A copy number variation breakpoint verification method based on a deep learning model
By employing a copy number mutation breakpoint verification method based on a deep learning model, this paper addresses the issues of poor robustness and high false positive rate of existing copy number mutation detection methods across platforms and sample types, achieving high accuracy and wide applicability in copy number mutation detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE FIRST AFFILIATED HOSPITAL OF MEDICAL COLLEGE OF XIAN JIAOTONG UNIV
- Filing Date
- 2026-04-20
- Publication Date
- 2026-07-14
AI Technical Summary
Existing copy number variation detection methods rely excessively on manually set fixed thresholds and specific experimental hypotheses, resulting in poor robustness, insufficient generalization ability, and high false positive rates across datasets, sequencing platforms, and sample types, which affects clinical interpretation and genetic counseling.
A copy number variant breakpoint verification method based on a deep learning model is adopted. By simulating genomic data with different allele fractions of variants, a high-quality standard set of copy number variants is generated. Then, a neural network model is used to train feature extraction and evaluation to improve the accuracy and generalization ability of detection.
It achieves high-accuracy copy number variation detection on different sequencing platforms and sample types, reduces false positive rate, simplifies the validation process, and improves the robustness and applicability of the detection.
Smart Images
Figure CN122392623A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of bioinformatics technology, and in particular to a method for verifying copy number variation breakpoints based on a deep learning model. Background Technology
[0002] Copy number variation (CNV) is a major class of structural variations characterized by the gain or loss of genomic DNA segments, typically spanning from thousands to trillions of bases. Numerous studies have demonstrated the important role of CNV in various genetic diseases and cancers. In genetic diseases, a single lineage of CNV can lead to developmental abnormalities and rare genetic syndromes, while in tumor genomes, CNV often manifests as somatic alterations during tumor development and progression. With the continuous development of next-generation sequencing technologies, copy number variation detection based on whole-genome sequencing (WGS) data has become a central analytical task in clinical and research settings.
[0003] In recent years, various computational algorithms, such as CNVkit, FACETS, and Sequenza, have been widely used in copy number variant analysis. These methods typically extract one or more signals from sequence alignment, such as read depth, paired end mapping, and split reads, and then model these signals to identify copy number variants. However, the short read lengths of NGS data and the abundance of repetitive sequences in the genome affect signal resolution, leading to significant differences in sensitivity, specificity, and breakpoint localization accuracy. Systematic biases (such as the GC effect) can distort read depth signals and hinder reliable copy number variant inference. These problems are more severe in tumor samples, where tumor purity variants, aneuploidy, and clonal heterogeneity often lead to an increase in false-positive copy number variants.
[0004] Furthermore, in practical applications, most existing copy number variant detection tools are designed for high sensitivity to capture a wide range of candidate events, which inevitably leads to an increase in the number of false positives. Therefore, inconsistent copy number variant calls and high false positive rates can propagate into downstream workflows, complicating variant assessment and posing significant challenges to clinical interpretation and genetic counseling.
[0005] Experimental validation methods such as quantitative PCR or fluorescence in situ hybridization can be used to validate copy number variations with high confidence, but they generally have low throughput and are time-consuming and labor-intensive, and cannot be extended to whole genome analysis.
[0006] Recent studies have explored post-processing and filtering strategies to improve the quality of copy number variant detection by introducing additional evidence or heuristic criteria. However, many existing methods rely on manually adjusted thresholds or specific assumptions, which leads to a lack of robustness and versatility when used across datasets and sequencing platforms. Summary of the Invention
[0007] To address the aforementioned shortcomings in existing technologies, this application provides a copy number mutation breakpoint verification method based on a deep learning model. This method solves the problem that existing copy number mutation verification methods rely excessively on manually set fixed thresholds and specific experimental hypotheses, resulting in poor robustness and insufficient generalization ability across datasets, sequencing platforms, and sample types.
