A deep learning-based genomic detection data analysis method and system

Abstract

The application discloses a kind of based on deep learning's genomic detection data analysis method and system, it is related to bioinformatics data processing technical field;The method comprises: obtaining sequencing reads alignment file, reference genome sequence and candidate variation set, extract the local coverage reads of candidate variation site and divide into reference allele reads set and alternative allele reads set;According to the local uniqueness of the reference sequence on both sides of candidate variation site, determine candidate anchor side and low alignable side;According to candidate anchor side and low alignable side, extract anchor-related read evidence, construct one-side anchor evidence tensor;It is input into double-branch deep learning model, and the probability of real variation and one-side anchor false variation probability are output, and generate one-side anchor imbalance interpretation evidence, output genomic detection data analysis result.The application can improve the reliability of distinguishing real variation and suspected false variation in complex genomic region.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics data processing technology, and in particular to a method and system for analyzing genome detection data based on deep learning. Background Technology

[0002] Genome sequencing typically requires obtaining sequencing reads, aligning them to a reference genome sequence, and then generating candidate variant results through processes such as candidate variant identification, quality filtering, and result annotation. In this process, candidate variant sites are not directly equivalent to the actual genotype differences in the sample. Their reliability is affected by sequencing quality, read length, alignment quality, local sequence structure of the reference genome, and the comparability of the candidate variant sites on both sides.

[0003] Existing genomic sequencing data analysis methods typically filter candidate variants based on indicators such as variant quality, sequencing depth, base quality, strand bias, read position bias, or alignment quality. Some methods also employ deep learning models to classify and judge read stacking diagrams or candidate variant characteristics. While these methods can identify some low-quality candidate variants, sequencing reads are prone to alignment instability, local soft shearing, insertion / deletion slippage, or end mismatches in homopolymer regions, low-complexity sequence regions, fragment repetition regions, pseudogene homologous regions, regions with abnormal coverage due to high or low GC, and near regions of local collapse in the reference genome. Especially at low-alignment boundaries, substituted allele reads may rely mainly on the unique sequence on one side of the candidate variant site for alignment anchoring, while the other side falls into the low-alignment region, thus forming suspected pseudo-variants with some read depth support and local consistency. However, in the analysis of genome detection data, it is difficult to fully explain and identify suspected spurious variants caused by the imbalance of unilateral anchoring of substitution alleles at low-comparability boundary based solely on the overall quality indicators of candidate variant sites or the stacking characteristics of ordinary reads. This can easily lead to confusion between real variants and spurious variants driven by regional bias. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies where, at low-comparability boundaries, substituted allele reads are prone to forming suspected pseudo-variants with a certain read depth support due to unilateral anchoring imbalance, making it difficult to reliably distinguish between real variants and region-biased pseudo-variants. Therefore, this invention proposes a genome detection data analysis method and system based on deep learning.

[0005] To address the problems existing in the prior art, the present invention adopts the following technical solution:

[0006] A deep learning-based method for analyzing genome detection data includes:

[0007] S1. Obtain the sequencing read alignment file, reference genome sequence, and candidate variant set of the sample to be analyzed. Extract local coverage reads from the candidate variant sites in the candidate variant set and divide them into a reference allele read set and a substitute allele read set.

[0008] S2. Based on the local uniqueness of the reference sequences on both sides of the candidate variant site, determine the candidate anchoring side and the low comparability pair side;

[0009] S3. Based on the candidate anchoring side and the low comparable side, extract anchoring-related read evidence from the reference allele read set and the alternative allele read set, and construct a one-sided anchoring evidence tensor in combination with the local uniqueness.

[0010] S4. Input the one-sided anchored evidence tensor into the dual-branch deep learning model and output the true mutation probability and the one-sided anchored pseudo mutation probability.

[0011] S5. Generate one-sided anchoring imbalance explanation evidence based on the true mutation probability, the one-sided anchoring pseudo mutation probability, and the one-sided anchoring evidence tensor, and output the genome detection data analysis results based on the true mutation probability, the one-sided anchoring pseudo mutation probability, and the one-sided anchoring imbalance explanation evidence.

[0012] Preferably, the locally covered read is a read that covers the candidate variant site in the sequencing read alignment file; locally covered reads supporting the reference allele are assigned to the reference allele read set, and locally covered reads supporting the alternative allele are assigned to the alternative allele read set.

[0013] Preferably, the method further includes determining bilateral analysis windows on both sides of the candidate variant site based on the median length of the reads covering the candidate variant site, the length of the reference allele, and the length of the alternative allele; the bilateral analysis windows include a left window located to the left of the candidate variant site and a right window located to the right of the candidate variant site.

[0014] Preferably, the local uniqueness is obtained in the following manner:

[0015] The reference sequence in the left and right windows is split into continuous subsequences whose length is determined by the effective alignment length of the locally covered read. The occurrence frequency of the continuous subsequence covering the corresponding position in the reference genome sequence is counted, and the local uniqueness of the corresponding position is obtained based on the occurrence frequency. The more occurrences, the lower the local uniqueness of the corresponding position, and the fewer occurrences, the higher the local uniqueness of the corresponding position.

[0016] The average uniqueness of the left side and the average uniqueness of the right side are obtained based on the local uniqueness of each position in the left and right windows, respectively.

[0017] The side with higher average uniqueness is identified as the candidate anchoring side, and the side with lower average uniqueness is identified as the low comparable side.

[0018] When the average uniqueness of the left side and the average uniqueness of the right side are the same, a bilateral uniqueness balance mark is generated, and the bilateral uniqueness balance mark is used as a regional context channel to participate in the construction of the subsequent unilateral anchored evidence tensor.

[0019] Preferably, the anchoring-related read evidence includes the matching length of the locally covering read in the left window, the matching length in the right window, the soft-splitting length in the left window, the soft-splitting length in the right window, the insertion or deletion length in the bilateral analysis window, the base error probability at the candidate variant site, the alignment quality in the bilateral analysis window, and the supported allele type.

[0020] The base error probability is calculated from the base quality fraction of the locally covered read at the candidate variant site; the supported allele types include reference allele types and alternative allele types, the reference allele type corresponds to the reference allele read set, and the alternative allele type corresponds to the alternative allele read set.

[0021] Preferably, the one-sided anchored evidence tensor is constructed as follows: anchored relevant read evidence in the reference allele read set is arranged into a reference allele evidence matrix; anchored relevant read evidence in the alternative allele read set is arranged into an alternative allele evidence matrix; and left-side average uniqueness, right-side average uniqueness, and bilateral uniqueness balance markers are used as regional context channels, which are combined with the reference allele evidence matrix and the alternative allele evidence matrix to form the one-sided anchored evidence tensor; wherein, when bilateral uniqueness balance markers are not generated, the regional context channels include left-side average uniqueness and right-side average uniqueness.

[0022] Preferably, the dual-branch deep learning model includes a read evidence branch and a boundary context branch; the read evidence branch receives the reference allele evidence matrix and the alternative allele evidence matrix, and extracts the differences between the reference allele read and the alternative allele read in terms of left-right matching length, left-right soft splicing length, insertion or deletion length, base error probability, and alignment quality; the boundary context branch receives the regional context channel and extracts the difference between the left-side average uniqueness and the right-side average uniqueness; after fusing the outputs of the read evidence branch and the boundary context branch, the true mutation probability and the one-sided anchored pseudo-mutation probability are obtained.

