CN113963746B - A deep learning-based genome structural variation detection system and method - Google Patents

Abstract

The invention provides a genome structure variation detection system and method based on deep learning, wherein the method mainly comprises four steps: (1) Extracting a structural variation characteristic sequence based on the existing sequence alignment technology; (2) Coding the structural variation feature sequence similarity image by utilizing the RGB image; (3) Predicting structural variation contained in the structural variation feature sequence similarity image by utilizing a multi-target recognition framework; (4) The complex structure variant type is systematically represented through the graph data structure. The invention realizes the simultaneous detection of simple and complex structural variations from sequence difference encoded images without relying on any structural variation model.

Description

A deep learning-based genome structural variation detection system and method

Technical field

The invention belongs to the technical field of precision medicine and relates to a genome structure variation detection system and method based on deep learning.

Background technique

In the past decade, large-scale international cooperation projects based on second-generation sequencing data, such as TCGA, ICGC and HGSVC, have continuously revealed the differences in genomic structural variation among populations and their relationship with the occurrence of genetic diseases, tumors and other diseases. close relationship between them. In the past five years, with the development and increasing popularity of third-generation long-read sequencing, the number of known structural variations in human germ cells has been 2.5 times the number of structural variations detected by second-generation sequencing. These structural variations provide the basis for subsequent evolution. and provide an important foundation for research on related diseases. What's more important is that more and more simple structural variations are found to be complex structural variations through further analysis. For example, in 2015, "Nature" comprehensively introduced complex structural variations in the genome for the first time. The particularity of complex structural variations is first manifested in a completely different formation method from simple structural variations. As an unexplored part of the genome, they provide new evidence for researchers to study the genome damage repair mechanism. On the other hand, complex structural variation has a strong correlation with genetic diseases and developmental diseases. Related research has greatly enriched researchers' understanding of complex structural variation, such as the research published in "Genome Biology" in 2017. , discovered 16 different complex mutation types and conducted an in-depth analysis of their role in the development of autism. However, these complex structural variations that occur in germ cells often cannot be detected by existing traditional clinical methods. At the same time, due to the complexity of the structural variations themselves and the limitations of existing detection methods, methods based on third-generation sequencing cannot accurately detect Detect these complex events. Compared with germ cell structural variations, tumor genomes have experienced multiple and rapid selections, so there are more large-scale complex structural variations, such as chromothripsis and chromoplexy. These complex structural variations are considered to be events that occur rapidly in a short period of time during tumor development and contribute to tumor development to a great extent.

Generally speaking, with the overall concept of "big health" being proposed, and the problem of my country's aging population gradually becoming more and more prominent, the incidence of genetic diseases, developmental diseases, and cancer is getting higher and higher. Therefore, as the price of third-generation sequencing data increases, As the number continues to decline, whole-genome testing based on third-generation sequencing technology will become an inevitable trend in clinical diagnosis. In this context, a large amount of sequencing data will be generated, and the interpretation of these data, especially data related to clinical diseases, will become the key to restricting the development of the entire industry.

The current main steps for whole-genome structural variant detection based on third-generation sequencing technology still continue the detection theory based on second-generation sequencing data, which mainly includes three steps: (1) establishing a model of known genome structural variants; ( 2) Infer the abnormal alignment features that the model may reflect in the sequencing data comparison results; (3) Based on the constructed abnormal alignment feature models of different structural variation types, match the sequencing read comparison results, and finally get the detection result. Detection tools developed based on the above detection ideas, such as PBSV, Sniffles, SVIM, NanoVar, CuteSV and other methods, have been widely used to detect germline genome structural variations, as well as analyze a small number of disease and tumor samples. In order to detect complex structural variations, most detection tools use patching, that is, abnormal alignment models corresponding to new structural variation types are added to the original tools. The most representative of these is Sniffles, which is the first detection tool to detect two complex structural variant types by adding additional abnormal alignment models. However, with the development of sequencing technology, researchers' understanding of genome structural variation is still the tip of the iceberg. This method of detecting structural variation through patching only treats the symptoms but not the root cause, and still cannot explore the unknown types of structural variation that exist in the genome. On the other hand, tools developed based on modeling ideas have to write specific codes for each mutation type, so the code of such tools is particularly complex and has poor readability, which directly leads to low computational efficiency and difficulty in maintenance. This is mainly due to the complexity of the abnormal alignment characteristics of complex structural variants, which results in widely varying detection sensitivities for different size ranges and different variant types. For example, as shown in Figure 1, for simple deletion variants and deletion inversion complex structural variants, now Some tools will detect complex structural variations as individual deletions or inversions, and some will even miss this event. In the past two or three years, as more and more complex structural variations have been discovered through tedious manual analysis, biomedical researchers have gradually realized that complex structural variations play an important role in some undiagnosed diseases; at the same time, in order to To achieve better all-round structural variation detection results, the new detection system is a key technology to promote future clinical detection. In addition to model limitations, repetitive sequences have long been a key factor affecting structural variant detection, and there is still no effective solution. In addition, the characterization methods of complex structural variations have not been unified. Different studies mostly use a combination of simple variations to characterize the types of complex structural variations, and at the same time match detailed text explanations. The biggest problem with this method is that it is not conducive to comparison and detection between different studies. complex structural variations.

In summary, despite nearly 10 years of development, researchers have used genome sequencing data to detect simple types of mutations and applied this information to study human evolution, population migration and fusion, disease mechanisms and treatment options, which has greatly Promoted the development of biomedicine. However, this structural variation detection theory based on modeling and patching is no longer able to meet the needs of future scientific research, hospitals and genetic testing service providers for variation detection, especially cannot support the transition from targeted testing to whole-genome testing.

Contents of the invention

In view of the problems existing in the existing genome-wide structural variation detection technology, the present invention proposes a genome structure variation detection system and method based on deep learning, which realizes the simultaneous detection of simple and complex sequences from sequence difference encoding images without relying on any structural variation model. Structural variation.

In order to achieve the above objects, the technical solutions of the present invention are as follows:

A method for detecting genomic structural variation based on deep learning, including:

Step 1. Structural variation feature sequence extraction: compare the sample sequence with the reference genome sequence to obtain the global alignment result, and extract the structural variation feature sequence based on the global alignment result; the matching fragments in the structural variation feature sequence are called main fragments; According to the alignment characteristics of the structural variation characteristic sequence, the sequence of the unmatched fragment in the structural variation characteristic sequence is subjected to local Kmer re-alignment with the reference genome sequence. The matching fragment obtained after the local Kmer re-alignment is called a secondary fragment;

Step 2. Structural variation characteristic sequence similarity image coding: Use RGB image three-channel coding method, combine the main fragments and secondary fragments, encode the structural variation characteristic sequence and the reference genome sequence, and obtain the similarity between the reference genome sequence and the sample sequence. RGB image, while encoding the reference genome sequence to obtain the similarity image of the reference genome sequence itself; subtracting the two images to obtain the structural variation characteristic sequence similarity image;

Step 3. Segmentation of the structural variation feature sequence similarity image: According to the order of the main fragments on the reference genome sequence, combine the two adjacent main fragments in the structural variation feature sequence similarity image to obtain a sub-image containing only a single structural variation. ;According to the order of the major fragments and minor fragments on the reference genome sequence, combine adjacent major fragments and minor fragments in pairs in the sub-image to obtain the fragment of interest;

Step 4. Identification of structural variation feature sequence similarity images and structural variation characterization: Use the pre-trained structural variation detection CNN model to identify all the fragments of interest in the sub-image containing a single structural variation to obtain complex structural variation fragments; Systematic characterization and classification of complex structural variant fragments using graph data structures.