[0008] To achieve the aforementioned objectives, the technical solution adopted in this application is as follows: This application provides a method for verifying copy number mutation breakpoints based on a deep learning model, including: S1: Simulate a genome containing copy number variations with different allele fractions as a tumor sample, simulate a normal control sample containing only single nucleotide polymorphisms, downsample the tumor sample and the normal control sample to obtain samples with different sequencing depths, and treat all simulated copy number variations as simulated positive variations. S2: For samples including tumor samples and normal control samples, establish a high-quality copy number variation standard set as true positive variations; S3: Perform random offsetting of artificial breakpoints on simulated positive variants and real positive variants to generate several copy number variants with inaccurate breakpoints, which are used as simulated negative variants; S4: Using the real genome of healthy human beings as tumor-normal control samples, several intervals are generated as pseudocopy number variations, which are then used as true negative variations; S5: Extract abnormal read pair information and sequencing depth information from simulated positive variants, real positive variants, simulated negative variants, and real negative variants, and convert them into digital features; S6: Input the digital features of simulated positive variants, real positive variants, simulated negative variants, and real negative variants into the neural network model for training; S7: Evaluate the new copy number mutation set based on the trained neural network model and obtain the evaluation results.
[0009] Further, S1 includes: S101: A variant detection format file based on single nucleotide polymorphisms in healthy human subjects, which simulates the generation of two haplotypes, including homozygous and heterozygous single nucleotide polymorphisms in healthy human subjects. S102: Insert the homozygous single nucleotide polymorphism and heterozygous single nucleotide polymorphism in the variant detection format file into two haplotypes respectively to simulate sequencing, and merge the sequencing results of the two haplotypes to obtain a normal control sample; S103: Several non-overlapping copy number variants are generated by random simulation and inserted into two haplotypes respectively for simulated sequencing. The sequencing results of the two haplotypes are then combined to obtain a tumor sample. S104: Tumor samples and normal control samples were downsampled and merged at different ratios to generate tumor samples with different allele fractions but the same sequencing depth. S105: Normal control samples and tumor samples with different allele fractions but the same sequencing depth were downsampled to generate samples with different sequencing depths, and copy number variations in these samples were used as simulated positive variations.
[0010] Further, S2 includes: S201: Several different tools for detecting structural variations based on third-generation sequencing data were used to detect variations in tumor samples and normal control samples, and then filtered. S202: Based on the filtered variant detection results, retain the variants that exist in tumor samples but not in normal control samples based on each tool as the basic somatic copy number variant set; S203: Merge the basic somatic copy number variant sets from multiple tools and perform manual verification to obtain a high-quality copy number variant standard set as the true positive variant.
[0011] Further, S4 includes: The genomes of twins from Chinese pedigrees were used as tumor samples and normal control samples, respectively, to simulate paired samples, and several intervals were randomly generated as true negative variants. Multiple second-generation methods for detecting copy number variations were used to detect the genomes of twins in Chinese families, and the detected results were also considered as true negative variations.
[0012] Further, S5 includes: For simulated positive variants, real positive variants, simulated negative variants, and real negative variants, a window is selected for tumor samples and normal control samples, centered on the left and right breakpoints of the interval. Within the window, the number of abnormal read pairs, the total number of read pairs, the number of split reads, and the total number of reads are counted. The proportion of abnormal read pairs to the total read pairs and the proportion of split reads to the total reads within the left and right breakpoint windows are calculated. For simulated positive variants, real positive variants, simulated negative variants, and real negative variants, a neighborhood is taken to the left of the left breakpoint and to the right of the right breakpoint for both tumor samples and normal control samples, and the ratio of the average sequencing depth of the interval to the average sequencing depth of the left and right neighborhoods of the interval is calculated.
[0013] Further, S6 includes: The numerical features of simulated positive variants, real positive variants, simulated negative variants, and real negative variants are robustly regularized, and the processed numerical features are used as input, with the classification of the three labels as output, to train the neural network model. The three labels are categorized as follows: whether the copy number mutation event is true, and whether the left and right breakpoints are accurate.
[0014] Further, S7 includes: Abnormal read pair information and sequencing depth information are extracted from the new copy number variant set and converted into digital features; The numerical features are input into the trained neural network model for evaluation, and the output includes evaluation results on whether the copy number mutation event is true and whether the left and right breakpoints are accurate.
[0015] The beneficial effects of this application are: This application provides a copy number variant breakpoint verification method based on a deep learning model. It collects features related to real copy number variant breakpoints in a posterior manner, while generating a rich set of negative variants that closely match actual data. Through training a machine learning model, the trained model has the ability to verify new copy number variant detection sets. It has a wide range of applications, is not limited by any sequencing platform, has strong transferability, and is easy to use. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.