[0023] Preferably, the evidence for interpreting the unilateral anchoring imbalance includes: evidence characterizing the set of alternative allele reads that depend on the candidate anchoring side for matching; evidence characterizing the presence of soft-splitting sets of alternative allele reads on the low-comparable side; evidence characterizing the presence of insertion or deletion sets of alternative allele reads on the low-comparable side; evidence characterizing the directional difference between the alternative allele reads and the reference allele reads in terms of left and right matching lengths; and evidence characterizing the candidate variant site located at the boundary of a region with different left-side average uniqueness and right-side average uniqueness.

[0024] Preferably, the genomic detection data analysis results include candidate variant sites, variant types, reference alleles, substitute alleles, candidate anchoring sides, low-comparable counterparts, true variant probability, one-sided anchoring pseudo-variable probability, classification results, and evidence explaining one-sided anchoring imbalance. Specifically, when the evidence explaining one-sided anchoring imbalance indicates that substitute allele reads are concentrated and match the candidate anchoring side, and exhibit a concentrated distribution of soft splicing or insertion / deletion on the low-comparable counterpart relative to the reference allele reads, the classification result is output as a suspected one-sided anchoring pseudo-variable. When the evidence explaining one-sided anchoring imbalance indicates that substitute allele reads match on both sides of the candidate variant site, and do not exhibit one-sided anchoring imbalance relative to the reference allele reads, the classification result is output as a true variant. When there is no single higher probability between the true variant probability and the one-sided anchoring pseudo-variable probability, or when the probability classification result determined by the higher of the true variant probability and the one-sided anchoring pseudo-variable probability differs from the evidence classification result corresponding to the evidence explaining one-sided anchoring imbalance, the classification result is output as a variant awaiting verification.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] 1. This invention obtains the locally covering reads corresponding to candidate variant sites and classifies them into a set of reference allele reads and a set of alternative allele reads. Then, based on the local uniqueness of the reference sequences on both sides of the candidate variant site, it determines the candidate anchoring side and the low-comparability side. This makes the analysis process no longer dependent on the overall quality, read depth, or a single alignment quality indicator of the candidate variant. Instead, it explicitly introduces the sources of read alignment instability in complex genomic regions into the candidate variant judgment process. It can further identify whether alternative allele reads mainly rely on a unique sequence on one side to complete the alignment anchoring, addressing the read mismatch, soft splicing, and insertion / deletion slippage problems that are prone to occur in regions such as homopolymers, low-complexity sequences, fragment duplications, pseudogene homology, and local collapse of the reference genome. This solves the problem of reliably distinguishing between true variants and pseudo-variants driven by regional bias in complex regions.

[0027] 2. This invention further constructs a one-sided anchoring evidence tensor by using the matching length, soft-splitting length, insertion or deletion length, base error probability, alignment quality, and local uniqueness information of the reference allele read and the alternative allele read on the candidate anchoring side and the low-comparability side. This tensor is then input into a two-branch deep learning model to output the true variant probability and the one-sided anchoring pseudo-variable probability. Simultaneously, it generates one-sided anchoring imbalance explanation evidence through channel masking. This not only utilizes the deep learning model to learn the directional anchoring differences between the reference allele read and the alternative allele read, but also outputs the basis for judging candidate variants as true variants, suspected one-sided anchoring pseudo-variables, or variants to be verified. This avoids simply filtering or directly retaining candidate variants in complex regions, improving the reliability, interpretability, and verification convenience of the genomic detection data analysis results. Attached Figure Description

[0028] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0029] Figure 1 This is a flowchart illustrating a deep learning-based genome detection data analysis method according to an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0031] Example 1, as Figure 1 As shown, this embodiment provides a deep learning-based method for analyzing genome detection data; this method is used to analyze candidate variant sites obtained during genome detection, especially to identify suspected spurious variants caused by unilateral anchoring imbalance of substitution alleles at low-comparability boundary.

[0032] In this embodiment, the genome detection data includes a sequencing read alignment file of the sample to be analyzed, a reference genome sequence, and a candidate variant set. The sequencing read alignment file is a file that records the alignment relationship between the sequencing read and the reference genome sequence, and the sequencing read alignment file includes, but is not limited to, a BAM file or a CRAM file. The reference genome sequence is a reference sequence corresponding to the sequencing read alignment file. The candidate variant set is a set of candidate variant sites formed after preliminary variant identification, and the candidate variant set includes candidate single nucleotide variants, candidate insertion variants, and candidate deletion variants.

[0033] In this embodiment, the candidate variant set refers to the set of candidate variant records obtained from the preliminary variant identification process that have not yet undergone one-sided anchorage imbalance analysis using the method of this invention; each candidate variant record includes at least a candidate variant site, a reference allele, a substitute allele, and a variant type; the candidate variant site refers to the genomic location or genomic region in the candidate variant record whose authenticity needs to be determined; when the variant type is a single nucleotide variant, the candidate variant site is a single base position; when the variant type is an insertion variant or a deletion variant, the candidate variant site is the reference genomic anchorage point or region corresponding to the insertion variant or deletion variant;

[0034] In this embodiment, a partially covered read refers to a sequencing read that covers a candidate variant site in the sequencing read alignment file. In this embodiment, the coverage is not limited to the read bases falling directly on the candidate variant site, but also includes the aligned fragment of the read crossing the candidate variant site, the read supporting the candidate variant site through insertion or deletion alignment, and the read forming alignment evidence in the vicinity of the candidate variant site that can determine the support relationship of the reference allele or the alternative allele. Reads that only have alignment in the vicinity of the candidate variant site but cannot determine the allele support relationship are not considered as valid reads in the reference allele read set or the alternative allele read set.

[0035] The reference allele read set refers to the set of locally covering reads that support the reference allele at the candidate variant site; the alternative allele read set refers to the set of locally covering reads that support the alternative allele at the candidate variant site. In this embodiment, supporting the reference allele means that the base, fragment continuity, or deletion spanning relationship of the locally covering read at the candidate variant site is consistent with the reference genome sequence; supporting the alternative allele means that the base, insertion, deletion connection, or substitution fragment of the locally covering read at the candidate variant site is consistent with the alternative allele recorded in the candidate variant set. For the same locally covering read, it is not included in both the reference allele read set and the alternative allele read set at the same time. When a read cannot clearly support any allele due to low-quality bases, severe soft splicing, or multiple alignment, the read is marked as an unclassified read and is not included as a member of either type of read set.

[0036] Candidate anchoring side refers to the side of the reference sequence on both sides of the candidate variant site with high local uniqueness, which can provide a relatively stable alignment anchor for the read; low comparability side refers to the side of the reference sequence on both sides of the candidate variant site with low local uniqueness, which is prone to duplicate alignment, homology confusion, soft splicing or local insertion / deletion instability.

[0037] The unilateral anchoring imbalance of alternative alleles at low-comparability boundaries, as referred to in this embodiment, means that when a candidate variant site is adjacent to a homopolymer region, a low-complexity sequence region, a fragment repetition region, a pseudogene homologous region, a region with abnormal coverage due to high or low GC, or a region with local collapse of the reference genome, the alternative allele read mainly relies on the unique sequence on one side of the candidate variant site to form a matching anchor, while the other side of the candidate variant site falls into a low-comparability region, accompanied by soft splicing, local insertions or deletions, end mismatches, or decreased alignment quality; at the same time, the reference allele read does not show the same degree of unilateral anchoring imbalance; this feature is used to explain why some suspected pseudovariables have a certain read depth support but are not true genomic variations.

[0038] The method in this embodiment includes the following steps;

[0039] S1. Obtain the sequencing read alignment file, reference genome sequence, and candidate variant set of the sample to be analyzed. Extract local coverage reads for candidate variant sites in the candidate variant set and divide them into a reference allele read set and a substitute allele read set.