Preferably, in step 1, according to the alignment characteristics of the structural variation characteristic sequence, perform a local Kmer re-alignment of the sequence of the unmatched fragment in the structural variation characteristic sequence with the reference genome sequence, specifically: according to the CIGAR characters of the structural variation characteristic sequence , extract the fragments that do not match the reference genome sequence, perform local Kmer re-alignment of the sequence of the mismatched fragments with the reference genome sequence, and obtain the secondary fragments;

Preferably, step 2 specifically includes:

1) RGB three-channel sequence similarity encoding: encode the structural variation characteristic sequence and the reference genome sequence into the sequence matching channel (255, 0, 0), the sequence repetition channel (0, 0, 255) and the sequence inversion channel (0, 255,0), output the RGB image of the similarity between the reference genome sequence and the sample sequence; encode the reference genome sequence into the sequence matching channel (255,0,0), the sequence repeat channel (0,0,255) and the sequence inversion In channel (0, 255, 0), the similarity image of the reference genome sequence is obtained, and the position information of each matching fragment is recorded at the same time;

2) Remove reference genome sequence repetitive fragments: According to the RGB image of the similarity between the reference genome sequence and the sample sequence and the coordinate position of the matching fragment in the reference genome sequence self-similarity image relative to the reference genome sequence, in the reference genome sequence self-similarity image Find the fragments corresponding to the fragments in the RGB image of the similarity between the reference genome sequence and the sample sequence. If the corresponding fragment is found, remove it from the RGB image of the similarity between the reference genome sequence and the sample sequence to obtain the structural variation characteristic sequence. Similarity images.

Preferably, step 3 specifically includes:

1) Single structural variation segmentation: According to the order of the main fragments on the reference genome sequence, two adjacent main fragments are combined in the structural variation characteristic sequence similarity image to obtain a sub-image containing only a single structural variation;

2) Multi-target segmentation of structural variation images: Sort the main fragments and secondary fragments according to their coordinates on the structural variation feature sequence, combine all fragments in the sub-image in pairs according to the sorting results, and use the combination of the main fragments and secondary fragments as Fragments of interest.

Preferably, the CNN model training method for structural variation detection described in step 4 specifically includes:

1) Construct a structural variation training data set: The real data uses the structural variation characteristic sequences of 2500 samples in the 1000Genome Project as training data, and the virtual data uses VISOR virtual noise-free training samples as training data. The two together form a training data set. ;

2) Training data set encoding: Use the method described in step 2 to encode the training data in the training data set to obtain the interesting fragments of the training data set;

3) Model training: Input the training data set into the convolutional neural network, train the convolutional neural network, and obtain the structural variation detection CNN model after the training is completed.

Preferably, in step 4, a pre-trained structural variation detection CNN model is used to identify all the fragments of interest in the sub-image containing a single structural variation to obtain complex structural variation fragments; a graph data structure is used for complex structural variation fragments. Systematic characterization and classification, specifically:

The structural variation detection CNN model identifies the fragments of interest in the sub-image to obtain complex structural variation fragments, constructs a structural variation representation map based on the complex structural variation fragments, and calculates whether different structural variations belong to the same type based on the topology of the structural variation representation map. Each node in the structural variation representation graph is a fragment contained in all the fragments of interest in the sub-image, and each edge connects two consecutive fragments on the sample sequence.

A deep learning-based genome structural variation detection system, including:

The structural variation characteristic sequence extraction module is used to compare the sample sequence with the reference genome sequence to obtain the global comparison result, and extract the structural variation characteristic sequence based on the global comparison result; the matching fragment in the structural variation characteristic sequence is called the main fragment; According to the alignment characteristics of the structural variation characteristic sequence, the sequence of the unmatched fragment in the structural variation characteristic sequence is subjected to local Kmer re-alignment with the reference genome sequence. The matching fragment obtained after the local Kmer re-alignment is called a secondary fragment;

The structural variation characteristic sequence encoding module is used to encode the structural variation characteristic sequence and the reference genome sequence using the RGB image three-channel encoding method to obtain an RGB image of similarity between the reference genome sequence and the sample sequence; at the same time, the reference genome sequence is encoded to obtain the reference Genome sequence self-similarity image; two images are subtracted to obtain a sequence similarity image of structural variation characteristics;

The structural variation characteristic sequence similarity image segmentation module is used to combine the two adjacent main fragments in the structural variation characteristic sequence similarity image according to the order of the main fragments on the reference genome sequence to obtain a sub-image containing only one structural variation; According to the order of the major fragments and minor fragments on the reference genome sequence, the adjacent major fragments and minor fragments are combined in the sub-image to obtain the fragment of interest;

The structural variation identification and characterization module uses a pre-trained structural variation detection CNN model to identify all interesting fragments in sub-images containing a single structural variation to obtain complex structural variation fragments; the graph data structure is used to perform complex structural variation fragments Systematic characterization and classification.

Preferably, the structural variation characteristic sequence coding module includes:

The RGB three-channel sequence similarity coding module encodes the structural variation characteristic sequence and the reference genome sequence into the sequence matching channel (255, 0, 0), the sequence repeat channel (0, 0, 255) and the sequence inversion channel (0, 255 , 0), output the RGB image of the similarity between the reference genome sequence and the sample sequence; encode the reference genome sequence into the sequence matching channel (255, 0, 0), the sequence repeat channel (0, 0, 255) and the sequence inversion channel (0, 255, 0), obtain the similarity image of the reference genome sequence itself, and record the position information of each matching fragment;

Remove the reference genome sequence repetitive fragment module, and based on the RGB image of the similarity between the reference genome sequence and the sample sequence and the coordinate position of the matching fragment in the reference genome sequence self-similarity image relative to the reference genome sequence, search in the reference genome sequence self-similarity image Fragments corresponding to the fragments in the RGB image of the similarity between the reference genome sequence and the sample sequence. If the corresponding fragment is found, it is removed from the RGB image of the similarity between the reference genome sequence and the sample sequence, and the structural variation characteristic sequence similarity is obtained. Sexual images.

Preferably, the structural variation feature sequence similarity image segmentation module includes:

A single structural variation segmentation module is used to combine the two adjacent main fragments in the structural variation characteristic sequence similarity image according to the order of the main fragments on the reference genome sequence to obtain a sub-image containing only one structural variation;

The structural variation image segmentation module is used to sort the main fragments and the secondary fragments according to the coordinates on the structural variation feature sequence. According to the sorting results, the two adjacent fragments are combined to obtain the combined fragment. The filter consists of two secondary fragments and two A combined segment formed by combining linear segments, and a combined segment composed of a primary segment and a secondary segment is used as the segment of interest.

