A method for detecting genomic structural variations based on long-read sequencing data and false positive filtering model

Abstract

A method for detecting genomic structural variations based on long-read sequencing data and a false-positive filtering model belongs to the field of bioinformatics. The present invention solves the problem of low accuracy of SV detection in long-read sequencing data by existing methods. The present invention uses individual long-read sequencing data as input, preliminarily obtains SV signals by comparing with the reference genome, and then uses the random forest network and the convolutional neural network as false-positive filtering models, and redesigns the clustering method of the SV signal to filter out false-positive SV signals from the preliminarily obtained SV signals, which can effectively reduce the possibility of calling false-positive events and has the ability to quickly and accurately detect SV signals from a variety of long-read sequencing platforms and various types. The method of the present invention can be applied to the detection of genomic structural variations in long-read sequencing data.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, and specifically relates to a method for detecting genomic structural variations based on long-read sequencing data and a false-positive filtering model. Background Art

[0002] Structural variation (SV) is a fundamental form of genomic variation, referring to DNA rearrangements larger than 50 base pairs in the genome. It includes insertions, deletions, inversions, duplications, and translocations. Growing evidence indicates that SVs are closely associated with numerous human diseases (e.g., cognitive neurological disorders, obesity, and cancer) and significantly impact many other phenotypic traits of organisms (e.g., growth and development, and adaptive traits). Therefore, comprehensive characterization of the diverse forms of SVs is crucial to fully understanding their contributions to genetic diversity, species differences, and some phenotypes.

[0003] A deeper understanding of the characteristics, location, and function of SVs can better understand their role in human health and disease, and provide an important basis for disease diagnosis, treatment, and prevention. With the rapid development of long-read sequencing technologies such as the Pacific Biosciences (PacBio) and Oxford Nanopore Technologies (ONT) platforms, long-read sequencing data provides opportunities for higher-resolution and more comprehensive SV detection. Currently, mainstream methods usually manually design heuristic rules from alignment data based on the characteristics of the sequencing platform and the types of different SV events to infer SVs. However, due to the complexity of SV events, manually designed heuristic rules can easily mistake data noise generated by sequencing or alignment for real SV signals. Therefore, accurately detecting SVs in noisy long-read sequencing data remains an important challenge.

[0004] In summary, the accuracy of existing methods for SV detection in long-read sequencing data is still low, and it is very necessary to propose a new detection method to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem of low accuracy of SV detection in long-read sequencing data by existing methods, and to propose a genomic structural variation detection method based on long-read sequencing data and a false positive filtering model.

[0006] The technical solution adopted by the present invention to solve the above technical problems is:

[0007] A method for detecting genomic structural variations based on long-read sequencing data and a false positive filtering model, the method specifically comprising the following steps:

[0008] Step 1: Align the long-read sequencing data to be tested with the reference genome to obtain CIGAR strings and gap information, and then extract SV signals based on the CIGAR strings and gap information;

[0009] And record the features of the SV signal extracted based on the CIGAR string and the features of the SV signal extracted based on the gap information respectively;

[0010] Step 2: Use the features of the SV signals extracted based on the gap information to construct an original feature matrix, perform minimum and maximum normalization on the constructed original feature matrix to obtain a processed feature matrix; input the processed feature matrix into the random forest network, and finally determine whether each SV signal extracted based on the gap information is a true SV signal and the type of the true SV signal based on the output of the random forest network;

[0011] Step 3: Cluster the SV signals extracted from the CIGAR string in Step 1 and the true SV signals determined in Step 2 to obtain clusters containing SV signals;

[0012] Step 4: For any cluster obtained in step 3, if the cluster contains the real SV signal determined in step 2, the SV signal in the cluster is not processed. Instead, the median of the starting point coordinates of all SV signals in the cluster and the median of the ending point coordinates of all SV signals in the cluster are directly calculated, and the calculated median is used as the final breakpoint coordinate of the candidate SV signal;

[0013] Otherwise, the feature matrix is constructed using the features of the SV signal in the cluster, and the constructed feature matrix is input into the trained convolutional neural network. The output of the convolutional neural network is used to finally determine whether the cluster is a true SV cluster;

[0014] If the cluster is a true SV cluster, the median of the starting point coordinates and the median of the ending point coordinates of all SV signals in the cluster are calculated, and the calculated median is used as the final breakpoint coordinate of the candidate SV signal;

[0015] If the cluster is not a true SV cluster, no processing is required;

[0016] Similarly, after processing each cluster obtained in step 3 separately, the final breakpoint coordinates of all candidate SV signals are obtained, completing the detection of genomic structural variation.

[0017] Furthermore, the CIGAR character string is obtained as follows:

[0018] When directly aligning the long-read sequencing data to be detected with the reference genome, the alignment record obtained is the CIGAR string;

[0019] The specific process of extracting SV signals according to the CIGAR string is as follows:

[0020] The CIGAR string is a string composed of numbers and letters. The SV signals of the deletion and insertion types are located according to the letters in the CIGAR string, and the lengths of the SV signals of the deletion and insertion types are determined according to the numbers in the CIGAR string.

[0021] Furthermore, the way to obtain the gap information is as follows:

[0022] After splitting the long-read sequencing data to be detected into multiple segments, the bases at both ends of each segment are the breakpoints of the segment, and the coordinates of the breakpoints are the gap information;

[0023] The extraction of SV signals according to the gap information includes identifying SV signals and extracting SV signals; where:

[0024] The specific process of identifying SV signals is as follows:

[0025] (1) For any two adjacent segments, one of the segments is taken as the main alignment segment, and the other segment is taken as the supplementary alignment segment. The distance between the main alignment segment and the supplementary alignment segment in the long-read sequencing data to be detected is denoted as dis_read, and the distance between the main alignment segment and the supplementary alignment segment in the reference genome is denoted as dis_ref;

[0026] Compare the magnitudes of dis_read and dis_ref:

[0027] If dis_read < dis_ref, there is a deletion SV signal between the main alignment segment and the supplementary alignment segment;

[0028] If dis_read > dis_ref, there is an insertion SV signal between the main alignment segment and the supplementary alignment segment;

[0029] If dis_read = dis_ref, there is no SV signal between the main alignment segment and the supplementary alignment segment;

[0030] (2) If there is an overlap between the alignment position of the main alignment segment on the reference genome and the alignment position of the supplementary alignment segment on the reference genome, there is a duplication SV signal between the main alignment segment and the supplementary alignment segment;

[0031] (3) If the orientation of the main alignment fragment relative to the reference genome is different from that of the supplementary alignment fragment relative to the reference genome, there is an inversion SV signal between the main alignment fragment and the supplementary alignment fragment;

[0032] (4) If the chromosome number corresponding to the main alignment fragment in the reference genome is different from the chromosome number corresponding to the supplementary alignment fragment in the reference genome, there is a translocation SV signal between the main alignment fragment and the supplementary alignment fragment;

[0033] The specific process of extracting SV signals is: extracting the identified SV signals.

