotes a continuous region of the sequencing read that matches the reference genome sequence, denotes a soft clip in the CIGAR string, denotes a hard clip in the CIGAR string, denotes that there are bases in the sequencing read that match the reference genome sequence, denotes that there are bases in the sequencing read that do not align to the reference genome sequence but are still retained in the sequencing read, denotes that there are bases in the sequencing read that do not align to the reference genome sequence and are discarded in the sequencing read; and denotes that the clipped portion is located after the matched region, and denotes that the clipped portion is located before the matched region.

[0041] In step six, in order to make the mapping quality signal of each segmented region more objectively reflect the actual mapping quality, eliminate the systematic bias caused by the difference of sequencing depth, it is necessary to normalize the mapping quality signal, that is,

[0042] wherein, denotes the mapping quality signal of the th segmented region after normalization, denotes the average of the mapping quality signal of the th segmented region before normalization, denotes the average of the read depth signal of the th segmented region before normalization.

[0043] Step seven, the read signal of all segmented regions is smoothed and denoised by using the total variation model. By minimizing the difference between the original read signal and the smoothed and denoised read signal, the difference between the read signals of adjacent segmented regions is controlled to highlight the copy number variation region and suppress local noise. In the smoothing and denoising process, the difference between the original read depth signal and the smoothed and denoised read depth signal is calculated to construct the fidelity term, which is used to constrain the smoothing and denoising process to ensure that the main features of the original read depth signal are maintained during the smoothing and denoising process, thereby avoiding the smoothing and denoising result deviating too much from the original read depth signal. The sum of the absolute values of the differences between the read signals of adjacent segmented regions is calculated as the total variation regular term, which quantifies the changes in the read depth signals between adjacent segmented regions. Larger differences usually correspond to potential copy number variations, while more subtle fluctuations are often noise. According to the size of the read depth signal difference between adjacent segmented regions and the noise level, different optimization weights are assigned to different segmented regions. Higher weights are assigned to segmented regions that detect true breakpoints or significant read depth signal mutations to enhance the preservation of key changes. In segmented regions with more noise fluctuations, lower weights are assigned to smooth out meaningless noise fluctuations. The formula for smoothing and denoising is as follows,

[0044] wherein represents the read depth signal of the segmented region after smoothing and denoising, represents the read depth signal of the segmented region before smoothing and denoising, i.e. the original read depth signal, represents the regularization parameter, and in the smoothing process, the regularization parameter is set to 0.25 to balance the fidelity and smoothness of the read depth signal, represents the number of segmented regions, is the fidelity term of the total variation model, which prevents the read depth signal after smoothing and denoising from deviating too much from the original read depth signal, represents the deviation of the read depth signal of the th segmented region from the read depth signal of the th segmented region, represents the total variation regular term, which penalizes the sharp fluctuations in the read depth signal, smooths the read depth signal, and preserves the true inflection points or mutations.

[0045] Step eight, improve the identification accuracy of copy number variation by using the results of existing detection tools, and introduce seed files to preliminarily label the segmented regions with pseudo labels. The seed file is constructed by the intersection region of the output results of multiple representative existing copy number variation detection tools (CNVkit, GROM-RD, Matchclips2 and OTSUCNV). If the segmented region labeled with the pseudo label meets the following two screening conditions, it is considered as a high-confidence pseudo label region: the segmented region labeled with the pseudo label belongs to the intersection region of the detection results of the existing copy number variation detection tools, ensuring that the segmented region has strong cross-method consistency; the read depth signal corresponding to the segmented region labeled with the pseudo label needs to exceed the principle that the read depth signal of the segmented region labeled with the pseudo label needs to exceed the average value of the read depth signals of all segmented regions by more than three standard deviations. That is,

[0046] wherein represents the read depth signal of the i-th segmented region, represents the average value of the read depth signals of all segmented regions, represents the standard deviation of the read depth signals of all segmented regions.