[0040] Specifically, the sequencing read alignment file, reference genome sequence, and candidate variant set corresponding to the sample to be analyzed are obtained; each candidate variant record in the candidate variant set includes at least the chromosome number, candidate variant site, reference allele, substitution allele, and variant type; variant types include single nucleotide variants, insertion variants, and deletion variants;

[0041] For any candidate variant site in the candidate variant set, read the sequencing reads that overlap with the candidate variant site from the sequencing read alignment file and treat these reads as local coverage reads. If the base or fragment state of a local coverage read at the candidate variant site is consistent with the reference allele, then the local coverage read is assigned to the reference allele read set. If the base or fragment state of a local coverage read at the candidate variant site is consistent with the alternative allele, then the local coverage read is assigned to the alternative allele read set. For local coverage reads that cannot clearly support the reference allele or alternative allele, retain their alignment position and quality information, but do not include them in the reference allele read set or alternative allele read set for reference during subsequent verification.

[0042] When the candidate variant is a single nucleotide variant, if the bases of the read at the candidate variant site are the same as those of the reference allele, then the read is a reference allele read; if the bases of the read at the candidate variant site are the same as those of the substitution allele, then the read is a substitution allele read. When the candidate variant is an insertion variant, if the read has an insertion fragment near the candidate variant site that corresponds to the substitution allele, then the read is a substitution allele read; if the read does not contain the insertion fragment and matches the reference allele consecutively, then the read is a reference allele read. When the candidate variant is a deletion variant, if the read shows a deletion connection near the candidate variant site that corresponds to the substitution allele, then the read is a substitution allele read; if the read covers the region and retains the reference allele fragment, then the read is a reference allele read.

[0043] To determine the analysis range flanking the candidate variant site, this embodiment determines the bilateral analysis window based on the median read length covering the candidate variant site, the reference allele length, and the alternative allele length; let the median read length covering the candidate variant site be... Reference allele length is The length of the alternative allele is The analytical radius on both sides of the candidate variant site is ,but:

[0044] ;

[0045] in, The left and right windows represent the lengths of the candidate variant sites; the left window represents the length to the left of the candidate variant site. The reference genome region, with the right window showing candidate variant sites of length [missing information]. The reference genome region;

[0046] In this embodiment, the bilateral analysis window refers to two local analysis intervals extending to the left and right sides of the reference genome sequence around the candidate variant site, respectively. The bilateral analysis window includes a left window and a right window. The left and right sides are determined according to the increasing direction of the reference genome coordinates, with the side whose coordinates are less than the candidate variant site being the left side and the side whose coordinates are greater than the candidate variant site being the right side. For insertion or deletion variants, the left and right windows are determined based on the reference genome anchor point or interval boundary corresponding to the candidate variant site.

[0047] The above method adapts the two-sided analysis window to the actual read length and variation length, avoiding the inconsistency in the analysis range between samples with different read lengths caused by using a fixed window.

[0048] S2. Based on the local uniqueness of the reference sequences on both sides of the candidate variant site, determine the candidate anchoring side and the low comparability pair side;

[0049] Specifically, the reference genome sequence is read in the left and right windows obtained in S1, and the reference sequence in the left and right windows is split into continuous subsequences; the length of the continuous subsequence is determined by the effective alignment length of the local coverage read.

[0050] Furthermore, the effective alignment length is obtained by reading the alignment description information of the locally covered reads; the alignment description information is used to record the matching, mismatch, insertion, deletion, soft splicing, and hard splicing status of the locally covered reads relative to the reference genome sequence; for any locally covered read, the length of the fragment corresponding to the matching or equivalent matching in the alignment description information is included in the effective alignment length, while soft splicing fragments, hard splicing fragments, and unaligned fragments are not included in the effective alignment length; for inserted and deleted fragments, they are used for subsequent anchoring of related read evidence extraction and are not used as the main basis for determining the length of continuous subsequences;

[0051] When multiple locally covered reads exist for the same candidate variant site, the effective alignment length of each locally covered read is calculated separately, and the median of the effective alignment lengths of each locally covered read is taken as the length of the continuous subsequence. If there is a difference in the distribution of effective alignment lengths between the reference allele read set and the alternative allele read set, the locally covered reads in the reference allele read set and the alternative allele read set are merged and the median of the effective alignment length is calculated so that the length of the continuous subsequence reflects the local alignment of the two types of allele reads simultaneously.

[0052] In this way, the length of the continuous subsequence is derived from the effective alignment relationship of the sequencing read itself. This setting can match the local uniqueness calculation with the comparable length of the real read near the candidate variant site, thereby avoiding the misalignment evaluation caused by using a fixed continuous subsequence length under different sequencing read lengths, different soft splicing ratios and different variant lengths.

[0053] For any reference genome location within the window The number of occurrences of the continuous subsequence covering this position in the reference genome sequence is counted and denoted as . The local uniqueness of a position is obtained based on the frequency of its occurrence. :

[0054] ;

[0055] in, Indicates position Local uniqueness, Indicates the coverage location The number of times a continuous subsequence appears in the reference genome sequence; the more times it appears, the more easily the sequence at that position is confused with other regions in the reference genome, and the lower the local uniqueness of that position; the fewer times it appears, the more unique the sequence at that position is, and the higher the local uniqueness of that position.

[0056] In practical implementation, the number of occurrences of the continuous subsequence in the reference genome sequence can be obtained through a pre-established reference genome continuous subsequence index table or a pre-established reference genome comparability index. The reference genome continuous subsequence index table records the correspondence between the continuous subsequence and its position in the reference genome sequence. By calling the reference genome continuous subsequence index table or the reference genome comparability index, it is possible to avoid repeatedly performing site-by-site scanning of the entire reference genome when analyzing each candidate variant site, thereby improving the feasibility and operational efficiency of local uniqueness calculation.

[0057] Calculate the average local uniqueness of each position within the left and right windows respectively; let the left window be... The right window is The average uniqueness on the left is The average uniqueness on the right is ,but:

[0058] ;

[0059] ;

[0060] in, This indicates the number of positions within the left-hand window. Indicates the number of positions within the right-hand window;

[0061] When the average uniqueness on the left is greater than the average uniqueness on the right, the left window is designated as the candidate anchoring side and the right window as the low-comparability side; when the average uniqueness on the right is greater than the average uniqueness on the left, the right window is designated as the candidate anchoring side and the left window as the low-comparability side; when the average uniqueness on the left and the average uniqueness on the right are the same, a bilateral uniqueness balance marker is generated, and the average uniqueness on the left, the average uniqueness on the right, and the bilateral uniqueness balance marker are retained as regional context information to participate in the subsequent construction of the unilateral anchoring evidence tensor.

[0062] Local uniqueness in this step is used to describe the localization ability of the reference sequences flanking the candidate variant site to the read. In this embodiment, local uniqueness refers to the degree to which the reference sequences flanking the candidate variant site are uniquely located in the reference genome sequence. The higher the local uniqueness, the lower the repetition of the continuous subsequence in the region in the reference genome sequence, and the easier it is for the sequencing read to be stably aligned to the region. The lower the local uniqueness, the higher the repetition of the continuous subsequence in the region in the reference genome sequence, and the more likely the sequencing read is to generate multiple alignments, incorrect alignments, or unstable local alignments.

[0063] By defining the side with higher average uniqueness as the candidate anchoring side and the side with lower average uniqueness as the low comparable side, subsequent analysis can focus on whether the alternative allele reads exhibit an imbalance pattern of anchoring on one side and instability on the other.