[0034] Furthermore, the distance dis_read between the main alignment fragment and the supplementary alignment fragment in the long-read sequencing data to be detected is:

[0035] If the main alignment fragment is located before the supplementary alignment fragment in the long-read sequencing data to be detected, the distance dis_read is the difference between the starting base coordinate of the supplementary alignment fragment and the ending base coordinate of the main alignment fragment;

[0036] If the main alignment fragment is located after the supplementary alignment fragment in the long-read sequencing data to be detected, the distance dis_read is the difference between the starting base coordinates of the main alignment fragment and the ending base coordinates of the supplementary alignment fragment.

[0037] Furthermore, the features of the SV signal extracted based on the CIGAR character string include global alignment features, local CIGAR features, and base distribution features; wherein:

[0038] Global alignment features include the alignment quality of the SV signal, the soft-clip ratio, the soft-clip length, the FLAG value, the length of the long-read sequencing data to be detected, the number of variants, the partition number of the alignment, the starting position of the SV signal in the long-read sequencing data to be detected, the ending position of the SV signal in the long-read sequencing data to be detected, the chromosome number of the long-read sequencing data to be detected in the reference genome, the starting position of the long-read sequencing data to be detected in the reference genome, the ending position of the long-read sequencing data to be detected in the reference genome, the edit distance of the alignment, and the length of the mismatched bases;

[0039] The local CIGAR features include the starting point of the SV signal in the reference genome, the end point of the SV signal in the reference genome, the length of the SV signal, the type of the SV signal, the starting position of the SV signal in CIGAR, and the ending position of the SV signal in CIGAR in the long-read sequencing data to be detected;

[0040] The base distribution characteristics include the frequencies of occurrence of the four bases in the SV signal and the frequency of occurrence of repeated bases.

[0041] Furthermore, the features of the SV signal extracted based on the gap information include global alignment features, split alignment features and base distribution features; wherein:

[0042] The global alignment features are the same as those of the SV signal extracted based on the CIGAR string;

[0043] The base distribution features are the same as those of the SV signal extracted from the CIGAR string;

[0044] The split alignment features include the consistency of the corresponding chromosome numbers of the main alignment and the supplementary alignment in the reference genome, the consistency of the directions of the main alignment and the supplementary alignment in the reference genome, the distance d1 between the main alignment and the supplementary alignment in the long-read sequencing data to be tested, the distance d2 between the main alignment and the supplementary alignment in the reference genome, and the difference between the distances d1 and d2.

[0045] Furthermore, the training process of the random forest network is:

[0046] Step 1: Align the known long-read sequencing data with the reference genome to obtain gap information, and then extract SV signals based on the gap information; and record the characteristics of the SV signals extracted based on the gap information;

[0047] Step 2: Divide all SV signals on chromosomes 1 to 11 extracted in step 1 into two parts at a ratio of 7:3. Use one part as the training set and the other as the validation set. Annotate the SV signals in the training set and validation set according to the benchmark set. Use all SV signals on chromosomes 12 to 22 and sex chromosomes extracted in step 1 as the test set.

[0048] The marking method of SV signal is:

[0049] The annotation method for SV signals of deletion, insertion, duplication, and inversion is the same. Take the insertion SV signal as an example:

[0050] The starting coordinate of any extracted insertion SV signal is marked as s1, and the length of the insertion SV signal is recorded as size1. It is determined whether there is an insertion SV signal in the benchmark set that is on the same chromosome as the insertion SV signal and satisfies the following conditions:

[0051]

[0052] Where s2 is the starting coordinate of the inserted SV signal in the benchmark set, |·| represents the absolute value, and size2 is the length of the inserted SV signal in the benchmark set;

[0053] If there is an inserted SV signal in the benchmark set that is on the same chromosome as the inserted SV signal and meets the conditions, then the extracted inserted SV signal is a true inserted SV signal and the extracted inserted SV signal is marked as an inserted SV signal; otherwise, if there is no inserted SV signal in the benchmark set that is on the same chromosome as the inserted SV signal and meets the conditions, then the extracted inserted SV signal is marked as a false detection;

[0054] The annotation method for the translocated SV signal is:

[0055] The starting coordinate of any extracted translocation SV signal is marked as s′1, and it is determined whether there is a translocation SV signal in the reference set that is on the same chromosome as the translocation SV signal and satisfies the following conditions:

[0056] |s′1-s′2|≤1000

[0057] Where s′2 is the starting coordinate of the translocated SV signal in the benchmark set;

[0058] If there is a translocation SV signal in the benchmark set that is on the same chromosome as the translocation SV signal and meets the conditions, the extracted translocation SV signal is a true translocation SV signal, and the extracted translocation SV signal is marked as a translocation SV signal; otherwise, there is no translocation SV signal in the benchmark set that is on the same chromosome as the translocation SV signal and meets the conditions, and the extracted translocation SV signal is marked as a false detection;

[0059] Step 3: construct a feature matrix using the SV signal features extracted according to the gap information in step 1, and then perform minimum and maximum normalization processing on the constructed feature matrix to obtain a processed feature matrix;

[0060] The random forest network is trained using the labeled SV signals and processed feature matrix in the training set.

[0061] Furthermore, the clustering process in step 3 is performed on each type of SV signal separately; taking the insertion of SV signal as an example:

[0062] Step 31: Represent each inserted SV signal extracted from the CIGAR string in step 1 and determined in step 2 as a four-tuple (t, c, s, e), where t represents the SV signal type, c represents the chromosome number where the SV signal is located, s represents the starting coordinate of the SV signal in the reference genome, and e represents the ending coordinate of the SV signal in the reference genome;

[0063] Sort each insertion SV signal: that is, for all insertion SV signals on a chromosome, sort each insertion SV signal in ascending order according to the starting point coordinates of the insertion SV signal in the reference genome. If the spatial distance between two adjacent insertion SV signals is less than 1000bp after sorting, the two adjacent insertion SV signals are divided into the same partition, otherwise the two adjacent insertion SV signals are divided into different partitions;

[0064] Similarly, the inserted SV signals on each chromosome were sorted and partitioned separately;

[0065] Step 3.2: Get the SV local depth of each partition and the SV global depth of all partitions:

[0066]

[0067] Where a is the SV global depth, b i is the SV local depth of the i-th partition, N is the total number of partitions;

[0068] Step 3: Cluster the inserted SV signals in the i-th partition, where i = 1, 2, ..., N. The specific clustering process is as follows:

[0069] Step 331: For the k-th inserted SV signal and the l-th inserted SV signal in the i-th partition, calculate the similarity S(k,l) between the k-th inserted SV signal and the l-th inserted SV signal:

[0070]

[0071] Among them, pos_dis is the site distance between the kth inserted SV signal and the lth inserted SV signal, span_dis is the span distance between the kth inserted SV signal and the lth inserted SV signal, s k and e k are the starting and ending coordinates of the kth insertion SV signal in the reference genome, s l and e l are the starting and ending coordinates of the lth inserted SV signal in the reference genome, respectively, and λ is the correction parameter;

[0072]

[0073] span_dis=|(e k -s k )-(e l -s l )|

[0074]

[0075] Step 332: Perform hierarchical clustering on the inserted SV signals in the i-th partition according to the calculated similarity to obtain each cluster;

[0076] Step 34: Perform the process of step 31 to step 33 respectively on the missing SV signal, repeated SV signal, inverted SV signal and transposed SV signal extracted according to the CIGAR character string in step 1 and determined in step 2.