[0047] Step nine, use the high-confidence pseudo label region as the clustering core and the read depth signal of all segmented regions as the clustering feature to perform clustering analysis on all segmented regions based on the clustering core and the improved density clustering algorithm. After obtaining the high-confidence pseudo label region, all segmented regions are clustered and analyzed by combining the improved density clustering algorithm with the introduction of z-score constraint to identify potential copy number variation regions. The traditional density clustering algorithm relies on two key parameters: neighborhood radius and minimum neighbor number. By introducing z-score as a third constraint condition into the density clustering algorithm, the z-score quantifies the deviation of the read depth signal of a certain segmented region from the global average read depth signal. Under the conditions of meeting the neighborhood radius and the minimum neighbor number, by setting the deviation threshold of z-score (such as ), the clustering label is used to mark the segmented region with a read depth signal significantly deviating from the global average read depth signal as an abnormal segmented region, and the segmented region with a read depth signal less deviating from the global average read depth signal is marked as a normal segmented region.

[0048]

[0049] is the z-score of the i-th segmented region, represents the z-score of the i-th segmented region, represents the z-score of the i-th segmented region, ​a read depth signal of the segmented region, an average of all segmented region read depth signals, a standard deviation of all segmented region read depth signals.

[0050] Step ten, the abnormal segmented regions screened out from the clustering analysis results are sorted according to the starting positions. The sorted abnormal segmented regions are traversed in turn, and for two adjacent abnormal segmented regions, if the difference between the ending position of the former abnormal segmented region and the starting position of the latter abnormal segmented region is not more than 1, the two abnormal segmented regions are merged, and the read depth signal, GC content and mapping quality signal of the merged abnormal segmented region are each averaged to serve as the characteristic information of the new abnormal segmented region, replacing the original characteristic. If the interval between the two adjacent abnormal segmented regions is greater than 1, the former abnormal segmented region is regarded as an independent variant region for recording, and a new merging process is started. For each merged abnormal segmented region, if the read depth signal of the abnormal segmented region is higher than the overall average of all segmented region read depth signals, it is determined that the variant type is duplication; otherwise, it is determined that the variant type is deletion. The starting position, ending position, read depth signal and determined variant type of each merged abnormal segmented region are sorted, and saved in the form of detection results to a TXT file for output.

[0051] To verify the effectiveness and fairness of the technical scheme of the application, the detection tools were evaluated from multiple angles on simulation sequencing data samples and real sequencing data samples, and uniform evaluation indexes were used for comparative analysis. For the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV, the default parameters were used for detection to ensure the objectivity and consistency of the comparative experiment.

[0052] In the simulation sequencing data experiment, 50 groups of samples were generated for each experimental configuration to reduce the randomness of the experimental results. In the simulation sequencing data experiment, the copy number variation detection method (SSLCNV) based on semi-supervised learning of the application was compared with the existing detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV in terms of precision, recall rate and F1 score. In the experimental evaluation, the precision is defined as the proportion of the number of true positive CNVs to the total number of all CNVs detected, and the recall rate is defined as the proportion of the number of true positive CNVs to the total number of real CNV samples. The F1 score is an index that considers the precision and recall rate of the detection tool comprehensively, and can be regarded as the harmonic mean of the two, which is used to evaluate the detection performance of the detection tool comprehensively.

[0053]

[0054] Wherein, TP represents true positive CNV, defined as the CNV whose detection result is consistent with the true CNV type, and the overlap ratio of the detection CNV length range and the true CNV length range is not less than 80%; FP represents false positive CNV, defined as the CNV whose detection result is not consistent with the true CNV type, or the overlap ratio of the detection CNV length range and the true CNV length range is less than 80%; FN represents false negative CNV, defined as the true CNV which is not recognized by the detection tool. TP+FP represents the total number of CNVs detected by the detection tool, and TP+FN represents the total number of true CNVs.