[0064] In this embodiment, the candidate anchoring side does not necessarily mean that there are no sequencing errors or alignment errors on that side. Rather, it means that when comparing the two sides of the candidate variant site, the reference sequence on that side has higher local uniqueness and can provide a relatively stable alignment and localization basis for locally covered reads. The low comparability side does not mean that the reads on that side are completely uncomparable. Rather, it means that when comparing the two sides of the candidate variant site, the reference sequence on that side has lower local uniqueness and is more prone to soft splicing, insertion / deletion slip, multiple alignments, or decreased alignment quality.

[0065] S3. Extract anchoring-related read evidence from the reference allele read set and the alternative allele read set, and construct a one-sided anchoring evidence tensor in combination with the local uniqueness.

[0066] Specifically, for candidate variant sites, alignment information of locally covered reads is read one by one from the reference allele read set and the alternative allele read set; for any locally covered read, its matching length in the left window, matching length in the right window, soft splice length in the left window, soft splice length in the right window, insertion or deletion length in the bilateral analysis window, base quality fraction at the candidate variant site, alignment quality in the bilateral analysis window, and supported allele type are extracted;

[0067] In this embodiment, the anchor-related read evidence refers to read evidence that reflects the alignment and localization status of locally covered reads on both sides of the candidate variant site; the anchor-related read evidence is not an indiscriminate summary of all sequencing information of the candidate variant site, but includes at least evidence that can characterize whether the read depends on one side of the candidate variant site to form a match, whether soft splicing occurs on the other side, whether insertion or deletion occurs within the bilateral analysis window, and whether it is affected by the probability of base error and alignment quality;

[0068] Among them, the matching length refers to the length of the local coverage read that matches the reference genome sequence within the corresponding window; the soft splice length refers to the length of the local coverage read that is marked as soft spliced ​​by the alignment program within the corresponding window; the insertion or deletion length refers to the length of the inserted or deleted bases of the local coverage read relative to the reference genome sequence within the bilateral analysis window; the alignment quality is used to characterize the reliability of the read alignment position; the supported allele types include the reference allele type and the alternative allele type, with the reference allele type corresponding to the reference allele read set and the alternative allele type corresponding to the alternative allele read set.

[0069] For the base mass fraction at candidate variant sites in partially covered reads Convert it to base error probability :

[0070] ;

[0071] in, This indicates the probability of a sequencing error occurring at a candidate variant site. This represents the base quality fraction; the base error probability can be used to convert base quality into a numerical feature that directly corresponds to the probability of error, enabling the model to distinguish between random noise caused by low-quality bases and anchoring imbalance caused by low-comparability boundary.

[0072] In a candidate variant site, anchor-related read evidence from the reference allele read set is arranged into a reference allele evidence matrix, and anchor-related read evidence from the alternative allele read set is arranged into an alternative allele evidence matrix; each row in the reference allele evidence matrix corresponds to one reference allele read, and each row in the alternative allele evidence matrix corresponds to one alternative allele read; each column corresponds to one type of anchor-related read evidence.

[0073] When arranging reads, they are sorted according to their start position, end position, or relative distance from candidate variant sites, so that the model can preserve the local read distribution relationship. For the reference allele evidence matrix and the alternative allele evidence matrix within the same candidate variant site, the same sorting rule is used to avoid distortion of the comparison relationship between alleles due to different read arrangement.

[0074] Furthermore, to ensure that the one-sided anchoring evidence tensors corresponding to different candidate variant sites can be input into the same bi-branch deep learning model, this embodiment performs unified dimensional processing on the reference allele evidence matrix and the alternative allele evidence matrix. For any candidate variant site, the local coverage reads are first sorted according to the relative distance between the local coverage reads and the candidate variant site, the alignment quality, and the allele support relationship. Among them, the local coverage reads with higher overlap with the candidate variant site, higher alignment quality, and clear support for the reference allele or alternative allele are ranked first.

[0075] When the number of reads in the reference allele read set or the alternative allele read set exceeds the uniform number of rows required for the model input, the read evidence at the beginning is truncated according to the sorting result; when the number of reads in the reference allele read set or the alternative allele read set is less than the uniform number of rows required for the model input, padding rows are added to the corresponding evidence matrix; the matching length, soft shearing length, insertion or deletion length, base error probability, and alignment quality corresponding to the padding rows are set to read-free placeholder values, and placeholder marker channels are set in the one-sided anchored evidence tensor; the placeholder marker channels are used to identify whether each row in the evidence matrix is ​​a true read row or a padding read row.

[0076] In this embodiment, the placeholder value for no reads refers to a value used to fill empty rows in the evidence matrix and does not represent actual sequencing read evidence; the placeholder marker channel refers to a channel used to distinguish between actual read rows and filled read rows; when a row corresponds to an actual read, the placeholder marker channel marks that row as an actual read row; when a row corresponds to a filled row, the placeholder marker channel marks that row as a filled read row; when the dual-branch deep learning model reads the unilateral anchoring evidence tensor, it identifies and excludes the influence of filled read rows on the unilateral anchoring imbalance judgment based on the placeholder marker channel;

[0077] The uniform row number can be determined based on the distribution of local coverage reads of candidate variant sites in the training samples, or based on the target coverage depth of the sequencing platform. After determining the uniform row number, the same uniform row number, the same sorting rules, the same truncation rules, and the same padding rules are used in the training and inference processes. The uniform row number is determined before model training and is fixedly saved as the input configuration parameter of the two-branch deep learning model. In the model inference stage, the samples to be analyzed are truncated or padded using the same uniform row number as in the training stage. Through this processing, different candidate variant sites can still form a one-sided anchored evidence tensor with consistent dimensions and meaning, even when the read depth, the number of reference allele reads, and the number of substitute allele reads are different.

[0078] Furthermore, to avoid the padding rows affecting the model's learning of real read evidence, the dual-branch deep learning model reads the placeholder marker channels simultaneously when reading the reference allele evidence matrix and the alternative allele evidence matrix; for matrix rows marked as padding read rows by the placeholder marker channels, they are not used as valid evidence of alternative allele reads or reference allele reads in the one-sided anchoring imbalance judgment.

[0079] The left-side average uniqueness, right-side average uniqueness, and bilateral uniqueness balance markers are used as regional context channels, which are combined with the reference allele evidence matrix and the alternative allele evidence matrix to form a one-sided anchored evidence tensor. When bilateral uniqueness balance markers are not generated, the regional context channels include left-side average uniqueness and right-side average uniqueness; when bilateral uniqueness balance markers are generated, the regional context channels include left-side average uniqueness, right-side average uniqueness, and bilateral uniqueness balance markers.

[0080] In this embodiment, the one-sided anchoring evidence tensor is not a regular variant pileup image, but a structured input constructed around the comparability of both sides of the candidate variant site, the reference allele read anchoring method, and the alternative allele read anchoring method.

[0081] In this embodiment, the one-sided anchoring evidence tensor refers to a model input data structure constructed on a per-candidate variant site basis, which includes at least a reference allele evidence matrix, an alternative allele evidence matrix, and a regional context channel. The reference allele evidence matrix represents anchoring-related read evidence in the reference allele read set, the alternative allele evidence matrix represents anchoring-related read evidence in the alternative allele read set, and the regional context channel represents the local uniqueness state on both sides of the candidate variant site. The one-sided anchoring evidence tensor enables the two-branch deep learning model to simultaneously obtain allele read differences and comparability differences on both sides of the candidate variant site.

[0082] This tensor includes at least the following information: left-side matching information of the reference allele read, right-side matching information of the reference allele read, left-side matching information of the alternative allele read, right-side matching information of the alternative allele read, soft-splitting information of the reference allele read, soft-splitting information of the alternative allele read, insertion or deletion information, base error probability information, alignment quality information, left-side average uniqueness information, right-side average uniqueness information, bilateral uniqueness balance marker information, and placeholder marker channel information.