[0077] Furthermore, after the sorting, the spatial distance between two adjacent inserted SV signals is s k+1 -s k , where s k is the starting coordinate of the kth inserted SV signal, s k+1 is the starting coordinate of the k+1th inserted SV signal;

[0078] The method for calculating the spatial distance between two adjacent translocated SV signals is the same as that for calculating the spatial distance between two adjacent inserted SV signals;

[0079] The spatial distance between two adjacent missing SV signals is s′ k+1 -e′ k , where e′ k is the end point coordinate of the kth missing SV signal, s′ k+1 is the starting coordinate of the k+1th missing SV signal;

[0080] The calculation method of the spatial distance between two adjacent inverted SV signals and the spatial distance between two adjacent repeated SV signals is the same as the calculation method of the spatial distance between two adjacent missing SV signals.

[0081] Furthermore, the training method of the convolutional neural network is:

[0082] Step S1: Align the known long-read sequencing data with the reference genome to obtain a CIGAR string, then extract the SV signal based on the CIGAR string; and record the characteristics of the SV signal extracted based on the CIGAR string;

[0083] Step S2: extracting SV signal features based on the gap information of the known long-read sequencing data in step S1, constructing a feature matrix using the features of the SV signals extracted based on the gap information, performing minimum and maximum normalization processing on the constructed feature matrix, inputting the processed feature matrix into a random forest network, and predicting the true SV signal from the extracted SV signals through the random forest network;

[0084] Step S3, clustering the SV signal extracted according to the CIGAR character string in step S1 and the true SV signal determined in step S2 together to obtain clusters containing SV signals;

[0085] Step S4: Determine the clusters of SV signals extracted from the comparison, and train the convolutional neural network based on the feature matrix and labels of each cluster; specifically:

[0086] Step S41: for any determined cluster, determine whether the number of SV signals in the cluster is within the range of [1, 20].

[0087] If it is not within the range of [1,20], execute step S42;

[0088] If it is within the range of [1,20], execute step S45;

[0089] Step S42: Continue to determine whether the number of SV signals in the cluster is less than or equal to 100;

[0090] If the number of SV signals in the cluster is less than or equal to 100, step S43 is executed;

[0091] If the number of SV signals in the cluster is greater than 100, step S44 is executed;

[0092] Step S43: directly use the SV signal features in the cluster to construct a feature matrix with a dimension of (100, 25). If the number of SV signals in the cluster is less than 100, it is necessary to obtain the feature matrix through zero filling operation and add labels to the feature matrix;

[0093] And perform minimum-maximum normalization processing on the constructed feature matrix to obtain a processed feature matrix;

[0094] Step S44: Randomly select 100 SV signals from the cluster to execute step S43, and determine whether the number of remaining SV signals in the cluster is in the range of [1, 20].

[0095] If it is not within the range of [1,20], execute step S42;

[0096] If it is within the range of [1,20], execute step S45;

[0097] Step S45: The number of SV signals in the cluster is recorded as w. Each SV signal in the cluster is truncated using a window size smaller than w. The data in the window obtained after each truncation is used as a new SV signal. A feature matrix is constructed using the features of the SV signal before truncation and the new SV signal after truncation, and a label is added to the feature matrix.

[0098] Then perform minimum-maximum normalization on the constructed feature matrix to obtain a processed feature matrix;

[0099] Step S5: process each determined cluster separately using the method of step S4;

[0100] Step S6: Use the obtained processed feature matrix and the added labels to train the convolutional neural network.

[0101] Step S6: Use the obtained processed feature matrix and the added labels to train the convolutional neural network.

[0102] The beneficial effects of the present invention are:

[0103] The present invention uses individual long-read sequencing data as input, preliminarily obtains SV signals by comparing with the reference genome, then uses the random forest network and convolutional neural network as false positive filtering models, and redesigns the clustering method of SV signals to filter out false-positive SV signals from the preliminarily obtained SV signals. This can effectively reduce the possibility of calling false-positive events and has the ability to quickly and accurately detect SV signals from multiple long-read sequencing platforms and various types. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 This is a flow chart of a method for detecting genomic structural variations based on long-read sequencing data and a false-positive filtering model of the present invention. DETAILED DESCRIPTION

[0105] Specific implementation method 1: Combination Figure 1 This embodiment describes a method for detecting genomic structural variations based on long-read sequencing data and a false positive filtering model, the method specifically comprising the following steps:

[0106] Step 1: Align the long-read sequencing data to be tested with the reference genome to obtain CIGAR strings and gap information, and then extract SV signals based on the CIGAR strings and gap information;

[0107] And record the features of the SV signal extracted based on the CIGAR string and the features of the SV signal extracted based on the gap information respectively;

[0108] Step 2: Use the features of the SV signals extracted based on the gap information to construct an original feature matrix, perform minimum and maximum normalization on the constructed original feature matrix to obtain a processed feature matrix; input the processed feature matrix into the random forest network, and finally determine whether each SV signal extracted based on the gap information is a true SV signal and the type of the true SV signal based on the output of the random forest network;

[0109] Step 3: Cluster the SV signals extracted from the CIGAR string in Step 1 and the true SV signals determined in Step 2 to obtain clusters containing SV signals;

[0110] Step 4: For any cluster obtained in step 3, if the cluster contains the real SV signal determined in step 2, the SV signal in the cluster is not processed. Instead, the median of the starting point coordinates of all SV signals in the cluster and the median of the ending point coordinates of all SV signals in the cluster are directly calculated, and the calculated median is used as the final breakpoint coordinate of the candidate SV signal;

[0111] Otherwise, the feature matrix is constructed using the features of the SV signal in the cluster, and the constructed feature matrix is input into the trained convolutional neural network. The output of the convolutional neural network is used to finally determine whether the cluster is a true SV cluster;

[0112] If the cluster is a true SV cluster, the median of the starting point coordinates and the median of the ending point coordinates of all SV signals in the cluster are calculated, and the calculated median is used as the final breakpoint coordinate of the candidate SV signal;

[0113] If the cluster is not a true SV cluster, no processing is required;

[0114] Similarly, after processing each cluster obtained in step 3 separately, the final breakpoint coordinates of all candidate SV signals are obtained, completing the detection of genomic structural variation.