[0055] The simulation sequencing data experiment results are shown in Figures 2-13 As shown in Figure 2 , Figure 5 , Figure 8 and Figure 11 , when the tumor purity is 0.4, the existing detection tools face the problem of low recall rate due to the weak CNV signal itself, the accuracy of OTSUCNV is close to 1, but the recall rate is obviously insufficient, resulting in F1 score only between 0.5-0.7, while the SSLCNV proposed in the present application can significantly improve the recall rate while maintaining high accuracy, and the final F1 score reaches 0.8, which is better than the existing four detection tools. As shown in Figure 3 , Figure 6 , Figure 9 and Figure 12 , when the tumor purity is 0.6, the accuracy and recall rate of the SSLCNV of the present application are further improved, and the F1 score is increased to 0.89, which is significantly higher than the four existing detection tools. This result shows that the SSLCNV can still maintain a better detection balance when the CNV signal gradually increases. As shown in Figure 4 , Figure 7 , Figure 10 and Figure 13 , when the tumor purity is 0.8, the accuracy and recall rate of the SSLCNV of the present application are both more than 0.80, and continue to maintain the optimal performance. Although the recall rate of GROM-RD is high under this condition, the accuracy is lower than that of SSLCNV, resulting in the overall F1 score still lower than that of SSLCNV, while CNVkit and OTSUCNV still mainly have high accuracy, but the recall rate is low, so the overall detection effect is not as good as that of the SSLCNV of the present application.

[0056] To further verify the adaptability and robustness of the SSLCNV proposed in the present application under different sequencing depths, the present application compares the performance of SSLCNV and the existing four detection tools CNVkit, GROM-RD, Matchclips2 and OTSUCNV under the conditions of 10x, 20x, 30x and 50x sequencing depths based on simulated sequencing data samples. For each sequencing depth, statistical analysis is performed based on 150 sets of simulated sequencing data samples, the average value of the F1 score of the 150 sets of simulated sequencing data samples is calculated, and the F1 score improvement of SSLCNV compared with the existing four detection tools is evaluated. The F1 score improvement of SSLCNV compared with the existing four detection tools is shown in Table 1: Table 1 F1 score improvement of SSLCNV compared with the existing four detection tools

[0057] As can be seen from Table 1, SSLCNV achieves the highest F1 score under each sequencing depth. When the sequencing depth is 10x, the F1 score of SSLCNV reaches 0.739, which is about 10% higher than GROM-RD, 14.09% higher than CNVkit, and more than 100 times higher than Matchclips2, fully embodying the robustness advantage of SSLCNV in the environment with larger noise and sparse CNV signal. Under the conditions of 20x and 30x medium sequencing depth, SSLCNV still leads, with the F1 score of 0.807 at 20x, which is about 0.86% higher than Matchclips2, and the improvement amplitude compared with GROM-RD, OTSUCNV and CNVkit is 9.44%, 15.37% and 23.96% respectively; the F1 score of SSLCNV reaches 0.856 at 30x sequencing depth, which is 6.29% higher than the suboptimal Matchclips2, and more than 36.75% higher than CNVkit. Under the condition of 50x high sequencing depth, the F1 score of SSLCNV is as high as 0.889, which is 14.18% higher than Matchclips2, and the improvement amplitude compared with GROM-RD, OTSUCNV and CNVkit is 17.75%, 18.73% and 50.45% respectively. Experiments show that no matter under low, medium or high sequencing depth, SSLCNV of the present application can achieve higher precision and recall rate, and the F1 score of SSLCNV is always better than that of the existing four detection tools. This fully verifies the stability and reliability of the present application under different sequencing depths and complex tumor purity, especially in the scene with weak CNV signal and high detection difficulty.