[0083] S4. Input the one-sided anchored evidence tensor into the dual-branch deep learning model and output the true mutation probability and the one-sided anchored pseudo mutation probability.

[0084] Specifically, the obtained one-sided anchored evidence tensor is input into a two-branch deep learning model; the two-branch deep learning model includes a read segment evidence branch and a boundary context branch;

[0085] The read evidence branch receives a reference allele evidence matrix and an alternative allele evidence matrix. This branch is used to extract differences between the reference allele reads and the alternative allele reads in terms of left-right matching length, left-right soft splicing length, insertion or deletion length, base error probability, and alignment quality. Specifically, the read evidence branch extracts features from both the reference allele evidence matrix and the alternative allele evidence matrix to obtain reference read features and alternative read features. These features are then differentially fused to obtain allele read difference features. These allele read difference features reflect whether the alternative allele read exhibits a unilateral anchoring imbalance relative to the reference allele read.

[0086] The boundary context branch receives the region context channel; the boundary context branch is used to extract the difference between the left average uniqueness and the right average uniqueness, as well as the bilateral comparability state represented by the bilateral uniqueness balance label; when the left average uniqueness is higher than the right average uniqueness, the boundary context branch outputs the boundary context features where the left side is the candidate anchor side and the right side is the low comparability side; when the right average uniqueness is higher than the left average uniqueness, the boundary context branch outputs the boundary context features where the right side is the candidate anchor side and the left side is the low comparability side; when generating the bilateral uniqueness balance label, the boundary context branch outputs the boundary context features with bilateral uniqueness balance.

[0087] The allele read difference features output by the read evidence branch and the boundary context features output by the boundary context branch are fused to obtain the fused features of candidate variant sites. The fused features are then classified through a classification layer to output the true variant probability. and one-sided anchored pseudo-variation probability ;

[0088] In this embodiment, the true mutation probability represents the probability that a candidate mutation site belongs to a true mutation; the one-sided anchorage pseudo mutation probability represents the probability that a candidate mutation site belongs to a suspected pseudo mutation caused by one-sided anchorage imbalance of the substitution allele at the low-comparability boundary.

[0089] In this embodiment, both the true mutation probability and the one-sided anchored pseudo mutation probability are classification probabilities output by a two-branch deep learning model for the same candidate mutation site. The true mutation probability is used to characterize the possibility that the candidate mutation site is formed by the difference in the true genotype of the sample. The one-sided anchored pseudo mutation probability is used to characterize the possibility that the candidate mutation site is formed by the one-sided anchoring imbalance of the substitute allele read at the low-comparability boundary. The above two probabilities are used to generate the probability classification results and, together with the evidence for the interpretation of the one-sided anchoring imbalance, determine the results of the genome detection data analysis.

[0090] Two-branch deep learning models can be trained using training samples with credible labels. The training samples include real variant samples and one-sided anchored pseudo-variable samples. Real variant samples can be confirmed by a set of standard sample truth values, manual verification results, long-read sequencing verification results, or pedigree consistency verification results. One-sided anchored pseudo-variable samples can be composed of candidate variants located in low-alignment regions, homopolymer regions, low-complexity regions, repetitive regions, or homology collapse regions that are not supported by a set of credible truth values, and confirmed by manual verification or supplementary verification results.

[0091] Furthermore, the training samples are constructed using the same data processing workflow as the samples to be analyzed. For each candidate variant site in the training samples, local coverage reads are extracted according to S1 and divided into a reference allele read set and a substitute allele read set. The left average uniqueness and right average uniqueness are calculated according to S2, and candidate anchoring sides and low-comparability opposite sides are determined. Anchoring-related read evidence is extracted according to S3, and a one-sided anchoring evidence tensor is constructed. Each training sample corresponds to a one-sided anchoring evidence tensor and a training label.

[0092] The training labels for real variant samples are determined through a set of standard sample truth values, manual verification results, long-read sequencing verification results, or family consistency verification results. When a candidate variant site is confirmed by any of the above verification methods to belong to the true genotype difference of the sample, the candidate variant site is marked as a real variant category.

[0093] The standard sample true value set refers to the reference variant set formed after confirming the variant sites in the standard genome sample; the long read sequencing verification result refers to the verification result obtained by using sequencing data that can span candidate variant sites and their adjacent low alignment regions using read lengths; the family consistency verification result refers to the result obtained by verifying the authenticity of candidate variants based on the genetic transmission relationship between parents and offspring.

[0094] The training labels for unilaterally anchored pseudovariant samples are determined jointly by regional attributes and validation results. For candidate variant sites located in homopolymer regions, low-complexity sequence regions, fragment repetition regions, pseudogene homology regions, regions with abnormal coverage due to high or low GC, regions with local collapse of the reference genome, or low-alignment regions, if they are not supported by the standard sample ground truth set and are confirmed by manual review, long-read sequencing validation, duplicate detection validation, or family consistency validation to not belong to the true genotype difference of the sample, then the candidate variant site is marked as a unilaterally anchored pseudovariant category. Candidate variant sites located only in low-alignment regions but not validated to rule out the possibility of true variants are not directly used as unilaterally anchored pseudovariant samples, but are retained as samples to be reviewed or weakly labeled samples.

[0095] Before model training, the one-sided anchoring evidence tensor in the training samples is subjected to the same uniform dimension processing, sorting processing, truncation processing, padding processing, and placeholder marking processing as the samples to be analyzed. During training, the one-sided anchoring evidence tensor is input into the two-branch deep learning model. The read evidence branch extracts the anchoring differences between the reference allele read and the alternative allele read, and the boundary context branch extracts the local uniqueness differences on both sides of the candidate variant site. After fusion, the true variant probability and the one-sided anchoring pseudo-variable probability are output.

[0096] During training, the one-sided anchored evidence tensor corresponding to the training samples is input into the two-branch deep learning model, and the model outputs the true mutation probability and the one-sided anchored pseudo mutation probability of the training samples; the cross-entropy loss is calculated based on the training labels and the model output.

[0097] ;

[0098] in, For training loss, Candidate variant sites, The categories include true variant categories and one-sided anchored pseudo-variable categories. Candidate variant sites In category The real labels on it Candidate mutation sites output by a two-branch deep learning model Category The probability, Indicates the true mutation category, This represents the single-sided anchored pseudo-variant category; by minimizing the cross-entropy loss, the trained bi-branch deep learning model is obtained.

[0099] During the model inference phase, for candidate variant sites in the sample to be analyzed, the same S1 to S3 processing method as in the training phase is used to construct a one-sided anchored evidence tensor, and the trained two-branch deep learning model is used to output the true variant probability and the one-sided anchored pseudo variant probability. The same evidence item definition, the same window construction method, the same local uniqueness calculation method, and the same tensor construction method are used in the training and inference phases to ensure that the meaning of the model input is consistent.

[0100] This embodiment does not limit the specific network type of the dual-branch deep learning model; the reading evidence branch can use convolutional neural networks, recurrent neural networks, attention networks, or Transformer networks to extract reading evidence features; the boundary context branch can use fully connected networks, convolutional networks, or attention networks to extract region context features; the fusion of the two branches can use concatenation fusion, weighted fusion, or attention fusion; as long as the model input is a one-sided anchored evidence tensor and the output is the true mutation probability and the one-sided anchored pseudo mutation probability, the technical solution of this embodiment can be achieved.