Preferably, the structural variation identification and characterization module specifically includes:

Construct a structural variation training data set module to combine real data and virtual data into a training data set. The real data uses the structural variation of 2500 samples in the 1000Genome Project as training data, and the virtual data uses VISOR virtual training data without noise interference;

The training data set encoding module encodes all the training data in the training data set according to the structural variation feature sequence encoding module to obtain the interesting fragments of the training data set;

The convolutional neural network training module is used to input the training data set into the convolutional neural network, train the convolutional neural network, and obtain the structural variation detection CNN model after the training is completed;

The structural variation graph representation module is used to identify the fragments of interest in the sub-image through the structural variation detection CNN model, obtain the complex structural variation fragments, construct the structural variation representation graph based on the complex structural variation fragments, and calculate the topology structure based on the structural variation representation graph. Whether different structural variations belong to the same type, each node in the structural variation representation graph is a fragment contained in all the fragments of interest in the sub-image, and each edge connects two consecutive fragments on the sample sequence.

Compared with the existing technology, the present invention has the following beneficial technical effects:

The present invention proposes for the first time a model-independent genome structural variation detection method for long-read sequencing. This method mainly includes four steps (Figure 2): (1) Based on existing sequence comparison technology, extract structural variation characteristic sequences; ( 2) Use RGB images to encode structural variation feature sequence similarity images; (3) Use a multi-objective recognition framework to predict the structural variations contained in structural variation feature sequence similarity images; (4) Systematically represent complex structural variations through graph data structures type. The present invention innovatively optimizes the existing image coding technology and multi-target recognition frame. In step (2), the optimized image coding technology is used to code the similarity between the reference genome sequence and the sample sequence, and in step (3), the similarity based on The multi-objective recognition framework of convolutional neural network (CNN) detects structural variations of different levels of complexity carried in sample sequences from structural variation feature sequence similarity images. First, a three-channel (RGB) encoding scheme is used to encode the similarity RGB image of the reference genome sequence (REF) and the sample sequence (ALT), which is defined as a REF-to-ALT image. Secondly, in order to remove the impact of background noise (such as short tandem repeats, etc.) in the reference genome sequence on structural variation detection, the reference genome sequence's own similarity image (REF-to-REF image) was also encoded, and then the image subtraction method was used Remove the influence of background noise, especially to avoid the influence of detecting complex structural variations. In the multi-target recognition framework based on CNN, the present invention proposes a two-step image segmentation method. The first step is mainly used to split multiple structural variations contained in a structural variation characteristic sequence similarity image into sub-images containing only one structural variation. The sub-image containing one structural variation is defined as a similarity image ( Similarity Image (SI); the second step is to segment the similarity image into segments of interest (SOI), and finally use the pre-trained CNN model to identify all SOIs contained in the SI and obtain the final prediction result. The image encoding method and multi-target recognition framework used in the present invention have good scalability, so that they can be applied to various long-read sequencing data, especially data generated by PacBio and Oxford Nanopore sequencing platforms; at the same time, the The invention can be applied to contig sequences assembled based on long-read sequencing. In addition to complex structural variation detection, different related studies in the past have used different naming methods for the same complex structural variation types. These naming methods often depend on the subjective understanding of the researchers themselves, and there is a lack of systematic and unified naming methods, which also hinders the development of complex structural variation. Variation-related research. Another major innovation point of the present invention is that it is proposed for the first time that graph data structures can be used to represent complex structural variations, and different types of complex structural variations can be classified by calculating the topological structure similarity of graphs. Based on the above characteristics, the present invention can greatly improve the detection rate and accuracy of structural variations of different levels of complexity, further promote disease diagnosis and early screening based on third-generation test data, and provide an efficient and reliable detection system for related fields. On the other hand, from the perspective of identifying characteristic sequences of structural variation, the present invention is more conducive to detecting structural variation from different clones in tumors, and provides a new detection system for studying the role of structural variation in tumor evolution.

Furthermore, due to the low detection rate of complex structural variation in the past, its characterization and comparison methods have always relied on artificial definitions by relevant researchers. While improving the detection rate of complex structural variation, the present invention proposes a graph-based complex structural variation The essence of the representation method is to construct the connection relationship between different SOIs relative to the reference genome sequence. The nodes in the graph are the fragments contained in the SOI, and the edges connect two consecutive fragments on the sample sequence.

The invention is a genome structure variation detection system based on multi-target recognition. It takes the model-independent structural variation detection theory as the core and uses a structural variation feature sequence extraction module, a structural variation feature sequence similarity image coding module, and a structural variation feature sequence. The similarity image segmentation module and the structural variant identification and characterization module realize structural variant detection for long-read sequencing without relying on any mutation model. Sequence similarity encoding and multi-target recognition based on RGB images capture two fundamental characteristics of structural variation detection. First, structural variation is manifested as the difference between the reference genome sequence and sample sequence. Second, complex structural variation is manifested as a variety of simple A combination of structural variations; then, the present invention achieves the detection of structural variations of different levels of complexity by identifying SOI in similarity images. The present invention does not rely on any mutation model. For the detection of known simple structural variations, its sensitivity is higher than the existing model-based detection methods. On the premise of a low false positive rate, the present invention greatly improves the detection of complex structural variations. Out rate.

In summary, the genome structure variation detection system involved in the present invention is the core technology to achieve accurate diagnosis. At the same time, it seizes the major opportunities for the development of precision medicine brought by third-generation sequencing technology and innovatively proposes from a technical perspective. Model-independent structural variation detection theory, and a new structural variation detection system was designed and implemented based on this theory. This system not only ensures high detection rates and accuracy of simple structural variants, but also greatly improves the detection rate of complex structural variants, providing important technical support for promoting the clinical application of third-generation sequencing technology. The present invention is oriented to the country's major needs and studies the core issues in the national strategic emerging industry "precision medicine", which is beneficial to our country in breaking the situation where major key core technologies are controlled by others in the strategic field of genome structural variation detection, and is more conducive to Open up new development directions for "precision medicine" related industries and cultivate new economic growth points.

Description of drawings

Figure 1 shows models of simple and complex structural variations constructed with long-read sequencing technology;

Figure 2 shows the process of a deep learning-based genomic structural variation detection system for long reads and assembled sequences;

Figure 3 shows a comparison of simple structural variation detection results for different long sequencing platforms;

Figure 4 is a comparison of detection structures for virtual complex structural variations;

Figure 5 shows a comparison of the detection results of structural variants of different sizes for high-fidelity (HiFi) sequencing;

Figure 6 is a comparison of the detection results of the present invention for sequencing reads and assembled contigs;

Figure 7 shows a comparison of structural variation detection results for genotyped genomes.

Detailed ways

The present invention will be further described in detail below with reference to specific examples, which are explanations rather than limitations of the present invention.

The present invention proposes a new model-independent genome structure variation detection theory, and designs a genome structure variation detection system for long-read sequencing based on the new theory.

The genome structural variation detection system and method based on deep learning proposed by the present invention is specifically stated as follows: for various long-read sequencing technologies, the fundamental characteristics of structural variation are expressed as sample sequences (including sequencing reads or assembly sequences) and the reference genome sequence. Therefore, the present invention encodes the differences between the sample sequence and the reference genome sequence through images, uses a multi-objective recognition framework to detect structural variations of different levels of complexity in the reference genome sequence and sample sequence similarity images, and then uses the data structure of the graph Systematically characterize and classify different types of complex structural variations, ultimately achieving the purpose of structural variation detection and characterization. The core of the algorithm designed based on this theory mainly includes: (1) Structural variation feature sequence extraction; (2) Structural variation feature sequence similarity image coding; (3) Structural variation feature sequence similarity image segmentation; (4) Structural variation feature sequence Identification and characterization of structural variations.