[0058] As an effective method to evaluate the performance of algorithms, simulation sequencing data experiment can compare the performance of different detection tools under controllable conditions. However, the simulation sequencing data sample is still difficult to fully reflect the complexity of the genomic structure and the diversity of the variation in the real sequencing data sample. In order to further verify the detection effect of the SSLCNV method proposed in the application in the real scene, the SSLCNV and the four existing detection tools are applied to the real sequencing data sample. Specifically, it includes three real sequencing data samples NA19238, NA19239 and NA19240 from the thousand genome project and the real sequencing data sample HG002 from the NCBI database.

[0059] The real sequencing data experiment results are as shown in Figures 14-18 , wherein the F1 score of the SSLCNV and the four existing detection tools in the real sequencing data sample is as shown in Table 2: Table 2 F1 score of SSLCNV and four existing detection tools in real sequencing data sample

[0060] As Figure 14 and Table 2 show, SSLCNV achieves the highest F1 score in the four real sequencing data samples, maintains the optimal overall detection effect, and exhibits the stability and adaptability of SSLCNV in different data sets and different data sample backgrounds. Among them, in the real sequencing data sample HG002, the F1 score of SSLCNV reaches 0.196, which is much higher than that of GROM-RD, and shows strong detection capability for complex genomic regions. In the two real sequencing data samples NA19238 and NA19239, SSLCNV achieves F1 scores of 0.133 and 0.149 respectively, which are also higher than the F1 scores of the four existing detection tools. In the NA19240 sample, although the overall detection difficulty is greater, SSLCNV still maintains an F1 score of 0.114, while the F1 scores of the four existing detection tools are all less than 0.05. Compared with GROM-RD, SSLCNV not only improves the F1 score by about 9% and 4% in the NA19238 and NA19239 samples respectively, but also has a more significant improvement in the HG002 and NA19240 samples. Compared with CNVkit and OTSUCNV, the detection performance advantage of SSLCNV in all real sequencing data samples is more obvious, and the F1 score is improved by several times. As shown in Figures 15-17 , Matchclips2 has very low precision and recall rate in the three real sequencing data samples with low sequencing depth NA19238, NA19239 and NA19240, while SSLCNV can still maintain high detection accuracy, showing good robustness for low CNV signal detection. As shown in Figure 18As shown, CNVkit and OTSU CNV exhibit extremely high precision, but due to the number of true positive CNVs detected is too small, resulting in too low recall rate, and the F1 score is lower than that of SSL CNV.

[0061] Figures 19-22 The UpSet chart is used to show the number of intersections between the true positive CNVs identified by each detection tool in the four real sequencing data samples. The results show that SSL CNV not only detects the largest number of true positive CNVs in all real sequencing data samples, but also has the largest number of intersections with the true positive CNVs of the existing four detection tools, reflecting a more comprehensive detection coverage and stronger consistency. For example, Figures 19-21 As shown, in sample NA19238, the number of intersections of true positive CNVs between SSL CNV and CNVkit reaches 300; in samples NA19239 and NA19240, the number of intersections of true positive CNVs between SSL CNV and CNVkit is 353 and 461, respectively; as shown in Figure 22 As shown, in HG002 sample, the number of intersections of true positive CNVs between SSL CNV and GROM-RD reaches 569. It is worth noting that SSL CNV also detects true positive CNVs that the existing four detection tools fail to identify, further demonstrating its higher detection precision and the ability to discover potential new CNVs. As shown in Figures 19-22 As shown, in sample NA19238, SSL CNV independently detects 312 true positive CNVs; in samples NA19239 and NA19240, 637 and 1699 true positive CNVs are detected, respectively; in HG002 sample, 2667 true positive CNVs are detected. Therefore, SSL CNV of the present application exhibits better recall rate and precision in real sequencing data samples, and finally the F1 score is always better than that of the existing four detection tools. This fully proves that the present application not only has remarkable effect in simulated sequencing data samples, but also has strong detection capability and reliability in real sequencing data samples, and is more suitable for actual scientific research and clinical detection needs.