[0017] Figure 1 This is a flowchart illustrating a copy number mutation breakpoint verification method based on a deep learning model, provided in an embodiment of this application.
[0018] Figure 2 This is another flowchart illustrating a copy number mutation breakpoint verification method based on a deep learning model, provided in an embodiment of this application.
[0019] Figure 3 This is a schematic diagram illustrating the training performance of a neural network model provided in an embodiment of this application.
[0020] Figure 4 This is a schematic diagram illustrating the evaluation of various methods on a simulated dataset using a neural network model provided in an embodiment of this application.
[0021] Figure 5This is a schematic diagram illustrating the evaluation of various methods on the HCC1395 dataset using a neural network model provided in an embodiment of this application. Detailed Implementation
[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.
[0023] Example 1: Existing methods for verifying copy number mutations include the following: Cytogenetic methods, such as fluorescence in situ hybridization (FISH) and stretched-fiber FISH, have low resolution and are time-consuming and labor-intensive, requiring individual experiments to verify their effectiveness.
[0024] Targeted analysis methods based on PCR technology, such as multiplex amplification probe hybridization (MAPH) and ligation-dependent multiplex probe amplification (MLPA), offer high resolution and low cost, but generally have low throughput.
[0025] Based on this, this application provides a method for verifying copy number mutation breakpoints based on a deep learning model. After training the model initially, this method only requires extracting breakpoint-related features from the reported copy number mutation set. The trained model can then be used to easily and conveniently verify the set. The method can be found in [reference needed]. Figure 1 and Figure 2 ,include: S1: A tumor sample is simulated by a genome containing copy number variations with different allele fractions, and a normal control sample is simulated by a sample containing only single nucleotide polymorphisms (SNPs). The tumor sample and the normal control sample are downsampled to obtain samples with different sequencing depths, and the CNVs are used as simulated positive variants.
[0026] Furthermore, S1 specifically includes: S101: VCF file, a variant detection format based on SNPs in healthy human subjects, uses existing tools to simulate and generate haplotypes, making them homozygous and heterozygous SNPs like those in healthy human subjects.
[0027] S102: Insert homozygous and heterozygous SNPs from the VCF file into two haplotypes respectively to simulate sequencing, and merge the sequencing results of the two haplotypes to obtain a normal control sample.
[0028] S103: Several non-overlapping copy number variations are generated by random simulation and inserted into two haplotypes respectively for simulated sequencing. The sequencing results of the two haplotypes are then merged to obtain a tumor sample.
[0029] S104: Tumor samples and normal control samples were downsampled at different ratios and merged to generate tumor samples with different allele fractions but the same sequencing depth.
[0030] S105: Normal control samples and tumor samples with different allele fractions but the same sequencing depth were downsampled to generate samples with different sequencing depths, and CNVs were used as simulated positive variants.
[0031] In one embodiment of this application, the VCF file of SNPs from healthy human subjects is downloaded, and the HACk command of the VISOR software is used to simulate two haplotypes. The homozygous and heterozygous SNPs in the VCF are inserted into the two haplotypes respectively. The two haplotypes are sequenced at 50x using Wgsim, and then aligned using BWA, and are denoted as hap1.bam and hap2.bam respectively. The samtools tool is then used to merge them to obtain a simulated 100x normal control sample.
[0032] Because copy number variants (CNVs) are generally large, the number of CNVs that can be inserted into a single genome is limited. Therefore, multiple independent beds containing CNVs of various sizes are generated. For the CNVs in each bed file, VISORHACk is used to seed a portion of the variant into haplotype 1 and another portion into haplotype 2. Sequencing is performed at 50x using Wgsim, and the results are compared and then merged to obtain tumor samples with a variant allele fraction of 0.5 and a sequencing multiplier of 100x. Assuming there are... m By generating multiple BED files, we obtain several tumor samples with different copy number variant sets, allele fractions of 0.5, and sequencing multipliers of 100x, denoted as cl1.bam, cl2.bam, ..., cl m .bam.
[0033] By downsampling tumor and normal controls at different ratios and then merging them, samples with different variant allele fractions but the same sequencing multiplier are generated. Variant allele fractions can be 0.3, 0.1, 0.05, 0.02, or 0.01, but are not limited to these values. For example, downsampling cl1.bam at 60x and normal at 40x results in a new sample with a variant allele fraction of 0.3 and a sequencing multiplier of 100x. Further downsampling yields samples with the same variant allele fraction but different sequencing multipliers; sequencing multipliers can be 80, 60, 40, 20, 10, or 5, but are not limited to these values.