The core method involved in the present invention mainly includes the following steps:

Step 1. Structural variation feature sequence extraction: Based on the sequencing data of a certain sample, locate the area where potential structural variation exists, and extract structural variation feature sequences that support structural variation; specifically:

Using existing technology, the sample sequence is compared with the reference genome sequence for whole genome comparison. After obtaining the alignment results, extract the abnormal alignment reads, and at the same time extract the read sequences spanning the structural variation site from these reads, and use the read sequences as structural variation characteristic sequences through sequence assembly or directly. Existing alignment technologies usually provide two main abnormal alignment read characteristics: (1) insertions and deletions of CIGAR characters in the alignment results; (2) supplementary alignment features retained in the alignment results.

Kmer-based variation feature sequence re-alignment: In step 1, the CIGAR characters of the structural variation feature sequence obtained based on the existing technology will record the matching fragments and non-matching fragments between the sample sequence and the reference genome sequence. This result is considered global Comparison results. First, the unmatched fragment sequence is extracted from the global alignment results. Secondly, the sequence of the unmatched fragment is re-aligned with the reference genome sequence and the new alignment result is recorded. During the alignment process, whether there are duplications in the fragment and the comparison will be recorded simultaneously. The alignment directions are the 5' to 3' end and the 3' to 5' end of the DNA respectively. In the next step, the matching fragment obtained from the CIGAR characters is defined as the primary fragment, and the re-aligned fragment obtained through re-alignment is defined as the secondary fragment. All primary fragments and secondary fragments are collectively referred to as the matching fragments of the sample sequence and the reference genome sequence. , each fragment contains three characteristics: alignment position, alignment direction and whether it is a repeated sequence.

Step 2. Structural variation feature sequence similarity image encoding: Based on the structural variation feature sequence obtained in step 1, the similarity between the structural variation feature sequence and the reference sequence is encoded through the comparison results of step 1 and the results based on Kmer re-alignment RGB image; at the same time, use the same encoding scheme to encode the similarity REF-to-REF image of the reference genome sequence itself, and then obtain the denoised REF-to-ALT image through image subtraction.

Step 2 specifically includes:

1) Coding of RGB three-channel sequence alignment results: Based on the three characteristics of each fragment, encode all the alignment results obtained in step 1 into the sequence matching channel (255, 0, 0). The sequence In the repeat channel (0, 0, 255) and the sequence inversion channel (0, 255, 0), each channel can encode sequences of different lengths by setting the minimum fragment threshold. Each pixel in the image can achieve a single Base resolution means that each pixel represents a pair of bases that is the same between the reference genome sequence and the sample sequence. Through the above three-channel encoding method, this step will output the RGB image of the similarity between the reference genome sequence and the sample sequence and the position information of each fragment in the image; at the same time, the re-alignment method in step 1 is called again to compare the reference genome sequence with Perform the alignment by itself. During the alignment process, the Kmer is continuously moved to obtain a single-base resolution alignment result, and then the same three-channel encoding method is used to obtain the REF-to-REF image.

2) RGB image denoising based on the similarity between the reference genome sequence and the sample sequence: Based on the similarity RGB image and REF-to-REF image obtained between the reference genome sequence and the sample sequence in 1), first extract the relative values of each fragment in the two images. Based on the start and end coordinates of the reference genome sequence, the fragments extracted from the two images are arranged in the order of the coordinates of the reference genome sequence. Secondly, the fragments extracted from the RGB image are selected for similarity between the reference genome sequence and the sample sequence, and their similarities with each other are detected one by one. The detection of overlap of segments in REF-to-REF images can be completed in O(logn) time complexity. After obtaining the overlap of the fragments, it is judged according to the preset threshold whether the fragment should be removed from the RGB image of the similarity between the reference genome sequence and the sample sequence. Through the above method, the two images are subtracted to obtain the structural variation characteristic sequence similarity image.

Step 3. Structural variation feature sequence similarity image segmentation: Based on the structural variation feature sequence similarity image obtained in step 2, first segment the structural variation feature sequence similarity image into a single structure based on the arrangement and combination of the main fragments in the image. Mutated similarity image (Similarity Image, SI); on the basis of SI, the similarity image is divided into different segments of interest (Segment of Interest, hereinafter referred to as SOI) again based on the combination of primary fragments and secondary fragments;

Step 3 specifically includes:

1) Single structural variation segmentation: used to separate multiple different structural variations appearing in the structural variation characteristic sequence similarity image to avoid predicting and characterizing continuous simple structural variations as complex structural variations. The main method is to first arrange the main fragments in the sequence similarity image of structural variation characteristics in order, and then combine the main fragments in pairs in order. At the same time, according to the coordinate position of the main fragments, the image containing the main fragments is changed from the original structural variation characteristic sequence. It is extracted from the similarity image, and finally a similarity image containing a single structural variation is obtained;

2) Multi-target segmentation of structural variation similarity images: Segmentation from the similarity images is performed to obtain multiple SOIs to be identified. The main steps are to first sort the major fragments and minor fragments in the similarity image according to the order of the reference genome sequence, and secondly to combine the two adjacent fragments according to the sorting results and extract the sub-image containing the two fragments. , called SOI. In the extraction process, the SOI composed of two secondary segments is first filtered; secondly, it is judged whether the two segments can form a linear segment. This process mainly checks whether the slope difference of the two segments in the two-dimensional space is within Within a fixed range, the slope difference is mainly calculated by the minimum structural variation size that the user wants to detect; at the same time, due to the different dimensions of each similarity image, the slope difference range is also different in each image. After the above two filtering steps, the SOI in each similarity image will be used for subsequent recognition;

Step 4. Identification of structural variation feature sequences and structural variation characterization: Based on multiple targets (i.e. SOI) in a similar image obtained in step 3, use a pre-trained convolutional neural network (Convolution NeuralNetwork, hereinafter referred to as CNN) Identify structural variations of different complexities in similarity images; after the identification is completed, use graphs to systematically characterize and classify different types of complex structural variations.

Step 4 specifically includes:

1) Construct structural variation training data set and encoding: Structural variation used for training mainly includes four simple structural variation types, deletion (DEL), insertion (INS), inversion (INV) and duplication (DUP). The real data uses the structural variation of 2500 samples in 1000GenomeProject as training data, and the virtual data uses VISOR virtual noise-free training samples. The two together form the training data set. Secondly, encode the training data set according to the method described in steps 2 and 3;

2) Convolutional neural network (CNN) training: Input the encoded training set in 1) into the convolutional neural network, and use existing technology to train the AlexNet neural network. AlexNet neural network training parameters use transfer learning method, use the best score parameters in the GoogleImageNet competition as model initialization parameters, and use cross-validation method to select the optimal model as the final structural variation detection model.