[0115] SV signal types include deletion variation, insertion variation, duplication variation, inversion variation, and translocation variation. Deletion variation refers to the loss of a DNA sequence at a specific location in the genome, which reduces the number of nucleotides relative to the reference genome; insertion variation refers to the insertion of an additional DNA sequence at a specific location in the genome, which increases the number of nucleotides relative to the reference genome; duplication variation refers to a DNA sequence in the genome being repeated one or more times. During the alignment process, the long-read sequencing data contains repeated sequences; inversion variation refers to a specific DNA sequence in the genome that has flipped 180 degrees in direction. During the alignment process, a sequence in the long-read sequencing data is in the opposite direction to the reference genome; translocation variation refers to a DNA sequence in the genome being moved to a different location on the same chromosome (and the distance between the starting positions before and after the movement is greater than the set threshold) or moved to a different chromosome.

[0116] The present invention utilizes random forest network and convolutional neural network as false positive filtering models, which can filter out false positive SV signals and improve the detection accuracy.

[0117] Specific embodiment 2: This embodiment differs from specific embodiment 1 in that the CIGAR character string is obtained as follows:

[0118] When the long-read sequencing data to be tested is directly aligned with the reference genome, the resulting alignment record is the CIGAR string;

[0119] The specific process of extracting the SV signal based on the CIGAR character string is as follows:

[0120] The CIGAR string consists of numbers and letters. The missing and inserted SV signals are located according to the letters in the CIGAR string, and the lengths of the missing and inserted SV signals are determined according to the numbers in the CIGAR string.

[0121] Other steps and parameters are the same as those in the first embodiment.

[0122] When a single long-read sequencing data (read) is directly aligned with a reference genome, abnormal alignment phenomena observed are internal alignment signals. These mismatches are generated within a continuous alignment region and are recorded in the alignment record as a CIGAR string. Different letters in the CIGAR string represent different SV signal types, and the preceding number represents the base length corresponding to that SV signal type. For example, "20=12D50=60I" means that the first 20 bases of the fragment completely match the reference genome, the next 12 bases are the length of the bases missing in the fragment, i.e., a deletion fragment, followed by 50 matching bases, and the last 60 bases are bases that do not exist in the reference genome but appear in the fragment, i.e., an insertion fragment. When extracting abnormal alignment information, the main focus is on the D and I characters in the CIGAR string, which represent deletions and insertions, respectively. By recording their position and length, both types of SV signals, deletions and insertions, are extracted.

[0123] Specific embodiment three: This embodiment differs from specific embodiment one or two in that the gap information is obtained in the following manner:

[0124] After splitting the long-read sequencing data to be tested into multiple fragments (meaning more than 2 fragments, split using alignment tools), the bases at both ends of each fragment are the breakpoints of the fragment, and the coordinates of the breakpoints are the gap information;

[0125] Extracting the SV signal according to the gap information includes identifying the SV signal and extracting the SV signal; wherein:

[0126] The specific process of identifying SV signals is as follows:

[0127] (1) For any two adjacent fragments, one of the fragments is used as the main alignment fragment, and the other is used as the supplementary alignment fragment. The distance between the main alignment fragment and the supplementary alignment fragment in the long-read sequencing data to be detected is recorded as dis_read, and the distance between the main alignment fragment and the supplementary alignment fragment in the reference genome is recorded as dis_ref;

[0128] Compare the magnitudes of dis_read and dis_ref:

[0129] If dis_read < dis_ref, there is a missing SV signal between the main alignment fragment and the supplementary alignment fragment;

[0130] If dis_read > dis_ref, there is an inserted SV signal between the main alignment fragment and the supplementary alignment fragment;

[0131] If dis_read = dis_ref, there is no SV signal between the main alignment fragment and the supplementary alignment fragment;

[0132] (2) If the alignment positions of the main alignment fragment and the supplementary alignment fragment on the reference genome overlap, there is a duplicated SV signal between the main alignment fragment and the supplementary alignment fragment;

[0133] (3) If the directions of the main alignment fragment and the supplementary alignment fragment relative to the reference genome are different, there is an inverted SV signal between the main alignment fragment and the supplementary alignment fragment;

[0134] (4) If the chromosome numbers corresponding to the main alignment fragment and the supplementary alignment fragment in the reference genome are different, there is a translocated SV signal between the main alignment fragment and the supplementary alignment fragment;

[0135] The specific process of extracting the SV signal is as follows: Extract the identified SV signal.

[0136] Other steps and parameters are the same as those in the first or second specific implementation manner.

[0137] Specific implementation manner four: The difference between this implementation manner and one of the first to third specific implementation manners is that the distance dis_read between the main alignment fragment and the supplementary alignment fragment in the to-be-detected long-read sequencing data is as follows:

[0138] If in the to-be-detected long-read sequencing data, the main alignment fragment is before the supplementary alignment fragment, the distance dis_read is the difference between the starting base coordinate of the supplementary alignment fragment and the ending base coordinate of the main alignment fragment;

[0139] If in the to-be-detected long-read sequencing data, the main alignment fragment is after the supplementary alignment fragment, the distance dis_read is the difference between the starting base coordinate of the main alignment fragment and the ending base coordinate of the supplementary alignment fragment.

[0140] Other steps and parameters are the same as those in one of the first to third specific implementation manners.

[0141] For any two fragments, the method of determining the main alignment fragment and the supplementary alignment fragment includes but is not limited to: using the fragment containing more bases as the main alignment fragment and using the fragment containing fewer bases as the supplementary alignment fragment.

[0142] Specific embodiment 5: This embodiment differs from any one of specific embodiments 1 to 4 in that the features of the SV signal extracted based on the CIGAR string include global alignment features, local CIGAR features, and base distribution features; wherein:

[0143] Global alignment features include alignment quality of SV signals, soft shear ratio, soft shear length, FLAG value, length of long-read sequencing data to be detected, number of variants, aligned partition number, starting position of SV signal in long-read sequencing data to be detected, ending position of SV signal in long-read sequencing data to be detected, chromosome number of long-read sequencing data to be detected in the reference genome, starting position of long-read sequencing data to be detected in the reference genome, ending position of long-read sequencing data to be detected in the reference genome, edit distance of alignment, and length of mismatched bases; local CIGAR features include starting point of SV signal in long-read sequencing data to be detected in the reference genome, ending point of SV signal in the reference genome, length of SV signal, type of SV signal, starting position of SV signal in CIGAR, and ending position of SV signal in CIGAR;

[0144] The base distribution characteristics include the frequencies of occurrence of the four bases in the SV signal and the frequencies of occurrence of repeated bases (repeated bases are defined as appearing at least five times in a row).

[0145] The other steps and parameters are the same as those in the first to fourth embodiments.

[0146] Specific embodiment 6: This embodiment differs from any one of specific embodiments 1 to 5 in that the features of the SV signal extracted based on the gap information include global alignment features, split alignment features, and base distribution features; wherein:

[0147] The global alignment features are the same as those of the SV signal extracted based on the CIGAR string;

[0148] The base distribution features are the same as those of the SV signal extracted from the CIGAR string;

[0149] The split alignment features include the consistency of the corresponding chromosome numbers of the main alignment and the supplementary alignment in the reference genome, the consistency of the directions of the main alignment and the supplementary alignment in the reference genome, the distance d1 between the main alignment and the supplementary alignment in the long-read sequencing data to be tested, the distance d2 between the main alignment and the supplementary alignment in the reference genome, and the difference between the distances d1 and d2.