[0034] S2: Collect some cell lines with high-quality copy number variation benchmarks as true positive variants, or establish a high-quality copy number variation benchmark for some samples including tumor-normal controls using third-generation sequencing data, which can also be used as true positive variants.
[0035] Furthermore, S2 specifically includes: S201: Several different tools for detecting structural variations based on third-generation sequencing data were used to detect variations in tumor samples and normal control samples, and then filtered.
[0036] S202: Based on the filtered variant detection results, retain variants that exist in tumor samples but not in normal control samples based on the same data from each tool, as the basic somatic copy number variant set.
[0037] S203: Merge somatic variant detection sets obtained from multiple methods and multiple datasets, and perform manual verification to obtain a high-quality copy number variant standard set as the true positive variant.
[0038] In one embodiment of this application, four tools—cuteSV, deBreak, Sniffles, and PBSV—were used to detect deletion and replication variants in third-generation sequencing data of tumor and control samples from the HCC1395 cell line. PBSV was only applicable to PacBio data, while the other three methods were used for both PacBio and ONT data. The detection results of seven pairs of tumor-normal control samples were filtered. Results with variant sizes greater than 1 kbp, a PASS status in the VCF, a support read count greater than a certain threshold (10 here), and not located in regions of low genomic alignment reliability such as centromeres and telomeres were selected. Variants present in tumor samples but not in normal controls were retained as the basic somatic copy number variant set. All results from the seven basic somatic copy number variant sets were sorted by chromosome and origin coordinates and merged according to the following strategy: if a variant appeared in two or more basic sets, with a length difference of less than 10% and overlapping lengths exceeding 90% of the smaller event, it was merged into one event, and the coordinates were averaged; otherwise, it was considered unreliable and discarded. The merged events were then manually examined using IGV to obtain the final high-confidence copy number variation standard set for the cell line, which included 138 copies and 111 deletions.
[0039] S3: For simulated positive variants and real positive variants, the left and right breakpoints are shifted by a large size, here 5kp, to represent copy number variants with inaccurate breakpoints, generating several copy number variants with inaccurate breakpoints as simulated negative variants.
[0040] S4: Using the real genome of a healthy human body as a tumor-normal control sample, several intervals are generated as pseudocopy number variations, which are then used as true negative variations.
[0041] In one embodiment of this application, one of the twin genomes in a Chinese family is treated as a tumor and the other as a normal control to simulate paired samples. Some intervals are randomly generated as true negative variants. In addition, the twin genomes are detected using a variety of second-generation methods for detecting copy number variations, and the detected results are also used as true negative variants.
[0042] S5: Extract abnormal read pair information and sequencing depth information from simulated positive variants, real positive variants, simulated negative variants, and real negative variants, and convert them into digital features.
[0043] In one embodiment of this application, for the above positive and negative variants, a window, such as a 1kbp window, is taken for the tumor sample centered at the left and right breakpoints of the interval. Within the window, the number of abnormal read pairs, the total number of read pairs, the number of split reads, and the total number of reads are counted. A normal read pair is considered to be aligned at both ends on the same chromosome with head-to-head orientation, and the insertion length falls within three standard deviations of the sample's average insertion length. The alignment quality at both ends reaches a certain threshold (here, 20), and the aligned length exceeds 30% of the read length. An abnormal read pair refers to an abnormal alignment direction or an abnormal insertion length (for copy number variants greater than 1.5kbp, the insertion length threshold is 1kbp; for copy number variants less than 1.5kbp, the insertion length threshold is the average insertion length plus three times the standard deviation). Valid split reads also require alignment quality to reach the threshold.
[0044] Calculate the proportion of abnormal read-pairs within the left and right breakpoint windows to the total read-pairs, and denote them as follows: and ; Calculate the proportion of the number of split-reads within the left and right breakpoint windows to the total number of reads, and denot them as follows: and For tumor samples, the average sequencing depth is calculated by taking a neighborhood to the left of the left breakpoint of the interval and another neighborhood to the right of the right breakpoint of the interval. The average sequencing depth within the interval is then calculated. Finally, the ratio of the average sequencing depth of the interval to the average sequencing depth of its left and right neighborhoods is calculated and denoted as... and .