[0062] The above results show that the copy number variation detection method based on semi-supervised learning proposed by the application effectively solves the problems of insufficient detection accuracy, limited boundary recognition accuracy and excessive dependence on labeled data in complex sequencing environment. Under the premise of ensuring fairness and objectivity, the experiment was tested on 600 simulated sequencing data samples and 4 real sequencing data samples. Among them, the 600 simulated sequencing data samples come from the combination design of four different sequencing depths (10x, 20x, 30x and 50x) and three tumor purities (0.4, 0.6, 0.8). Specifically, by combining 4 sequencing depths and 3 tumor purities, 12 experimental conditions are formed, and 50 repeated samples are generated under each condition, totaling 4x3x50=600 simulated sequencing data samples, to comprehensively evaluate the performance of SSLCNV under different sequencing depths and tumor purities. In addition, 4 real sequencing data samples were selected, corresponding to real sequencing data samples NA19238, NA19239, NA19240 and HG002, to verify the applicability and stability of SSLCNV in the real genomic variation background. The test results show that whether in simulated sequencing data samples or in real sequencing data samples, the SSLCNV method proposed by the application achieves the best balance in terms of detection accuracy, recall rate and F1 score and other core indicators. Especially under complex conditions such as low sequencing depth or low tumor purity, it still maintains high detection accuracy and stability, which is significantly better than the existing four detection tools.

[0063] In summary, the experimental results fully prove that the application is a reliable tool for detecting copy number variations based on second-generation sequencing data samples, especially suitable for practical application scenarios that are greatly affected by noise, sequencing depth and tumor purity complexity, with high detection accuracy and wide application value.

Claims

1. A copy number variation detection method based on semi-supervised learning, characterized in that: The following steps are included: Step 1: Input the reference genome sequence, the sequencing data sample to be tested, and the seed file, align and sort the sequencing data sample with the reference genome sequence, and obtain the sorted BAM file; Step 2: In the reference genome coordinate system, the reference genome sequence is divided into fixed-size non-overlapping windows, and the sequencing reads in the sorted BAM file are obtained. The read counts of all windows of the reference genome sequence are counted based on the alignment information of the sequencing reads. Abnormal windows containing N bases in the reference genome sequence are identified and removed, and normal windows that do not contain N bases are retained. Missing values ​​in the alignment information of the sequencing reads are replaced by zero filling. Step 3: Obtain the read depth signal and mapping quality signal of each normal window of the reference genome sequence from the alignment information of the sequencing reads, and save the characteristics of all normal windows of the reference genome sequence; wherein the read depth signal is the RD signal, which is used to indicate the sequencing depth of the bases in the sequencing data sample; the mapping quality signal is the MQ signal, which is used to indicate the quality of the alignment between the sequencing data sample and the reference genome sequence; Step 4: Correct the GC content bias of the read depth signals of all normal windows; Step 5: Based on the corrected read depth signal, a cyclic binary segmentation algorithm is used to segment all normal windows into segmented regions with uniform read depth signals. Combined with the split read strategy, potential copy number variation breakpoint positions are identified. Step 6: normalize the mapping quality signals of all segmented regions; Step 7: Smoothing and noise reduction are performed on the read depth signals of all segmented areas; Step 8: Based on the start and end positions of the copy number variation in the seed file, pseudo labels are added to the corresponding segmented regions; Step 9: The segmented regions annotated with pseudo labels are used as cluster cores in the semi-supervised learning clustering process. The read depth signals of all segmented regions are used as cluster features. All segmented regions are clustered using an improved density clustering algorithm. The cluster analysis results are marked as normal segmented regions or abnormal segmented regions using cluster labels. Step 10: Integrate the abnormal segmented regions and determine the variation type, and output the copy number variation detection results.