[0150] The other steps and parameters are the same as those in the first to fifth embodiments.

[0151] Specific embodiment 7: This embodiment differs from any one of specific embodiments 1 to 6 in that the training process of the random forest network is:

[0152] Step 1: Align the known long-read sequencing data with the reference genome to obtain gap information, and then extract SV signals based on the gap information; and record the characteristics of the SV signals extracted based on the gap information;

[0153] Step 2: Divide all SV signals on chromosomes 1 to 11 extracted in step 1 into two parts at a ratio of 7:3. Use one part as the training set and the other as the validation set. Label the SV signals in the training set and validation set according to the benchmark set (which contains currently known abnormal signals). Use all SV signals on chromosomes 12 to 22 and sex chromosomes extracted in step 1 as the test set.

[0154] The marking method of SV signal is:

[0155] The annotation method for SV signals of deletion, insertion, duplication, and inversion is the same. Take the insertion SV signal as an example:

[0156] The starting coordinate of any extracted insertion SV signal is marked as s1, and the length of the insertion SV signal is recorded as size1. It is determined whether there is an insertion SV signal in the benchmark set that is on the same chromosome as the insertion SV signal and satisfies the following conditions:

[0157]

[0158] Where s2 is the starting coordinate of the inserted SV signal in the benchmark set, |·| represents the absolute value, and size2 is the length of the inserted SV signal in the benchmark set;

[0159] If there is an inserted SV signal in the benchmark set that is on the same chromosome as the inserted SV signal and meets the conditions, then the extracted inserted SV signal is a true inserted SV signal and the extracted inserted SV signal is marked as an inserted SV signal; otherwise, if there is no inserted SV signal in the benchmark set that is on the same chromosome as the inserted SV signal and meets the conditions, then the extracted inserted SV signal is marked as a false detection;

[0160] The annotation method for the translocated SV signal is:

[0161] The starting coordinate of any extracted translocation SV signal is marked as s′1, and it is determined whether there is a translocation SV signal in the reference set that is on the same chromosome as the translocation SV signal and satisfies the following conditions:

[0162] |s′1-s′2|≤1000

[0163] Where s′2 is the starting coordinate of the translocated SV signal in the benchmark set;

[0164] If there is a translocation SV signal in the benchmark set that is on the same chromosome as the translocation SV signal and meets the conditions, the extracted translocation SV signal is a true translocation SV signal, and the extracted translocation SV signal is marked as a translocation SV signal; otherwise, there is no translocation SV signal in the benchmark set that is on the same chromosome as the translocation SV signal and meets the conditions, and the extracted translocation SV signal is marked as a false detection;

[0165] Step 3: construct a feature matrix using the SV signal features extracted according to the gap information in step 1, and then perform minimum and maximum normalization processing on the constructed feature matrix to obtain a processed feature matrix;

[0166] The random forest network is trained using the labeled SV signals and processed feature matrix in the training set.

[0167] The other steps and parameters are the same as those in the first to sixth embodiments.

[0168] The random forest network was constructed using the sklearn library. During the hyperparameter optimization process, the maximum depth of each tree (between 2 and 10), the total number of trees in the forest (between 10 and 100), and the minimum number of leaves required to split an internal node (between 10 and 100) were varied to select the optimal model with the highest F1 value.

[0169] Specific embodiment eight: This embodiment differs from any one of specific embodiments one to seven in that the clustering process in step three is performed on each type of SV signal separately; taking the insertion of SV signals as an example:

[0170] Step 31: Represent each inserted SV signal extracted from the CIGAR string in step 1 and determined in step 2 as a four-tuple (t, c, s, e), where t represents the SV signal type, c represents the chromosome number where the SV signal is located, s represents the starting coordinate of the SV signal in the reference genome, and e represents the ending coordinate of the SV signal in the reference genome;

[0171] Sort each insertion SV signal: that is, for all insertion SV signals on a chromosome, sort each insertion SV signal in ascending order according to the starting point coordinates of the insertion SV signal in the reference genome. If the spatial distance between two adjacent insertion SV signals is less than 1000bp after sorting, the two adjacent insertion SV signals are divided into the same partition, otherwise the two adjacent insertion SV signals are divided into different partitions;

[0172] Similarly, the inserted SV signals on each chromosome were sorted and partitioned separately;

[0173] For the inserted SV signal on any chromosome, for example, after sorting, the spatial distance between the first inserted SV signal and the second inserted SV signal is less than 1000 bp, the spatial distance between the second inserted SV signal and the third inserted SV signal is less than 1000 bp, the spatial distance between the third inserted SV signal and the fourth inserted SV signal is not less than 1000 bp, the spatial distance between the fourth inserted SV signal and the fifth inserted SV signal is less than 1000 bp, the spatial distance between the fifth inserted SV signal and the sixth inserted SV signal is not less than 1000 bp, and the spatial distance between the sixth inserted SV signal and the seventh inserted SV signal is less than 1000 bp; then the first inserted SV signal, the second inserted SV signal and the third inserted SV signal are divided into the same partition, the fourth inserted SV signal and the fifth inserted SV signal are divided into the same partition, and the sixth inserted SV signal and the seventh inserted SV signal are divided into the same partition;

[0174] Step 32: Obtain the SV local depth of each partition (i.e., the number of inserted SV signals contained in each partition) and the SV global depth of all partitions:

[0175]

[0176] Where a is the SV global depth, b i is the SV local depth of the i-th partition, N is the total number of partitions;

[0177] Step 3: Cluster the inserted SV signals in the i-th partition, where i = 1, 2, ..., N. The specific clustering process is as follows:

[0178] Step 331: For the k-th inserted SV signal and the l-th inserted SV signal in the i-th partition, calculate the similarity S(k,l) between the k-th inserted SV signal and the l-th inserted SV signal:

[0179]

[0180] Among them, pos_dis is the site distance between the kth inserted SV signal and the lth inserted SV signal, span_dis is the span distance between the kth inserted SV signal and the lth inserted SV signal, s k and e k are the starting and ending coordinates of the kth insertion SV signal in the reference genome, s l and e lare the starting and ending coordinates of the lth inserted SV signal in the reference genome, respectively, and λ is the correction parameter;

[0181]

[0182] span_dis=|(e k -s k )-(e l -s l )|

[0183]

[0184] The relative importance of span features in signal similarity measurement can be increased by using the correction parameter λ. The λ value adjusts the similarity between SV signals according to the SV depth and SV type of the local area. By increasing the clustering difficulty of low-depth partitions and reducing the clustering difficulty of high-depth partitions, SV signals in low-confidence areas can be better filtered and SV signals in high-confidence areas can be enhanced.