3) Construct a complex structural variation representation map: Use the structural variation detection CNN model to identify multiple SOIs in the sub-image. First, obtain the complex structural variation representation map based on the recognition results. Secondly, calculate whether different complex structural variation representations are based on the topology of the structural variation representation map. Belong to the same type.

Construct a structural variation representation map: First, obtain matching fragments from all SOIs in the similarity image. Each fragment contains start and end coordinates, direction, and repeated fragment source information. These fragments are uniformly used as each segment in the structural variation representation map. Nodes, each node has a direction; secondly, each edge in the graph connects two adjacent fragments on the sample sequence. At the same time, for repeated fragments whose sources are known, repeated edges are used in the graph to connect the fragment and its source.

Structural variation representation graph topological structure similarity calculation: first calculate whether the number of nodes and edges of the two graphs match; secondly, for two graphs with the same number of nodes and edges, obtain each graph based on the edge connection relationship and node direction. The 5' to 3' end path and the 3' to 5' end path of each graph. If there are symmetric paths in two structural variation representation graphs, the two structural variations are considered to be of the same type.

The specific implementation process of the present invention is as follows, and the flow process is shown in Figure 2:

Step 1: Use existing technology to compare the long-read sample sequencing data (sample sequence) with the reference genome sequence, determine the coordinates of each read sequence on the reference genome, and extract the sample sequence and reference from the comparison result. For the mismatched parts of the genome sequence, read sequences with abnormal alignment characteristics are obtained, and the read sequences are used through sequence assembly or directly used as structural variation characteristic sequences.

After several years of development, the research on the long-read comparison in step 1 has been relatively thorough, and it is usually completed using seed-and-extension combined with dynamic programming. The comparison steps include seed generation and expansion, and expansion is mainly implemented using dynamic programming. Mainstream representative tools include minimap2, NGMLR and pbmm2.

After obtaining the alignment results, the key step to extract the structural variation characteristic sequence is to find the abnormal alignment sites. If the sample sequence does not contain any form of structural variation at a certain site, then the read sequence at this site is no different from the reference genome sequence, and the alignment direction is also normal. Otherwise, the read sequence at this site will not be aligned to the reference genome sequence. Read sequences may carry variant signatures. The variation characteristics left by the structural variation involved in the present invention mainly include: (1) Unmatched sequence (unmatched). If the sample sequence has a short sequence insertion or deletion at a certain position relative to the reference genome sequence, the comparison will Multiple read sequences at this site will have the characteristics of unmatched sequences, and their main manifestations are insertions (I) and deletions (D) in the alignment results. Here, the present invention only considers unmatched sequences with long reads greater than 50. For the unmatched sequences in (1), the present invention uses a Kmer-based local hash comparison method to refine the source of the unmatched sequence inserts, and at the same time discover variation characteristics other than insertions (I) and deletions (D). (2) Supplementary alignment. Since there is structural variation at a certain site in the sample, when the read alignment passes through this site, the alignment software will break a long read sequence into multiple subsequences. The subsequences are compared with the reference genome sequence to obtain supplementary alignment results. However, due to the incompleteness of the reference genome itself, the present invention does not process reads that generate 4 or more supplementary alignments by default. This parameter can be adjusted according to user needs.

Step 2, use the three-channel RGB image to encode the similarity between the structural variation characteristic sequence and the reference genome sequence:

After obtaining the structural variation feature sequence through the above features, the present invention converts it into fragment features based on the comparison results of each structural variation feature sequence, including matching fragments and non-matching fragments, and at the same time clusters fragment features with the same attributes. For the fragment characteristics in each cluster, the present invention encodes the structural variation characteristic sequence CIGAR characters, Kmer alignment results and reference genome sequence in step 1 into the sequence matching channel (255, 0, 0) and the sequence repetition channel (0, 0, 255) and the sequence inversion channel (0, 255, 0), each channel can encode different sequences by setting the minimum fragment threshold. Through the above three-channel encoding method, this step will output the sample sequence and the reference genome sequence. The RGB coding matrix and the position information of each matching fragment; according to the RGB coding matrix of the sample sequence and the reference genome sequence, an RGB image of the similarity between the reference genome sequence and the sample sequence is constructed; at the same time, the reference genome sequence is encoded into the sequence matching channel (255 , 0, 0), the sequence repetition channel (0, 0, 255) and the sequence inversion channel (0, 255, 0), the REF-to-REF image is obtained, and the reference genome sequence is removed by subtracting the two images. The impact of repeated fragments on subsequent structural variation detection is obtained, and a sequence similarity image of structural variation characteristics is obtained.

Step 3: Use a multi-target recognition framework to segment the structural variation feature sequence similarity image in step 2 to obtain a similarity image containing a single structural variation with SOI as multiple targets:

First of all, as the sequencing read length continues to increase, a read sequence is likely to span more than one structural variation at the same time. Therefore, the present invention first performs single structural variation segmentation on the structural variation characteristic sequence similarity image. Its main purpose is to segment the image Separate multiple different structural variations that appear in the structure to avoid predicting and characterizing continuous simple structural variations as complex structural variations; the specific steps are mainly to arrange and combine the main fragments in order to form a similarity image, that is, in a structural variation In the feature sequence similarity image, the sub-image composed of each two main fragments is considered to contain a structural variation. Combining two adjacent main fragments results in a similarity image containing only a single structural variation. Secondly, the SOI is obtained by combining the sequentially adjacent primary fragments and secondary fragments in the similarity image.

Step 4. Identification of structural variation feature sequences and representation of structural variation: Based on the multiple targets in the similarity image obtained in step 3, use the pre-trained CNN model to predict and classify, and finally use graphs to represent different types of structural variation. :

The present invention first constructs a structural variation training data set. The structural variation used for training mainly includes four simple structural variation types, deletion (DEL), insertion (INS), inversion (INV) and duplication (DUP). The real data uses the structural variation of 2,500 samples in the 1000Genome Project as training data, and the virtual data uses VISOR virtual noise-free training samples. The two together form the training data set. Encode structural variation in the training data set as described in step 2. The present invention selects the existing AlexNet neural network and uses the coded structural variation to train the AlexNet neural network. The AlexNet neural network training parameters use the transfer learning method, using the best score parameters in the Google ImageNet competition as model initialization parameters, and using cross The verification method selects the optimal model as the final structural variation detection CNN model. The target in the sub-image is identified through the structural variation detection CNN model, and finally structural variation of different complexity levels is identified without relying on any structural variation model. In addition, in view of the multi-breakpoint characteristics of complex structural variations, the present invention uses the data structure of graphs to characterize different types of complex structural variations. The mapping process first extracts all the fragments contained in the SOI in the structural variation feature sequence similarity image as nodes of the graph, and then each edge in the graph connects two adjacent fragments on the sample sequence. At the same time, based on the graph representation method, the present invention classifies different complex structural variations by judging the topological similarity of the graph. The specific steps include: 1) checking whether the two structural variation representation graphs have the same number of nodes and edges; 2) For representation graphs with the same number of edges and nodes, determine whether there are symmetric paths in the two graphs. If the above two conditions are met, the two complex structural variations being compared are considered to belong to the same type.