2. The copy number variation detection method based on semi-supervised learning according to claim 1, characterized in that: In step 2, the positions containing N bases in the reference genome sequence are identified and processed, the reference genome sequence is divided into non-overlapping windows of fixed size, the windows containing N bases are detected, and the windows containing N bases are defined as abnormal windows. The read counts of the abnormal windows are marked with specific negative values ​​for distinction. The read count refers to the number of times the sequencing reads in the BAM file appear in the reference genome sequence window, which is used to reflect the sequencing coverage of the window. If the read count of the window is negative, the window is determined to be an abnormal window and deleted; otherwise, it is determined to be a normal window, and the start and end positions of the normal window are recorded. For missing values ​​in the alignment information of the sequencing reads, zero filling is used to replace the missing values.

3. The copy number variation detection method based on semi-supervised learning according to claim 2, characterized in that: In step 3, the alignment information of the sequencing reads is obtained, and the read depth signal of each base in the normal window is counted according to the start and end positions of each normal window in the reference genome sequence, and the average read depth signal of all bases in the normal window is used as the read depth signal of the normal window, that is, in, Indicates the The read depth signal of a normal window, Indicates the length of the normal window, Indicates the In a normal window base-level read depth signal; Obtain the alignment information of the sequencing reads, count the mapping quality signals of each base in the normal window according to the start and end positions of each normal window in the reference genome sequence, and take the average mapping quality signal of all bases in the normal window as the mapping quality signal of the normal window, that is, in, Indicates the The mapping quality signal of a normal window, Indicates the length of the normal window, Indicates the In a normal window The mapping quality signal of each base is given.

4. The copy number variation detection method based on semi-supervised learning according to claim 3, characterized in that: In step 4, the GC content bias correction is performed on the read depth signals of all normal windows, that is, in, Indicates the corrected The read depth signal of a normal window, Indicates the first The read depth signal of a normal window, Represents the average reading depth signal of all normal windows before correction, Indicates the difference between before and after correction The average read depth signal of normal windows with highly consistent GC content.

5. The copy number variation detection method based on semi-supervised learning according to claim 4, characterized in that: In step 5, all normal windows are segmented using a cyclic binary segmentation algorithm, and the normal windows are divided into segmented regions with uniform read depth signals. During the segmentation process, for discontinuous windows containing gaps, the discontinuous windows are split and merged according to the fixed length of the segmented region to restore the continuity of the reference genome sequence coordinates. For wide windows whose window length exceeds the preset length threshold, a fixed-size sliding window is used, starting from the starting point of the wide window, and sliding in sequence according to the set step size to further subdivide the wide window into segmented regions of uniform length. A split read strategy based on CIGAR strings is used to identify the breakpoint position of copy number variation. The breakpoint is inferred by comparing the relative position of the truncated part in the CIGAR string with the matching region of the reference genome sequence. If the truncated part is located before the matching region, the breakpoint position is located at the end of the truncated part; if the truncated part is located after the matching region, the breakpoint position is located at the beginning of the truncated part; that is, in, represents the continuous region where the sequencing reads match the reference genome sequence, Indicates soft truncation in CIGAR string, Indicates a hard truncation in a CIGAR string, Indicates that there is bases matched the reference genome sequence, Indicates that the length of the sequencing read is bases that did not map to the reference genome sequence but are still retained in the sequencing reads. Indicates that the length of the sequencing read is bases were not aligned to the reference genome sequence and were discarded in the sequencing reads; and Indicates that the cutoff part is after the matched area. and Indicates that the truncated portion is before the matched area.

6. The copy number variation detection method based on semi-supervised learning according to claim 5, characterized in that: In step six, the mapping quality signals of all segmented regions are normalized to eliminate the systematic bias introduced by the difference in sequencing depth so that the mapping quality signals can more truly reflect the mapping quality, i.e. in, After normalization, The mapping quality signal of the segmented region, Represents the first The average value of the mapping quality signal of the segmented area, Represents the first The average of the read depth signals of the segmented regions.