[0101] S5. Generate one-sided anchoring imbalance explanation evidence based on the true mutation probability, one-sided anchoring pseudo mutation probability and the one-sided anchoring evidence tensor, and output the genome detection data analysis results based on the true mutation probability, one-sided anchoring pseudo mutation probability and the one-sided anchoring imbalance explanation evidence;

[0102] Specifically, based on the contribution of each evidence channel in the one-sided anchored evidence tensor to the true variation probability and the one-sided anchored pseudo variation probability, one-sided anchored imbalance explanation evidence is generated.

[0103] In this embodiment, the one-sided anchoring imbalance interpretation evidence refers to interpretable evidence generated based on the evidence channels in the one-sided anchoring evidence tensor and their contribution to the probability of one-sided anchoring pseudo-variables; the one-sided anchoring imbalance interpretation evidence is used to explain the basis for determining candidate variant sites as true variants, suspected one-sided anchoring pseudo-variables, or variants to be verified, rather than detection data regenerated independently of the one-sided anchoring evidence tensor;

[0104] Furthermore, the evidence for interpreting the one-sided anchoring imbalance is generated through channel masking. Specifically, the one-sided anchoring evidence tensor corresponding to the candidate mutation site is input into a two-branch deep learning model to obtain the one-sided anchoring pseudo-mutation probability in the unmasked state. Subsequently, the candidate anchoring side matching channel, low comparable side matching channel, candidate anchoring side soft shearing channel, low comparable side soft shearing channel, insertion or deletion channel, base error probability channel, alignment quality channel, and regional context channel in the one-sided anchoring evidence tensor are masked in sequence. After masking one evidence channel each time, the masked one-sided anchoring evidence tensor is re-inputted into the two-branch deep learning model to obtain the masked one-sided anchoring pseudo-mutation probability.

[0105] If, after obscuring a certain evidence channel, the probability of unilateral anchoring pseudo-variation decreases compared to the unobscurated state, then that evidence channel is identified as contributing to the judgment of unilateral anchoring pseudo-variation. The decrease in the probability of unilateral anchoring pseudo-variation is determined by the difference between the probability of unilateral anchoring pseudo-variation in the unobscurated state and the probability of unilateral anchoring pseudo-variation in the obscurated state. When the difference is positive, it is determined that obscuring the evidence channel reduces the probability of unilateral anchoring pseudo-variation. If, after obscuring a certain evidence channel, the probability of unilateral anchoring pseudo-variation does not decrease, then that evidence channel is not used as the main basis for generating evidence for interpreting unilateral anchoring imbalance. For evidence channels identified as contributing, the corresponding segment evidence value and region context value are further read, and an interpretation text corresponding to that evidence channel is generated.

[0106] For example, when the low-comparable contralateral soft-splitting channel is masked, the probability of unilateral anchoring spurious variants decreases, and the soft-splitting length distribution of the surrogate allele read on the low-comparable contralateral is higher than that of the reference allele read on the same side, explanatory evidence is generated that "the surrogate allele read shows a soft-splitting concentration on the low-comparable contralateral." When the candidate anchoring side matching channel is masked, the probability of unilateral anchoring spurious variants decreases, and the matching length distribution of the surrogate allele read on the candidate anchoring side is higher than that on the low-comparable contralateral, explanatory evidence is generated that "the surrogate allele read concentration depends on the candidate anchoring side to form a match." When the regional context channel is masked, the probability of unilateral anchoring spurious variants decreases, and the average uniqueness on both sides of the candidate variant site is different, explanatory evidence is generated that "the candidate variant site is located at the boundary of a region with different average uniqueness on the left and right sides."

[0107] Through the aforementioned channel masking method, the evidence for the one-sided anchoring imbalance explanation is not directly inferred from the model classification results, but is determined by the influence of specific evidence channels in the one-sided anchoring evidence tensor on the probability of one-sided anchoring pseudo-variation. This enables the model judgment results to correspond with the reading anchoring patterns of candidate variant sites.

[0108] Evidence for unilateral anchoring imbalance includes at least the following:

[0109] First, evidence is used to characterize the reliance of alternative allele reads on candidate anchoring sides to form matches. This evidence comes from the match length distribution of alternative allele reads on candidate anchoring sides and the match length distribution of alternative allele reads on low comparable sides. When alternative allele reads have relatively stable matches on candidate anchoring sides, while the match length on low comparable sides is significantly shortened or unstable, it indicates that alternative allele reads may mainly rely on candidate anchoring sides to complete alignment and localization.

[0110] Second, evidence used to characterize the soft splice concentration of alternative allele reads on the low comparable side; this evidence comes from the soft splice length distribution of alternative allele reads on the low comparable side and is compared with the soft splice length distribution of reference allele reads on the same side; when alternative allele reads show soft splice concentration on the low comparable side, while the reference allele reads do not show the same distribution, it indicates that the alternative allele reads are alignment unstable on the low comparable side;

[0111] Third, it is used to characterize evidence of insertion or deletion concentrations in alternative allele reads on the low-comparable side; this evidence comes from the length distribution of insertions or deletions of alternative allele reads within the bilateral analysis window, and is judged in combination with the positional relationship of the low-comparable side; when the insertions or deletions of alternative allele reads are mainly distributed on the low-comparable side, and the reference allele reads do not show the same insertion or deletion distribution, it indicates that the candidate variant may be affected by local slippage of low-complexity sequences, homopolymer sequences, or repetitive sequences;

[0112] Fourth, evidence to characterize the directional difference in left and right matching lengths between alternative allele reads and reference allele reads; this evidence is obtained by comparing the matching lengths of alternative allele reads and reference allele reads in the left and right windows; when the alternative allele reads show sufficient matching on one side and insufficient matching on the other side, while the reference allele reads show relatively balanced matching on both sides, it indicates that there is a unilateral anchoring imbalance specific to alternative alleles at the candidate variant site;

[0113] Fifth, evidence used to characterize candidate variant sites located at the boundary of regions with different left and right average uniqueness; this evidence comes from the left and right average uniqueness obtained in S2; when the average uniqueness on both sides of the candidate variant site is different, and the direction of the abnormal anchoring of the alternative allele read is consistent with the direction between the candidate anchoring side and the low comparable pair side, it indicates that there is an association between the unilateral anchoring imbalance of the alternative allele and the low comparable pair boundary.

[0114] After generating evidence to explain one-sided anchoring imbalance, the genomic detection data analysis results are output based on the true variant probability, the one-sided anchoring pseudo variant probability, and the evidence to explain one-sided anchoring imbalance. The genomic detection data analysis results include candidate variant sites, variant types, reference alleles, substitute alleles, candidate anchoring sides, low-comparability opposite sides, true variant probability, one-sided anchoring pseudo variant probability, classification results, and evidence to explain one-sided anchoring imbalance.

[0115] The classification results include true variants, suspected one-sided anchorage pseudo-variables, and variants to be verified.

[0116] Furthermore, the probability classification result is determined by comparing the true mutation probability and the one-sided anchored pseudo mutation probability; when the true mutation probability is the only higher probability, the probability classification result is a true mutation; when the one-sided anchored pseudo mutation probability is the only higher probability, the probability classification result is a suspected one-sided anchored pseudo mutation; when there is no single higher probability between the true mutation probability and the one-sided anchored pseudo mutation probability, the probability classification result is a mutation to be verified.