[0185] Step 332: hierarchical clustering is performed on the inserted SV signals in the i-th partition according to the calculated similarity, that is, the partition is refined by hierarchical clustering to further obtain each cluster;

[0186] The threshold of hierarchical clustering was set to 0.3, and clustering was stopped when the similarity between any two clusters was greater than the threshold, obtaining the final clustering result;

[0187] Step 34: Perform the process of step 31 to step 33 respectively on the missing SV signal, repeated SV signal, inverted SV signal and transposed SV signal extracted according to the CIGAR character string in step 1 and determined in step 2.

[0188] The other steps and parameters are the same as those in the first to seventh embodiments.

[0189] SV detection generally clusters the signals of the same SV event to enhance the SV signal. However, due to the ambiguity of the clustering threshold, true events and false-positive events are often confused in the same cluster space. To address this problem, the present invention proposes a two-step clustering algorithm based on spatial distance and hierarchical structure. The method of the present invention can significantly improve the accuracy of SV detection.

[0190] Specific embodiment 9: The difference between this embodiment and any one of the specific embodiments 1 to 8 is that after the sorting, the spatial distance between two adjacent inserted SV signals is s k+1 -s k , where s k is the starting coordinate of the kth inserted SV signal, s k+1 is the starting coordinate of the k+1th inserted SV signal;

[0191] The method for calculating the spatial distance between two adjacent translocated SV signals is the same as that for calculating the spatial distance between two adjacent inserted SV signals;

[0192] The spatial distance between two adjacent missing SV signals is s′ k+1 -e′ k , where e′ k is the end point coordinate of the kth missing SV signal, s′ k+1 is the starting coordinate of the k+1th missing SV signal;

[0193] The calculation method of the spatial distance between two adjacent inverted SV signals and the spatial distance between two adjacent repeated SV signals is the same as the calculation method of the spatial distance between two adjacent missing SV signals.

[0194] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0195] Specific embodiment 10: This embodiment differs from any one of specific embodiments 1 to 9 in that the training method of the convolutional neural network is:

[0196] Step S1: Align the known long-read sequencing data with the reference genome to obtain a CIGAR string, then extract the SV signal based on the CIGAR string; and record the characteristics of the SV signal extracted based on the CIGAR string;

[0197] Step S2: extracting SV signal features based on the gap information of the known long-read sequencing data in step S1, constructing a feature matrix using the features of the SV signals extracted based on the gap information, performing minimum and maximum normalization processing on the constructed feature matrix, inputting the processed feature matrix into a random forest network, and predicting the true SV signal from the extracted SV signals through the random forest network;

[0198] Step S3, clustering the SV signal extracted according to the CIGAR character string in step S1 and the true SV signal determined in step S2 together to obtain clusters containing SV signals;

[0199] Step S4: Determine the cluster of SV signals extracted from the comparison (i.e., determine the cluster that only contains the SV signals extracted in step S1), and train a convolutional neural network (CNN) based on the feature matrix and label of each cluster; specifically:

[0200] Convolutional layers of varying sizes are used to extract the differences in local features between real and false-positive samples. The CNN backbone consists of three convolutional layers (with kernel sizes of 3×25, 3×8, and 3×1, respectively), three max pooling layers, and a fully connected layer. The final three layers are fully connected, followed by normalization and dropout layers to output the classification results.

[0201] Step S41: for any determined cluster, determine whether the number of SV signals in the cluster is within the range of [1, 20].

[0202] If it is not within the range of [1,20], execute step S42;

[0203] If it is within the range of [1,20], execute step S45;

[0204] Step S42: Continue to determine whether the number of SV signals in the cluster is less than or equal to 100;

[0205] If the number of SV signals in the cluster is less than or equal to 100, step S43 is executed;

[0206] If the number of SV signals in the cluster is greater than 100, step S44 is executed;

[0207] Step S43: directly use the SV signal features within the cluster to construct a feature matrix with a dimension of (100, 25). If the number of SV signals in the cluster is less than 100, it is necessary to obtain the feature matrix through zero-filling operation and add labels to the feature matrix (first, label each SV signal corresponding to the feature matrix (the labeling method is the same as when training the random forest network), and then the SV signal type label of the feature matrix is the signal type with the largest number of corresponding signals in the labeling results);

[0208] The constructed feature matrix is normalized to the minimum and maximum values (i.e., the eigenvalues in the matrix are scaled to the range of [0, 1]) to obtain a processed feature matrix.

[0209] Step S44: Randomly select 100 SV signals from the cluster to execute step S43, and determine whether the number of remaining SV signals in the cluster is in the range of [1, 20].

[0210] If it is not within the range of [1,20], execute step S42;

[0211] If it is within the range of [1,20], execute step S45;

[0212] Step S45: The number of SV signals in the cluster is recorded as w. Each SV signal in the cluster is truncated using a window size smaller than w. The data in the window obtained after each truncation is used as a new SV signal (data enhancement is achieved in this way). A feature matrix is constructed using the features of the SV signal before truncation and the new SV signal after truncation in the cluster, and a label is added to the feature matrix.

[0213] Then perform minimum-maximum normalization on the constructed feature matrix to obtain a processed feature matrix;

[0214] Step S5: process each determined cluster separately using the method of step S4;

[0215] Step S6: Use the obtained processed feature matrix and the added labels to train the convolutional neural network.

[0216] The training used cross entropy as the loss function, Adam as the optimizer algorithm, and a fixed-step learning rate decay strategy. The batch size of the training was set to 256 and the training was stopped after 15 iterations.

[0217] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0218] The collection of SV signals within the alignment relies on the CIGAR string matching rule. Since this stage generates relatively little noise, filtering is unlikely to achieve optimal results. In contrast, the ambiguity of signal similarity measurement and the presence of alleles in clustering lead to a high number of false-positive SVs during the clustering stage. Furthermore, the noise in the data after clustering is amplified, further facilitating the model's learning of false-positive SV features. Therefore, a second round of filtering, targeting the noise within the SV signal clusters within the alignment after clustering, can eliminate false-positive SV signals within the alignment.

[0219] The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. A method for detecting genomic structural variations based on long-read sequencing data and a false positive filtering model, characterized in that: The method specifically comprises the following steps: Step 1: Align the long-read sequencing data to be tested with the reference genome to obtain CIGAR strings and gap information, and then extract SV signals based on the CIGAR strings and gap information; And record the features of the SV signal extracted based on the CIGAR string and the features of the SV signal extracted based on the gap information respectively; Step 2: Use the features of the SV signals extracted based on the gap information to construct an original feature matrix, perform minimum and maximum normalization on the constructed original feature matrix to obtain a processed feature matrix; input the processed feature matrix into the random forest network, and finally determine whether each SV signal extracted based on the gap information is a true SV signal and the type of the true SV signal based on the output of the random forest network; Step 3: Cluster the SV signals extracted from the CIGAR string in Step 1 and the true SV signals determined in Step 2 to obtain clusters containing SV signals; Step 4: For any cluster obtained in step 3, if the cluster contains the real SV signal determined in step 2, the SV signal in the cluster is not processed. Instead, the median of the starting point coordinates of all SV signals in the cluster and the median of the ending point coordinates of all SV signals in the cluster are directly calculated, and the calculated median is used as the final breakpoint coordinate of the candidate SV signal; Otherwise, the feature matrix is constructed using the features of the SV signal in the cluster, and the constructed feature matrix is input into the trained convolutional neural network. The output of the convolutional neural network is used to finally determine whether the cluster is a true SV cluster; If the cluster is a true SV cluster, the median of the starting point coordinates and the median of the ending point coordinates of all SV signals in the cluster are calculated, and the calculated median is used as the final breakpoint coordinate of the candidate SV signal; If the cluster is not a true SV cluster, no processing is required; Similarly, after processing each cluster obtained in step 3 separately, the final breakpoint coordinates of all candidate SV signals are obtained, completing the detection of genomic structural variation.