7. The copy number variation detection method based on semi-supervised learning according to claim 6, characterized in that: In step seven, the total variation model is used to smooth and denoise the read depth signals of all segmented regions. The fidelity term is constructed by calculating the difference between the original read depth signal and the read depth signal after smoothing and denoising to constrain the smoothing and denoising process and ensure that the main features of the original read depth signal are maintained during the smoothing and denoising process. The total variation regularization term is used to quantify the changes in the read depth signals between adjacent segmented regions by calculating the sum of the absolute values ​​of the differences between adjacent segmented regions in the read signal. Different optimization weights are assigned to different segmented regions according to the size of the difference in the read depth signals of adjacent segmented regions. That is, in represents the read depth signal of the segmented area after smoothing and noise reduction, represents the read depth signal of the segmented area before smoothing and noise reduction, that is, the original read depth signal, represents the regularization parameter, which is used to balance the read depth signal fidelity and smoothness, represents the number of segmented regions, It is the fidelity term of the total variation model, which prevents the smoothed and denoised read depth signal from deviating too much from the original read depth signal. Indicates the The segmented area and The degree of deviation of the read depth signal of each segmented region, represents the total variation regularization term, which penalizes the sharp fluctuations in the read depth signal, smoothes the read depth signal, and retains the true inflection points or mutations.

8. The copy number variation detection method based on semi-supervised learning according to claim 7, characterized in that: In step eight, based on the start and end positions of the copy number variation in the seed file, the corresponding segmented regions are labeled with pseudo labels. The seed file is constructed by the intersection of the detection results of multiple existing copy number variation detection tools; the segmented regions labeled with pseudo labels meet the following two screening conditions: the segmented regions labeled with pseudo labels belong to the intersection of the detection results of existing copy number variation detection tools, and the read depth signal corresponding to the segmented regions labeled with pseudo labels exceeds The principle is that if the read depth signal of the segmented region marked with a pseudo-label exceeds three standard deviations of the average read depth signal of all segmented regions, it is regarded as a high-confidence pseudo-label region, that is, in Indicates the The read depth signal of each segmented region, represents the average value of the read depth signal of all segmented regions, Represents the standard deviation of the read depth signal across all segmented regions.

9. The copy number variation detection method based on semi-supervised learning according to claim 8, characterized in that: In step nine, high-confidence pseudo-label regions are used as clustering cores, and the read depth signals of all segmented regions are used as clustering features. Cluster analysis is performed on all segmented regions based on the clustering core and the improved density clustering algorithm. The improved density clustering algorithm is based on the traditional density clustering algorithm. On the basis of the original two parameters of neighborhood radius and minimum number of neighbors, the z-score is introduced as the third constraint of the improved density clustering algorithm to quantify the degree of deviation of the read depth signal of a segment relative to the average value of the read depth signal of all segmented regions. Under the conditions of satisfying the neighborhood radius and the minimum number of neighbors, the segmented region with a z-score higher than the deviation threshold is marked as an abnormal segmented region through the clustering label, otherwise it is marked as a normal segmented region, where, For the The z-score of each segmented region, Indicates the The read depth signal of each segmented region, represents the average value of the read depth signal of all segmented regions, Represents the standard deviation of the read depth signal across all segmented regions.

10. The copy number variation detection method based on semi-supervised learning according to claim 9, characterized in that: In step 10, the abnormal segmentation regions screened out from the cluster analysis results are sorted by starting position. If the interval between two adjacent abnormal segmentation regions is very small, the two abnormal segmentation regions are merged, and the average values ​​of the read depth signal, GC content, and mapping quality signal are taken as the new features of the merged segmentation region. If the interval is large, the abnormal segmentation region is recorded as an independent variant region. For the merged segmentation region, if the read depth signal of the abnormal segmentation region is higher than the average value of the read depth signals of all merged segmentation regions, the variant type of the abnormal segmentation region is defined as a duplication, otherwise the variant type is defined as a deletion. The starting position, ending position, read depth signal, and variant type of the merged abnormal segmented region are integrated and saved as the detection results in a TXT file for output.

Images