A deep learning-based genome structural variation detection system, including:

The structural variation characteristic sequence extraction module is used to compare the sample sequence with the reference genome sequence to obtain the global comparison result, and extract the structural variation characteristic sequence based on the global comparison result; the matching fragment in the structural variation characteristic sequence is called the main fragment; According to the alignment characteristics of the structural variation characteristic sequence, the sequence of the unmatched fragment in the structural variation characteristic sequence is subjected to local Kmer re-alignment with the reference genome sequence. The matching fragment obtained after the local Kmer re-alignment is called a secondary fragment;

The structural variation characteristic sequence encoding module is used to encode the structural variation characteristic sequence and the reference genome sequence using the RGB image three-channel encoding method to obtain an RGB image of similarity between the reference genome sequence and the sample sequence; at the same time, the reference genome sequence is encoded to obtain the reference Genome sequence self-similarity image; two images are subtracted to obtain a sequence similarity image of structural variation characteristics;

The structural variation characteristic sequence similarity image segmentation module is used to combine the two adjacent main fragments in the structural variation characteristic sequence similarity image according to the order of the main fragments on the reference genome sequence to obtain a sub-image containing only one structural variation; According to the order of the major fragments and minor fragments on the reference genome sequence, the adjacent major fragments and minor fragments are combined in the sub-image to obtain the fragment of interest;

The structural variation identification and characterization module uses a pre-trained structural variation detection CNN model to identify all interesting fragments in sub-images containing a single structural variation to obtain complex structural variation fragments; the graph data structure is used to perform complex structural variation fragments Systematic characterization and classification.

Structural variation feature sequence extraction module, specifically including:

The extraction module is used to compare the sample sequencing data with the reference genome sequence to obtain the comparison results; and extract the structural variation characteristic sequence based on the comparison results;

The low-quality variation signal removal module is used to remove structural variation signature sequences whose quality is not satisfactory, mainly including structural variation signature sequences with an alignment quality lower than 20 and those that generate more than 3 supplementary alignments;

Feature clustering and filtering module is used to cluster structural variation feature sequences based on feature similarity to obtain feature clusters, filter feature clusters whose feature values are less than a predetermined threshold, and retain structural variation feature sequences in the remaining feature clusters. and its CIGAR characters. That is, based on the similarity measure, signals that potentially support the same structural variation are clustered into one class through hierarchical clustering. After clustering, clusters that do not meet the minimum cluster size are removed, and the filtered clustering results are used to target potential variation areas. ;

The local alignment module based on the Kmer sequence is used to compare the unique mismatched fragments in the structural variation characteristic sequence according to the CIGAR characters of the structural variation characteristic sequence, and record the Kmer alignment results.

Structural variation characteristic sequence coding modules include:

The RGB three-channel sequence similarity coding module encodes the structural variation characteristic sequence and the reference genome sequence into the sequence matching channel (255, 0, 0), the sequence repeat channel (0, 0, 255) and the sequence inversion channel (0, 255 , 0), each channel can encode different sequences by setting the minimum fragment threshold. Through the above three-channel encoding method, the module will output an RGB image of the similarity between the reference genome sequence and the sample sequence; encode the reference genome sequence into the sequence In the matching channel (255, 0, 0), sequence repeat channel (0, 0, 255) and sequence inversion channel (0, 255, 0), the similarity image of the reference genome sequence itself is obtained, and the similarity of each matching fragment is recorded at the same time. Position information; the encoding process mainly includes encoding the main fragments through CIGAR character sequences of structural variation characteristics. Secondly, for the unmatched sequences existing in the alignment, a Kmer-based re-alignment encoding method is adopted. In this process, different sequence matching features will be Encoded into match channel, repeat channel and invert channel respectively.

The reference genome sequence repetitive segment module is removed, and based on the RGB image of the similarity between the reference genome sequence and the sample sequence and the coordinate position of the matching segment in the reference genome sequence's own similarity image relative to the reference genome sequence, find out each reference genome sequence and The fragment of the similarity RGB image of the sample sequence and the corresponding fragment in the similarity image of the reference genome sequence. After finding the corresponding fragment, remove it from the similarity RGB image of the reference genome sequence and the sample sequence to obtain the structural variation characteristic sequence. Similarity images. Used to solve the low detection rate and low accuracy of structural variations caused by repeated sequences in the reference genome. This module proposes for the first time to use the RGB coding matrix of the reference genome and sample sequence to subtract the RGB coding of the reference genome's own sequence to remove the impact of repeated sequences on structural variation detection and restore the true information of the sample sequence in this region. This method can effectively remove repeats. The impact of fragments on structural variation detection, especially complex structural variation, reduces the false positive rate and improves detection accuracy.

The fragment classification module is used to mark the main fragments and secondary fragments in the similarity image of the structural variation characteristic sequence after removing repeated fragments; the main fragment comes from the fragment marked M in the CIGAR character of the structural variation characteristic sequence, and the secondary fragment comes from Kmer-based local alignment results.

The structural variation feature sequence similarity image segmentation module includes:

A single structural variation segmentation module is used to combine the two adjacent main fragments in the structural variation feature sequence similarity image according to the order of the main fragments on the reference genome sequence to obtain a sub-image containing only one structural variation; used to combine the structure Separate multiple different structural variations that appear in the mutation signature sequence similarity image to avoid predicting and characterizing consecutive simple structural variations as complex structural variations;

The structural variation image segmentation module is used to sort the main fragments and the secondary fragments according to the coordinates on the structural variation feature sequence. According to the sorting results, the two adjacent fragments are combined to obtain the combined fragment. The filter consists of two secondary fragments and two A combined segment formed by combining linear segments, and a combined segment composed of a primary segment and a secondary segment is used as the segment of interest. It is used to identify structural variations of different levels of complexity, especially to identify the internal structures contained in complex structural variations. Compared with traditional computer vision multi-target recognition, the RGB images designed in the present invention are sparser and are not suitable for using the traditional sliding window method.

Structural variation identification and characterization module, including:

Construct a structural variation training data set module. The structural variation used for training mainly includes four simple structural variation types, deletion (DEL), insertion (INS), inversion (INV) and duplication (DUP). The real data uses the structural variation of 2500 samples in the 1000GenomeProject as training data. However, since DEL and INS account for the vast majority of the real data, it results in an unbalanced training data set. Therefore, the present invention adds noise-free interference virtual structural variation using the open source software VISOR to the training data set;

The training data set encoding module encodes all the training data in the training data set according to the structural variation feature sequence encoding module to obtain the interesting fragments of the training data set;

The convolutional neural network training module is used to train AlexNet neural network. The neural network training parameters use transfer learning method, using the best result parameters in the Google ImageNet competition as model initialization parameters. During the training process, the parameters are adjusted through backpropagation and gradient descent algorithms. to minimize the cross-entropy objective function. At the same time, cross-validation method is used to select the optimal model as the final structural variation detection CNN model;

The structural variation graph representation module is used to identify the fragments of interest in the sub-image through the structural variation detection CNN model, obtain the complex structural variation fragments, construct the structural variation representation graph based on the complex structural variation fragments, and calculate the topology structure based on the structural variation representation graph. Whether different structural variations belong to the same type, each node in the structural variation representation graph is a fragment included in all the fragments of interest in the sub-image, and each edge connects two consecutive fragments on the sample sequence; the nodes in the structural variation representation graph Represents the matching fragments contained in all SOIs in the structural variation characteristic sequence similarity image. These fragments are arranged from small to large according to the coordinates that appear in the sample sequence. The edges in the structural variation representation graph are used to connect two adjacent matching fragments; the structural variation representation graph is saved to a file according to the Graphical Fragment Assembly Format (GFA). In the output file, S represents the structural variation feature sequence. The matching segments obtained in the similarity image, L represents the connection relationship between different segments.