[0045] S6: Input the digital features of simulated positive variants, real positive variants, simulated negative variants, and real negative variants into the neural network model for training.
[0046] In one embodiment of this application, a simple fully connected neural network is used for training the classification task. The extracted features are taken as input, and robust regularization is applied to the input. The output is a three-label classification: the first label represents whether the copy number mutation event is true, and the second and third labels represent whether the left and right breakpoints are accurate, respectively. For the tumor samples with different copy number mutations simulated in S1, one-third are used as the test set, and the remaining portion is further divided into an 80% training set and a 20% validation set. Notably, each tumor sample with different copy number mutations simulates different allele scores and sequencing multipliers. The trained model is used to evaluate the simulated copy number mutations with different allele scores and sequencing multipliers in the test set. The results are shown in […]. Figure 3 For copy number variants with an allele fraction of 0.1 or higher, the model can make effective judgments at very low sequencing coverage. For copy number variants with an allele fraction of 0.02, the model can only make effective judgments at sequencing coverage of 50x or higher. For even lower allele fractions, a sufficiently high sequencing coverage multiplier can only achieve limited performance.
[0047] S7: Evaluate the new copy number mutation based on the trained neural network model and obtain the evaluation results.
[0048] In one embodiment of this application, for any copy number mutation detection set, feature extraction is performed on the copy number mutation interval of each report according to S5, and further evaluation is performed using the model trained in S6. The output will be an evaluation of each copy number mutation in the detection set, including whether the copy number mutation event is true and whether the left and right breakpoints are accurate.
[0049] Alternatively, some features can be redefined, training samples can be generated, a deep learning model can be rebuilt and trained, and the new copy number mutation set can be evaluated based on the trained model.
[0050] A sample with a variant allele fraction of 0.3 and a sequencing multiplier of 100x in the test set was tested using FACETS, CNVkit, and Sequenza. The trained model was used to evaluate the results. Figure 4 Taking FACETS as an example, FACETS detected 343 CNVs. Comparing these with bedtools' intersect command and bedtools, with an overlap ratio of 0.8, 196 CNVs were correctly identified. Of the 343 detected CNVs, the model predicted 203 as true, and compared with bedtools, 189 were verified. This demonstrates that the trained model can effectively distinguish between true and false copy number variations.
[0051] In one embodiment of this application, a trained model is used to evaluate the copy number variation detection of HCC1395 using three methods: FACETS, CNVkit, and DRAGEN. Figure 5 As shown. Taking FACETS as an example, a total of 1987 were detected after filtering. Using the `intersect` command in bedtools and comparing with a high-quality benchmark, 94 could be mapped with an overlap ratio of 0.8. The model predicted 127 true values, of which 70 matched the benchmark. After applying the trained model, the detection accuracy of FACETS improved from 4.73% (94 / 1,987) to 55.12% (70 / 127). Analysis of the remaining 57 false positives revealed that IGV validation showed 30 were true copy number variants, 2 were germline copy number variants, and the remaining 25 were false copy number variants. However, most of these had one correct copy number variant breakpoint.
[0052] This application starts with positive and negative variants, extracts breakpoint-related features, and trains the model using a deep learning framework. This enables the model to learn the ability to identify true CNVs, allowing for easy validation on other CNV detection sets. This method is simple to use, effectively improves the accuracy of CNV detection sets, and significantly enhances the reliability of copy number variant (CNV) reports. For new CNV detection sets, regardless of the detection method, the trained model can be easily evaluated. Evaluations on simulated samples and real datasets demonstrate a significant improvement in CNV detection accuracy.
[0053] It should be noted that those skilled in the art will recognize that the embodiments described herein are for the purpose of helping readers understand the principles of this application, and should be understood as not limiting the scope of protection of this application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this application without departing from the essence of this application, and these modifications and combinations are still within the scope of protection of this application.