2. The method for detecting genomic structural variations based on long-read sequencing data and a false positive filtering model according to claim 1, characterized in that: The CIGAR string is obtained as follows: When the long-read sequencing data to be tested is directly aligned with the reference genome, the resulting alignment record is the CIGAR string; The specific process of extracting the SV signal based on the CIGAR character string is as follows: The CIGAR string consists of numbers and letters. The missing and inserted SV signals are located according to the letters in the CIGAR string, and the lengths of the missing and inserted SV signals are determined according to the numbers in the CIGAR string.

3. The method for detecting genomic structural variations based on long-read sequencing data and a false positive filtering model according to claim 2, characterized in that: The gap information is obtained in the following manner: After splitting the long-read sequencing data to be tested into multiple fragments, the bases at both ends of each fragment are the breakpoints of the fragment, and the coordinates of the breakpoints are the gap information; Extracting the SV signal according to the gap information includes identifying the SV signal and extracting the SV signal; wherein: The specific process of identifying SV signals is as follows: (1) For any two adjacent segments, one segment is taken as the main alignment segment and the other segment is taken as the supplementary alignment segment. Denote the distance between the main alignment segment and the supplementary alignment segment in the long read sequencing data to be detected as dis_read, and the distance between the main alignment segment and the supplementary alignment segment in the reference genome as dis_ref; Compare the magnitudes of dis_read and dis_ref: If dis_read < dis_ref, there is a deletion SV signal between the main alignment segment and the supplementary alignment segment; If dis_read > dis_ref, there is an insertion SV signal between the main alignment segment and the supplementary alignment segment; If dis_read = dis_ref, there is no SV signal between the main alignment segment and the supplementary alignment segment; (2) If the alignment positions of the main alignment segment and the supplementary alignment segment on the reference genome overlap, there is a duplication SV signal between the main alignment segment and the supplementary alignment segment; (3) If the directions of the main alignment segment and the supplementary alignment segment relative to the reference genome are different, there is an inversion SV signal between the main alignment segment and the supplementary alignment segment; (4) If the chromosome numbers corresponding to the main alignment segment and the supplementary alignment segment in the reference genome are different, there is a translocation SV signal between the main alignment segment and the supplementary alignment segment; The specific process of extracting the SV signal is: extract the identified SV signal.

4. The method for detecting genomic structural variations based on long-read sequencing data and a false positive filtering model according to claim 3, characterized in that: The distance dis_read between the main alignment segment and the supplementary alignment segment in the long read sequencing data to be detected is: If in the long read sequencing data to be detected, the main alignment segment is before the supplementary alignment segment, the distance dis_read is the difference between the starting base coordinate of the supplementary alignment segment and the ending base coordinate of the main alignment segment; If in the long read sequencing data to be detected, the main alignment segment is after the supplementary alignment segment, the distance dis_read is the difference between the starting base coordinate of the main alignment segment and the ending base coordinate of the supplementary alignment segment.

5. The method for detecting genomic structural variations based on long-read sequencing data and a false positive filtering model according to claim 4, characterized in that: The features of the SV signal extracted according to the CIGAR string include global alignment features, local CIGAR features, and base distribution features; among which: The global alignment features include the alignment quality of the SV signal, the proportion of soft clipping, the length of soft clipping, the FLAG value, the length of the long read sequencing data to be detected, the number of variations, the alignment partition number, the starting position of the SV signal in the long read sequencing data to be detected, the ending position of the SV signal in the long read sequencing data to be detected, the chromosome number of the long read sequencing data to be detected in the reference genome, the starting position of the long read sequencing data to be detected in the reference genome, the ending position of the long read sequencing data to be detected in the reference genome, the edit distance of the alignment, and the length of the mismatched bases; The local CIGAR features include the starting point of the SV signal in the reference genome, the end point of the SV signal in the reference genome, the length of the SV signal, the type of the SV signal, the starting position of the SV signal in CIGAR, and the ending position of the SV signal in CIGAR in the long-read sequencing data to be detected; The base distribution characteristics include the frequencies of occurrence of the four bases in the SV signal and the frequency of occurrence of repeated bases.

6. The method for detecting genomic structural variations based on long-read sequencing data and a false positive filtering model according to claim 5, characterized in that: The features of the SV signal extracted based on the gap information include global alignment features, split alignment features and base distribution features; wherein: The global alignment features are the same as those of the SV signal extracted based on the CIGAR string; The base distribution features are the same as those of the SV signal extracted from the CIGAR string; The split alignment features include the consistency of the corresponding chromosome numbers of the main alignment and the supplementary alignment in the reference genome, the consistency of the directions of the main alignment and the supplementary alignment in the reference genome, the distance d1 between the main alignment and the supplementary alignment in the long-read sequencing data to be tested, the distance d2 between the main alignment and the supplementary alignment in the reference genome, and the difference between the distances d1 and d2.

7. The method for detecting genomic structural variations based on long-read sequencing data and a false positive filtering model according to claim 6, characterized in that: The training process of the random forest network is: Step 1: Align the known long-read sequencing data with the reference genome to obtain gap information, and then extract SV signals based on the gap information; and record the characteristics of the SV signals extracted based on the gap information; Step 2: Divide all SV signals on chromosomes 1 to 11 extracted in step 1 into two parts at a ratio of 7:

3. Use one part as the training set and the other as the validation set. Annotate the SV signals in the training set and validation set according to the benchmark set. Use all SV signals on chromosomes 12 to 22 and sex chromosomes extracted in step 1 as the test set. The marking method of SV signal is: The annotation method for SV signals of deletion, insertion, duplication, and inversion is the same. Take the insertion SV signal as an example: The starting coordinate of any extracted insertion SV signal is marked as s1, and the length of the insertion SV signal is recorded as size1. It is determined whether there is an insertion SV signal in the benchmark set that is on the same chromosome as the insertion SV signal and satisfies the following conditions: Where s2 is the starting coordinate of the inserted SV signal in the benchmark set, |·| represents the absolute value, and size2 is the length of the inserted SV signal in the benchmark set; If there is an inserted SV signal in the benchmark set that is on the same chromosome as the inserted SV signal and meets the conditions, then the extracted inserted SV signal is a true inserted SV signal and the extracted inserted SV signal is marked as an inserted SV signal; otherwise, if there is no inserted SV signal in the benchmark set that is on the same chromosome as the inserted SV signal and meets the conditions, then the extracted inserted SV signal is marked as a false detection; The annotation method for the translocated SV signal is: The starting coordinate of any extracted translocation SV signal is marked as s′1, and it is determined whether there is a translocation SV signal in the reference set that is on the same chromosome as the translocation SV signal and satisfies the following conditions: |s′1-s′2|≤1000 Where s′2 is the starting coordinate of the translocated SV signal in the benchmark set; If there is a translocation SV signal in the benchmark set that is on the same chromosome as the translocation SV signal and meets the conditions, the extracted translocation SV signal is a true translocation SV signal, and the extracted translocation SV signal is marked as a translocation SV signal; otherwise, there is no translocation SV signal in the benchmark set that is on the same chromosome as the translocation SV signal and meets the conditions, and the extracted translocation SV signal is marked as a false detection; Step 3: construct a feature matrix using the SV signal features extracted according to the gap information in step 1, and then perform minimum and maximum normalization processing on the constructed feature matrix to obtain a processed feature matrix; The random forest network is trained using the labeled SV signals and processed feature matrix in the training set.

8. The method for detecting genomic structural variations based on long-read sequencing data and a false positive filtering model according to claim 7, characterized in that: The clustering process in step 3 is performed on each type of SV signal separately; taking the insertion of SV signal as an example: Step 31: Represent each inserted SV signal extracted from the CIGAR string in step 1 and determined in step 2 as a four-tuple (t, c, s, e), where t represents the SV signal type, c represents the chromosome number where the SV signal is located, s represents the starting coordinate of the SV signal in the reference genome, and e represents the ending coordinate of the SV signal in the reference genome; Sort each insertion SV signal: that is, for all insertion SV signals on a chromosome, sort each insertion SV signal in ascending order according to the starting point coordinates of the insertion SV signal in the reference genome. If the spatial distance between two adjacent insertion SV signals is less than 1000bp after sorting, the two adjacent insertion SV signals are divided into the same partition, otherwise the two adjacent insertion SV signals are divided into different partitions; Similarly, the inserted SV signals on each chromosome were sorted and partitioned separately; Step 3.2: Get the SV local depth of each partition and the SV global depth of all partitions: Where a is the SV global depth, b i is the SV local depth of the i-th partition, N is the total number of partitions; Step 3: Cluster the inserted SV signals in the i-th partition, where i = 1, 2, ..., N. The specific clustering process is as follows: Step 331: For the k-th inserted SV signal and the l-th inserted SV signal in the i-th partition, calculate the similarity S(k,l) between the k-th inserted SV signal and the l-th inserted SV signal: Among them, pos_dis is the site distance between the kth inserted SV signal and the lth inserted SV signal, span_dis is the span distance between the kth inserted SV signal and the lth inserted SV signal, s k and e k are the starting and ending coordinates of the kth insertion SV signal in the reference genome, s l and e l are the starting and ending coordinates of the lth inserted SV signal in the reference genome, respectively, and λ is the correction parameter; span_dis=|(e k -s k )-(e l -s l )| Step 332: Perform hierarchical clustering on the inserted SV signals in the i-th partition according to the calculated similarity to obtain each cluster; Step 34: Perform the process of step 31 to step 33 respectively on the missing SV signal, repeated SV signal, inverted SV signal and transposed SV signal extracted according to the CIGAR character string in step 1 and determined in step 2.

9. The method for detecting genomic structural variations based on long-read sequencing data and a false positive filtering model according to claim 8, characterized in that: After the sorting, the spatial distance between two adjacent inserted SV signals is s k+1 -s k , where s k is the starting coordinate of the kth inserted SV signal, s k+1 is the starting coordinate of the k+1th inserted SV signal; The method for calculating the spatial distance between two adjacent translocated SV signals is the same as that for calculating the spatial distance between two adjacent inserted SV signals; The spatial distance between two adjacent missing SV signals is s′ k+1 -e′ k , where e′ k is the end point coordinate of the kth missing SV signal, s′ k+1 is the starting coordinate of the k+1th missing SV signal; The calculation method of the spatial distance between two adjacent inverted SV signals and the spatial distance between two adjacent repeated SV signals is the same as the calculation method of the spatial distance between two adjacent missing SV signals.

10. The method for detecting genomic structural variations based on long-read sequencing data and a false positive filtering model according to claim 9, characterized in that: The training method of the convolutional neural network is: Step S1: Align the known long-read sequencing data with the reference genome to obtain a CIGAR string, then extract the SV signal based on the CIGAR string; and record the characteristics of the SV signal extracted based on the CIGAR string; Step S2: extracting SV signal features based on the gap information of the known long-read sequencing data in step S1, constructing a feature matrix using the features of the SV signals extracted based on the gap information, performing minimum and maximum normalization processing on the constructed feature matrix, inputting the processed feature matrix into a random forest network, and predicting the true SV signal from the extracted SV signals through the random forest network; Step S3, clustering the SV signal extracted according to the CIGAR character string in step S1 and the true SV signal determined in step S2 together to obtain clusters containing SV signals; Step S4: Determine the clusters of SV signals extracted from the comparison, and train the convolutional neural network based on the feature matrix and labels of each cluster; specifically: Step S41: for any determined cluster, determine whether the number of SV signals in the cluster is within the range of [1, 20]. If it is not within the range of [1,20], execute step S42; If it is within the range of [1,20], execute step S45; Step S42: Continue to determine whether the number of SV signals in the cluster is less than or equal to 100; If the number of SV signals in the cluster is less than or equal to 100, step S43 is executed; If the number of SV signals in the cluster is greater than 100, step S44 is executed; Step S43: directly use the SV signal features in the cluster to construct a feature matrix with a dimension of (100, 25). If the number of SV signals in the cluster is less than 100, it is necessary to obtain the feature matrix through zero filling operation and add labels to the feature matrix; And perform minimum-maximum normalization processing on the constructed feature matrix to obtain a processed feature matrix; Step S44: Randomly select 100 SV signals from the cluster to execute step S43, and determine whether the number of remaining SV signals in the cluster is in the range of [1, 20]. If it is not within the range of [1,20], execute step S42; If it is within the range of [1,20], execute step S45; Step S45: The number of SV signals in the cluster is recorded as w. Each SV signal in the cluster is truncated using a window size smaller than w. The data in the window obtained after each truncation is used as a new SV signal. A feature matrix is constructed using the features of the SV signal before truncation and the new SV signal after truncation, and a label is added to the feature matrix. Then perform minimum-maximum normalization on the constructed feature matrix to obtain a processed feature matrix; Step S5: process each determined cluster separately using the method of step S4; Step S6: Use the obtained processed feature matrix and the added labels to train the convolutional neural network.

Images