[0117] The evidence classification results are determined through the interpretation of one-sided anchoring imbalance. When the evidence of one-sided anchoring imbalance simultaneously indicates that the alternative allele reads are concentrated in matching with the candidate anchoring side and show a concentrated distribution of soft splicing or insertion / deletion on the low comparable side relative to the reference allele reads, the evidence classification result is suspected one-sided anchoring spurious variant. When the evidence of one-sided anchoring imbalance indicates that the alternative allele reads are matched on both sides of the candidate variant site and do not show one-sided anchoring imbalance relative to the reference allele reads, the evidence classification result is true variant. When the evidence of one-sided anchoring imbalance is insufficient to support true variant or suspected one-sided anchoring spurious variant, the evidence classification result is variant pending verification.

[0118] When the probability classification result is consistent with the evidence classification result, output the consistent classification result; when the probability classification result is different from the evidence classification result, or when either the probability classification result or the evidence classification result is a variant to be verified, output the variant to be verified, and retain the probability result and explanatory evidence that led to the verification in the genome detection data analysis results.

[0119] When one-sided anchoring imbalance explains evidence that alternative allele reads are concentrated in a set of matches that depend on candidate anchoring sides, and show a concentrated distribution of soft splicing or insertion / deletion on low comparable sides relative to reference allele reads, the classification result is output as a suspected one-sided anchoring pseudovariable.

[0120] In this embodiment, the term "concentrated dependence on candidate anchoring side to form a match" means that, relative to the low comparable side, the substituted allele reads have a more continuous matching distribution on the candidate anchoring side, and this matching distribution contributes to the probability of unilateral anchoring pseudovariations. The term "concentrated distribution of soft splicing or insertion / deletion" means that, relative to the reference allele reads, the substituted allele reads exhibit a more concentrated soft splicing length distribution or insertion / deletion length distribution on the low comparable side, and the corresponding evidence channel, after channel masking, will affect the probability of unilateral anchoring pseudovariations. The term "unilateral anchoring imbalance" means that the substituted allele reads form a directional imbalance between the candidate anchoring side and the low comparable side, and this imbalance is more prominent than that of the reference allele reads.

[0121] When the evidence of unilateral anchoring imbalance indicates that the alternative allele reads match both sides of the candidate variant site and do not form a unilateral anchoring imbalance relative to the reference allele reads, the classification result is output as the true variant.

[0122] When there is no single higher probability between the true variant probability and the one-sided anchored pseudo variant probability, or when the probability classification result determined by the higher of the true variant probability and the one-sided anchored pseudo variant probability differs from the evidence classification result corresponding to the one-sided anchoring imbalance explanation evidence, the classification result is output as a variant to be reviewed. For variants to be reviewed, the one-sided anchoring evidence tensor summary, candidate anchoring side, low comparable side, left average uniqueness, right average uniqueness, and main explanatory evidence are retained in the genome detection data analysis results for subsequent manual review or supplementary detection.

[0123] In this embodiment, the variant to be verified refers to a candidate variant that cannot be directly classified as a true variant or a suspected spurious spurious variant based on the current one-sided anchoring evidence tensor, the probability of true variant, the probability of spurious one-sided anchoring, and the evidence of one-sided anchoring imbalance. The variant to be verified does not necessarily mean that the candidate variant is a true variant or a spurious variant, but rather that the candidate variant needs to be further confirmed by combining manual verification, supplemental sequencing, long read verification, pedigree consistency analysis, or other verification methods.

[0124] Example 2, based on Example 1, further illustrates the process of identifying suspected one-sided anchoring spurious variants at low-comparability pair boundaries using the method of the present invention;

[0125] In this embodiment, there is a candidate single nucleotide variant site in the candidate variant set of the sample to be analyzed; the reference sequence to the left of the candidate variant site consists of a relatively unique sequence fragment, and the reference sequence to the right contains a low-complexity sequence or a fragment repetitive sequence; according to the local uniqueness calculated by S2, the average uniqueness on the left is higher than that on the right, so the left window is determined as the candidate anchoring side, and the right window is determined as the low comparable pair side.

[0126] Based on S1, locally covering reads covering the candidate variant site were extracted and divided into a reference allele read set and a substitute allele read set. Subsequently, anchoring evidence was extracted based on S3. Analysis revealed that substitute allele reads had longer matching lengths in the left window and shorter matching lengths in the right window, and soft splicing and local insertion / deletion were present in the right window. Reference allele reads showed relatively stable matching in both the left and right windows, and no obvious soft splicing concentration was observed on the right.

[0127] After constructing the above evidence into a one-sided anchored evidence tensor, it is input into a two-branch deep learning model; the read evidence branch extracts the left and right matching length differences, right soft splicing differences, and local insertion / deletion differences between the alternative allele read and the reference allele read; the boundary context branch extracts the boundary context features where the average uniqueness on the left is higher than the average uniqueness on the right; after the two branches are fused, the model outputs that the one-sided anchored pseudo-variant probability of the candidate variant site is higher than the true variant probability;

[0128] The system generates evidence to explain the one-sided anchoring imbalance based on the one-sided anchoring evidence tensor. The evidence indicates that the alternative allele reads are matched based on the left candidate anchoring side, while the right low-comparability side has soft splicing and local insertion / deletion concentrations. The reference allele reads do not show the same degree of left-right anchoring imbalance. The classification result of the candidate variant site is output as a suspected one-sided anchoring pseudo-variable.

[0129] As can be seen from this embodiment, the present invention identifies pseudo-variables that have certain read depth support but may be driven by regional bias based on the one-sided anchoring imbalance pattern of the alternative allele read at the low comparability boundary relative to the reference allele read.

[0130] Example 3: Based on Example 1, this example further illustrates the application of the method of the present invention in real-world mutation identification scenarios;

[0131] In this embodiment, there is a candidate insertion mutation site in the candidate mutation set of the sample to be analyzed; according to S1, the local covering reads covering the candidate insertion mutation site are extracted from the sequencing read alignment file, and the local covering reads are divided into a reference allele read set and a substitute allele read set according to whether the reads contain an insertion fragment consistent with the alternative allele.

[0132] The local uniqueness of the reference sequences on both sides of the candidate insertion variant site is calculated according to S2. If there is a difference between the average uniqueness on the left and the average uniqueness on the right, the candidate anchoring side and the low comparable side are determined. If the average uniqueness on the left and the average uniqueness on the right are the same, a bilateral uniqueness balance marker is generated, and the average uniqueness on the left, the average uniqueness on the right, and the bilateral uniqueness balance marker are used as the regional context channel.

[0133] Based on S3, anchoring-related read evidence is extracted from the reference allele read set and the alternative allele read set. If the alternative allele read has a stable match in both the left and right windows, and there is no soft splicing concentration or insertion / deletion anomaly specific to the alternative allele read near the insertion fragment; and if the difference in left and right match lengths between the reference allele read and the alternative allele read does not show a directional imbalance, then there is no obvious alternative allele unilateral anchoring imbalance pattern in the unilateral anchoring evidence tensor.

[0134] After inputting the one-sided anchoring evidence tensor into the two-branch deep learning model, the read evidence branch did not extract the low-comparability contralateral anomaly of the alternative allele read relative to the reference allele read, and the boundary context branch also did not extract the one-sided boundary influence consistent with the alternative allele anomaly; the model output the true mutation probability was higher than the one-sided anchoring pseudo-mutation probability.

[0135] The system generates evidence to explain the one-sided anchoring imbalance. The evidence shows that the alternative allele reads match on both sides of the candidate variant site and do not form a one-sided anchoring imbalance relative to the reference allele reads. The alternative allele reads do not show soft splicing or insertion / deletion concentrations on the low-comparability contralateral side. Therefore, the classification result of the candidate insertion variant site is output as a true variant.