Claims
1. A copy number mutation breakpoint verification method based on a deep learning model, characterized in that, include: S1: Simulate a genome containing copy number variations with different allele fractions as a tumor sample, simulate a normal control sample containing only single nucleotide polymorphisms, downsample the tumor sample and the normal control sample to obtain samples with different sequencing depths, and treat all simulated copy number variations as simulated positive variations. S2: For samples including tumor samples and normal control samples, establish a high-quality copy number variation standard set as true positive variations; S3: Perform random offsetting of artificial breakpoints on simulated positive variants and real positive variants to generate several copy number variants with inaccurate breakpoints, which are used as simulated negative variants; S4: Using the real genome of healthy human beings as tumor-normal control samples, several intervals are generated as pseudocopy number variations, which are then used as true negative variations; S5: Extract abnormal read pair information and sequencing depth information from simulated positive variants, real positive variants, simulated negative variants, and real negative variants, and convert them into digital features; S6: Input the digital features of simulated positive variants, real positive variants, simulated negative variants, and real negative variants into the neural network model for training; S7: Evaluate the new copy number mutation set based on the trained neural network model and obtain the evaluation results.
2. The copy number mutation breakpoint verification method based on a deep learning model according to claim 1, characterized in that, S1 includes: S101: A variant detection format file based on single nucleotide polymorphisms in healthy human subjects, which simulates the generation of two haplotypes, including homozygous and heterozygous single nucleotide polymorphisms in healthy human subjects. S102: Insert the homozygous single nucleotide polymorphism and heterozygous single nucleotide polymorphism in the variant detection format file into two haplotypes respectively to simulate sequencing, and merge the sequencing results of the two haplotypes to obtain a normal control sample; S103: Several non-overlapping copy number variants are generated by random simulation and inserted into two haplotypes respectively to simulate sequencing. The sequencing results of the two haplotypes are then combined to obtain the tumor sample. S104: Tumor samples and normal control samples were downsampled and merged at different ratios to generate tumor samples with different allele fractions but the same sequencing depth. S105: Normal control samples and tumor samples with different allele fractions but the same sequencing depth were downsampled to generate samples with different sequencing depths, and copy number variations in these samples were used as simulated positive variations.
3. The copy number mutation breakpoint verification method based on a deep learning model according to claim 1, characterized in that, S2 includes: S201: Several different tools for detecting structural variations based on third-generation sequencing data were used to detect variations in tumor samples and normal control samples, and then filtered. S202: Based on the filtered variant detection results, retain the variants that exist in tumor samples but not in normal control samples for each tool as the basic somatic copy number variant set; S203: Merge the basic somatic copy number variant sets from multiple tools and perform manual verification to obtain a high-quality copy number variant standard set as the true positive variant.
4. The copy number mutation breakpoint verification method based on a deep learning model according to claim 1, characterized in that, The S4 includes: The genomes of twins from Chinese pedigrees were used as tumor samples and normal control samples, respectively, to simulate paired samples, and several intervals were randomly generated as true negative variants. Multiple second-generation methods for detecting copy number variations were used to detect the genomes of twins in Chinese families, and the detected results were also considered as true negative variations.
5. The copy number mutation breakpoint verification method based on a deep learning model according to claim 1, characterized in that, The S5 includes: For simulated positive variants, real positive variants, simulated negative variants, and real negative variants, a window is selected for tumor samples and normal control samples, centered on the left and right breakpoints of the interval. Within the window, the number of abnormal read pairs, the total number of read pairs, the number of split reads, and the total number of reads are counted. The proportion of abnormal read pairs to the total read pairs and the proportion of split reads to the total reads within the left and right breakpoint windows are calculated. For simulated positive variants, real positive variants, simulated negative variants, and real negative variants, a neighborhood is taken to the left of the left breakpoint and to the right of the right breakpoint for both tumor samples and normal control samples, and the ratio of the average sequencing depth of the interval to the average sequencing depth of the left and right neighborhoods of the interval is calculated.
6. The copy number mutation breakpoint verification method based on a deep learning model according to claim 1, characterized in that, The S6 includes: Robust regularization is applied to the numerical features of simulated positive variants, real positive variants, simulated negative variants, and real negative variants. The processed numerical features are used as input, and the classification of the three labels is used as output to train the neural network model. The three labels are categorized as follows: whether the copy number mutation event is true, and whether the left and right breakpoints are accurate.
7. The copy number mutation breakpoint verification method based on a deep learning model according to claim 1, characterized in that, The S7 includes: Abnormal read pair information and sequencing depth information are extracted from the new copy number variant set and converted into digital features; The numerical features are input into the trained neural network model for evaluation, and the output includes evaluation results on whether the copy number mutation event is true and whether the left and right breakpoints are accurate.