Simulation example

The simulation experiment evaluated and compared the present invention with four existing detection technologies from four aspects. The existing detection technologies included in the comparison include CuteSV, PbSV, Sniffles and SVIM. Both the present invention and the prior art use default parameters, in which the structural variation support threshold is 5, and only reads with alignment quality greater than or equal to 20 are selected.

Experiment 1: Since the present invention does not rely on any model to detect structural variation, it only detects structural variation of various levels of complexity by identifying sequence differences in image encodings. The first purpose of the experiment is to test the simple structural variation detection capabilities of the present invention for different sequencing platforms. For the two existing mainstream long-length sequencing platforms in the world, PacBio and Oxford Nanopore, this experiment is based on the detection software evaluation process in the Genome In A Bottle (GIAB) international genome authoritative project, using sample HG002 as the standard set to conduct a comprehensive evaluation. The present invention is aimed at the detection capabilities of sequencing data from different platforms and different amounts of sequencing data. The results are shown in Figure 3. The present invention can still achieve the same results as the most advanced detection software without relying on any model. The test results further reflect the universality of the present invention and can be applied to the detection of simple and complex structural variations.

Experiment 2: Since the present invention involves cutting-edge issues in this field, there is no industry-recognized set of complex structural variation standards for evaluating different methods. Therefore, based on the results of manual screening and inspection of complex structural variations published in Nature in 2015, the experiment virtualized 10 different complex structural variation types, each type containing 300 complex variation events. A virtual data set of complex structural variations was used to compare the performance of the present invention with existing detection technologies. In addition, unlike simple structural variations, complex structural variations usually contain multiple different types of breakpoints, called sub-elements. In the evaluation, two indicators are used to evaluate the performance of different detection methods: 1) regional matching; 2) complete match. Simply put, region matching requires that different detection methods can accurately detect genomic regions containing complex structural variations. Complete matching also requires accurate detection of complex structural variation regions and sub-elements of complex structural variation. The detection performance results are shown in Figure 4. For easier region matching, the recall rate of the present invention is 93%, while the detection method CuteSV with the second highest recall rate is only 70%. On the other hand, for complete matching, Sniffles is the only tool that can detect complex structural mutations. The recall rate of the present invention is 92%, which is twice that of Sniffles. Under the premise of high recall rate, the present invention can still maintain 90% The accuracy ensures the accuracy of detecting complex structural variations. However, other detection methods cannot detect complex structural variations and can only find the genomic interval where complex structural variations are located.

Experiment 3: This experiment aims to test the robustness of the sequence image encoding and recognition framework of the present invention by using sequence comparison results of different read lengths. This invention uses the structural variation of the HG00733 sample published in "Science" in 2021 as the standard set, and first selects the high-fidelity (HiFi) sequencing reads of the PacBio platform of the sample as the short DNA sequence input data set. On the other hand, in order to obtain a longer DNA sequence input data set, the experiment used contig sequences assembled from PacBio HiFi reads. These two reads of different DNA sequence lengths were aligned using existing sequence alignment technology. The input file is obtained on the human reference genome. The present invention first compared the detection performance of short DNA sequences. The results are shown in Figure 5. It can be seen that the detection performance of the present invention in detecting structural variations of different lengths is better than the model-based detection method. Since other detection solutions cannot detect long DNA sequence data sets, the present invention only compares the performance of the method of the present invention when short DNA sequences and long DNA sequences are used as input. As can be seen from the results in Figure 6, the detection method of the present invention The performance is positively related to the length of the input DNA sequence used for detection, which also meets the need for structural variation detection for longer and more accurate sequencing data in the future.

Experiment 4: With the development of genome assembly and typing technology in the past two years, more and more genome assembly results after typing have appeared. This experiment aims to compare whether this invention can meet the structural variation detection needs of future typing genomes. . The present invention still uses the different detection technologies for the classification and assembly results of the three samples HG00733, HG00514 and NA19240 published in "Science" in 2021 and the evaluation of the structural variation standard set after classification. The analysis and assembly results of these three samples are respectively passed through the current Alignment technology is available to the human reference genome. Since humans are diploid organisms, the standard set is divided into H1 and H2 for comparison. The results are shown in Figure 7. The Fscore of the present invention on both karyotypes is higher than that of the prior art. This experiment shows that the present invention can well meet the requirements for genome structure variation detection after typing and assembly.

Claims (8)

1. The genome structure variation detection method based on deep learning is characterized by comprising the following steps of:

step 1, extracting structural variation characteristic sequences: comparing the sample sequence with a reference genome sequence to obtain a global comparison result, and extracting a structural variation characteristic sequence according to the global comparison result; the matching fragments in the structural variation feature sequence are called main fragments; according to the comparison characteristics of the structural variation characteristic sequences, carrying out local Kmer weight comparison on the sequences of the unmatched fragments in the structural variation characteristic sequences and the reference genome sequences, and obtaining matched fragments called secondary fragments through the local Kmer weight comparison;

Step 2, structural variation characteristic sequence similarity image coding: adopting an RGB image three-channel coding mode, combining a main segment and a secondary segment, coding a structural variation characteristic sequence and a reference genome sequence to obtain a similarity RGB image of the reference genome sequence and a sample sequence, and simultaneously coding the reference genome sequence to obtain a similarity image of the reference genome sequence; subtracting the two images to obtain a structural variation characteristic sequence similarity image;

step 3, structural variation characteristic sequence similarity image segmentation: combining two adjacent main fragments in the structural variation characteristic sequence similarity image according to the sequence of the main fragments on the reference genome sequence to obtain a sub-image only containing single structural variation; combining adjacent primary fragments and secondary fragments in the sub-image in sequence in pairs according to the sequence of the primary fragments and the secondary fragments on the reference genome sequence to obtain a fragment of interest;

step 4, identifying structural variation characteristic sequence similarity images and carrying out structural variation characterization: identifying all interesting fragments in the sub-images containing single structural variation by using a pre-trained structural variation detection CNN model to obtain complex structural variation fragments; systematically characterizing and classifying the complex structure variant fragments by using a graph data structure;

The step 2 specifically comprises the following steps:

1) RGB three-way sequence similarity coding: coding the structural variation characteristic sequence and the reference genome sequence into a sequence matching channel (255, 0), a sequence repeating channel (0, 255) and a sequence reversing channel (0, 255, 0), and outputting a similarity RGB image of the reference genome sequence and the sample sequence; encoding the reference genome sequence into a sequence matching channel (255, 0), a sequence repeating channel (0, 255) and a sequence reversing channel (0, 255, 0) to obtain a similarity image of the reference genome sequence, and recording the position information of each matching segment;

2) Removal of reference genomic sequence repeat: and according to the RGB images of the similarity between the reference genome sequence and the sample sequence and the coordinate positions of the matching fragments in the similarity images of the reference genome sequence and the reference genome sequence relative to the reference genome sequence, fragments corresponding to the fragments in the similarity RGB images of the reference genome sequence and the sample sequence in the search of the similarity images of the reference genome sequence and the reference genome sequence, and if the corresponding fragments are found, removing the corresponding fragments from the RGB images of the similarity between the reference genome sequence and the sample sequence to obtain the structural variation characteristic sequence similarity images.

2. The deep learning-based genome structure variation detection method according to claim 1, wherein in step 1, according to the comparison feature of the structure variation feature sequence, the sequence of the unmatched fragment in the structure variation feature sequence is locally Kmer aligned with the reference genome sequence, specifically: and extracting unmatched fragments from the CIGAR characters of the structural variation characteristic sequence and the reference genome sequence, and carrying out local Kmer re-comparison on the sequences of the unmatched fragments and the reference genome sequence to obtain secondary fragments.

3. The method for detecting genomic structural variation based on deep learning according to claim 1, wherein step 3 specifically comprises:

1) Single structural variation segmentation: combining two adjacent main fragments in the structural variation characteristic sequence similarity image according to the sequence of the main fragments on the reference genome sequence to obtain a sub-image only containing single structural variation;

2) Structural variant image multi-objective segmentation: and sorting according to the coordinates of the main fragment and the secondary fragment on the structural variation characteristic sequence, and combining all fragments in the sub-images pairwise according to the sorting result, wherein the combination of the main fragment and the secondary fragment is used as the interested fragment.

4. The deep learning-based genomic structural variation detection method according to claim 1, wherein the structural variation detection CNN model training method in step 4 specifically comprises:

1) Constructing a structural variation training data set: the real data uses 2500 sample structural variation characteristic sequences in 1000Genome Project as training data, the virtual data uses VISOR virtual noise-free interference training samples as training data, and the training data set is formed by the real data and the virtual data;

2) Training data set encoding: encoding training data in the training data set by adopting the method in the step 2 to obtain an interested fragment of the training data set;

3) Model training: and inputting the training data set into a convolutional neural network, training the convolutional neural network, and obtaining a structural variation detection CNN model after training is completed.

5. The deep learning-based genome structure variation detection method according to claim 1, wherein in step 4, all interesting fragments in the sub-image containing the single structure variation are identified by using a pre-trained structure variation detection CNN model to obtain complex structure variation fragments; the complex structure variant fragments are systematically characterized and classified by using a graph data structure, and specifically:

Identifying interesting fragments in the sub-images through the structural variation detection CNN model to obtain complex structural variation fragments, constructing a structural variation characterization graph based on the complex structural variation fragments, and calculating whether different structural variations belong to the same type or not based on the topological structure of the structural variation characterization graph, wherein each node in the structural variation characterization graph is a fragment contained in all interesting fragments in the sub-images, and each side is connected with two continuous fragments on the sample sequence.

6. A deep learning-based genomic structural variation detection system comprising:

the structure variation characteristic sequence extraction module is used for comparing the sample sequence with the reference genome sequence to obtain a global comparison result, and extracting the structure variation characteristic sequence according to the global comparison result; the matching fragments in the structural variation feature sequence are called main fragments; according to the comparison characteristics of the structural variation characteristic sequences, carrying out local Kmer weight comparison on the sequences of the unmatched fragments in the structural variation characteristic sequences and the reference genome sequences, and obtaining matched fragments called secondary fragments through the local Kmer weight comparison;

the structure variation characteristic sequence coding module is used for coding the structure variation characteristic sequence and the reference genome sequence by adopting a three-channel coding mode of the RGB image to obtain a similarity RGB image of the reference genome sequence and the sample sequence; meanwhile, coding the reference genome sequence to obtain a self-similarity image of the reference genome sequence; subtracting the two images to obtain a structural variation characteristic sequence similarity image;

The structure variation characteristic sequence similarity image segmentation module is used for combining two adjacent main fragments in the structure variation characteristic sequence similarity image according to the sequence of the main fragments on the reference genome sequence to obtain a sub-image only comprising one structure variation; combining adjacent primary and secondary fragments in the sub-image in the order of the primary and secondary fragments on the reference genome sequence to obtain a fragment of interest;

the structural variation recognition and characterization module is used for recognizing all interesting fragments in the sub-images containing the single structural variation by using a pre-trained structural variation detection CNN model to obtain complex structural variation fragments; systematically characterizing and classifying the complex structure variant fragments by using a graph data structure;

the structural variation characteristic sequence coding module comprises:

the RGB three-channel sequence similarity coding module codes the structural variation characteristic sequence and the reference genome sequence into a sequence matching channel (255, 0), a sequence repeating channel (0, 255) and a sequence reversing channel (0, 255, 0) and outputs a similarity RGB image of the reference genome sequence and the sample sequence; coding the reference genome sequence into a sequence matching channel (255, 0), a sequence repeating channel (0, 255) and a sequence reversing channel (0, 255, 0) to obtain a similarity image of the reference genome sequence, and recording the position information of each matching fragment;

And removing the repeated segment module of the reference genome sequence, and removing the segment corresponding to the segment in the similarity RGB image of the reference genome sequence and the sample sequence in the search of the reference genome sequence self-similarity image according to the RGB image of the similarity of the reference genome sequence and the sample sequence and the coordinate position of the matched segment in the similarity image of the reference genome sequence self-similarity image relative to the reference genome sequence, if the corresponding segment is found, removing the segment from the RGB image of the similarity of the reference genome sequence and the sample sequence, thereby obtaining the structural variation characteristic sequence similarity image.

7. The deep learning based genomic structural variation detection system according to claim 6, wherein the structural variation feature sequence similarity image segmentation module comprises:

the single structure variation segmentation module is used for combining two adjacent main fragments in the structure variation characteristic sequence similarity image according to the sequence of the main fragments on the reference genome sequence to obtain a sub-image only comprising one structure variation;

the structure variation image segmentation module is used for sequencing according to the coordinates of the main segment and the secondary segment on the structure variation characteristic sequence, combining the two adjacent segments according to the sequencing result to obtain a combined segment, filtering the combined segment formed by combining the two secondary segments and the two linear segments, and taking the combined segment formed by combining the main segment and the secondary segment as the interested segment.

8. The deep learning based genomic structural variation detection system of claim 6, wherein the structural variation identification and characterization module specifically comprises:

the structure variation training data set module is used for forming a training data set from real data and virtual data, wherein the real data uses 2500 samples of structure variation in 1000Genome Project as training data, and the virtual data uses VISOR virtual noise-free training data;

the training data set coding module is used for coding all training data in the training data set according to the structural variation characteristic sequence coding module to obtain an interested fragment of the training data set;

the convolutional neural network training module is used for inputting the training data set into a convolutional neural network, training the convolutional neural network and obtaining a structural variation detection CNN model after training is completed;

the structure variation characterization module is used for identifying the interested fragments in the sub-images through the structure variation detection CNN model to obtain complex structure variation fragments, constructing a structure variation characterization graph based on the complex structure variation fragments, calculating whether different structure variations belong to the same type based on the topological structure of the structure variation characterization graph, wherein each node in the structure variation characterization graph is a fragment contained in all the interested fragments in the sub-images, and each side is connected with two continuous fragments on the sample sequence.