[0136] As can be seen from this embodiment, the present invention can not only identify suspected unilateral anchorage pseudovariables, but also output true variant results when candidate variants have bilateral stable read support, thereby avoiding the simple filtering of all complex region variants and improving the reliability of genome detection data analysis results.

[0137] Example 4: Based on Example 1, this example further illustrates the output method of the present invention for the variant to be verified;

[0138] In this embodiment, there is a candidate missing variant site in the candidate variant set of the sample to be analyzed; according to S1 to S3, the system obtains the one-sided anchored evidence tensor of the candidate missing variant site; according to S4, the two-branch deep learning model outputs the true variant probability and the one-sided anchored pseudo variant probability; if there is no unique higher one between the true variant probability and the one-sided anchored pseudo variant probability, it indicates that the model cannot form a clear probability classification based on the current one-sided anchored evidence tensor.

[0139] Alternatively, in another scenario, the probability of a true variant is higher than the probability of a spurious one-sided anchoring variant, but the one-sided anchoring imbalance explanation evidence generated by S5 indicates that the set of alternative allele reads depends on the candidate anchoring side to form a match, and there is a soft splicing or insertion / deletion set in the low-comparability side; in this case, the classification result determined by probability differs from the classification result determined by the explanation evidence; this situation indicates a conflict between the model's probability output and the mechanism explanation, and candidate variant sites should not be directly output as true variants or suspected spurious one-sided anchoring variants;

[0140] In the above cases, the system will output the classification results as variants to be reviewed, and retain the candidate variant sites, variant types, reference alleles, alternative alleles, candidate anchoring sides, low comparable counterparts, true variant probability, one-sided anchoring pseudo-variable probability, one-sided anchoring imbalance explanation evidence, and one-sided anchoring evidence tensor summary in the genome detection data analysis results.

[0141] The one-sided anchored evidence tensor abstract includes the matching length distribution, soft-splitting length distribution, insertion or deletion length distribution, base error probability distribution, alignment quality distribution, left-side average uniqueness, right-side average uniqueness, and bilateral uniqueness balance markers of the reference allele read and the substituted allele read within the left and right windows. By retaining the above information, the testing personnel can manually review the candidate variant sites, or further confirm the nature of the candidate variant sites by combining supplemental sequencing, long-read validation, pedigree consistency analysis, or other validation methods.

[0142] As can be seen from this embodiment, the present invention can output variants to be reviewed when there is a conflict between probabilistic judgment and interpretive evidence, thus avoiding the model from making overly certain judgments when there is insufficient or contradictory evidence.

[0143] Example 5: The present invention also provides a deep learning-based genome detection data analysis system for executing the deep learning-based genome detection data analysis method in Example 1. The system includes a data acquisition module, a local uniqueness analysis module, an evidence tensor construction module, a deep learning recognition module, and a result output module.

[0144] The data acquisition module is used to acquire the sequencing read alignment file, reference genome sequence and candidate variant set of the sample to be analyzed, extract local coverage reads from the candidate variant sites in the candidate variant set, and divide them into a reference allele read set and a substitute allele read set.

[0145] The local uniqueness analysis module is used to determine the candidate anchoring side and the low comparability pair side based on the local uniqueness of the reference sequences on both sides of the candidate variant site.

[0146] The evidence tensor construction module is used to extract anchor-related read evidence from the reference allele read set and the alternative allele read set according to the candidate anchoring side and the low-comparability pair side, and construct a one-sided anchoring evidence tensor in combination with the local uniqueness.

[0147] The deep learning recognition module is used to input the one-sided anchored evidence tensor into the dual-branch deep learning model and output the true mutation probability and the one-sided anchored pseudo mutation probability.

[0148] The results output module is used to generate one-sided anchoring imbalance explanation evidence corresponding to the candidate anchoring side and the low comparable pair based on the true mutation probability, the one-sided anchoring pseudo mutation probability, and the one-sided anchoring evidence tensor, and output the genome detection data analysis results based on the true mutation probability, the one-sided anchoring pseudo mutation probability, and the one-sided anchoring imbalance explanation evidence.

[0149] Example 6: Based on Examples 1 to 5, this example further illustrates that the present invention can be implemented in the form of a computer-readable storage medium or electronic device;

[0150] In one implementation, the electronic device includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, the following processes are implemented: acquiring sequencing read alignment files, a reference genome sequence, and a candidate variant set for the sample to be analyzed; extracting locally covering reads from candidate variant sites in the candidate variant set and dividing them into a reference allele read set and a substitute allele read set; determining candidate anchoring sides and low-comparability sides based on the local uniqueness of the reference sequences flanking the candidate variant sites; extracting anchoring-related read evidence from the reference allele read set and the substitute allele read set, and constructing a one-sided anchoring evidence tensor based on local uniqueness; inputting the one-sided anchoring evidence tensor into a two-branch deep learning model, and outputting the true variant probability and the one-sided anchoring pseudo-variant probability; generating one-sided anchoring imbalance explanation evidence based on the true variant probability, the one-sided anchoring pseudo-variant probability, and the one-sided anchoring imbalance explanation evidence; and outputting the genome detection data analysis results based on the true variant probability, the one-sided anchoring pseudo-variant probability, and the one-sided anchoring imbalance explanation evidence.

[0151] In another implementation, a computer-readable storage medium stores a computer program; when the computer program is executed by a processor, it implements the method steps of any one of Embodiments 1 to 5; the computer-readable storage medium includes, but is not limited to, a hard disk, a solid-state drive, a memory card, a read-only memory, a random access memory, an optical disk, or other media capable of storing a computer program.

[0152] As can be seen from the above embodiments, this invention determines candidate anchoring sides and low-comparability pairs through local uniqueness, and constructs a unilateral anchoring evidence tensor based on the left-right anchoring difference between the reference allele read and the alternative allele read. It then uses a bi-branch deep learning model to output the true mutation probability and the unilateral anchoring pseudo-mutation probability, and finally combines the unilateral anchoring imbalance interpretation evidence to output the analysis results. This method can identify suspected pseudo-mutations caused by unilateral anchoring imbalance of alternative alleles at low-comparability pair boundaries, avoiding misjudgments caused by relying solely on overall quality, read depth, or single alignment quality, thereby improving the reliability of distinguishing between true and pseudo-mutations in complex genomic regions.

[0153] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for analyzing genome detection data based on deep learning, characterized in that, Includes the following steps: S1. Obtain the sequencing read alignment file, reference genome sequence, and candidate variant set of the sample to be analyzed. Extract local coverage reads from the candidate variant sites in the candidate variant set and divide them into a reference allele read set and a substitute allele read set. S2. Based on the local uniqueness of the reference sequences on both sides of the candidate variant site, determine the candidate anchoring side and the low comparability pair side; S3. Based on the candidate anchoring side and the low comparable side, extract anchoring-related read evidence from the reference allele read set and the alternative allele read set, and construct a one-sided anchoring evidence tensor in combination with the local uniqueness. S4. Input the one-sided anchored evidence tensor into the dual-branch deep learning model and output the true mutation probability and the one-sided anchored pseudo mutation probability. S5. Generate one-sided anchoring imbalance explanation evidence based on the true mutation probability, the one-sided anchoring pseudo mutation probability, and the one-sided anchoring evidence tensor, and output the genome detection data analysis results based on the true mutation probability, the one-sided anchoring pseudo mutation probability, and the one-sided anchoring imbalance explanation evidence.

2. The method for analyzing genome detection data based on deep learning according to claim 1, characterized in that, In S1, the locally covered read is a read that covers the candidate variant site in the sequencing read alignment file; locally covered reads supporting the reference allele are assigned to the reference allele read set, and locally covered reads supporting the alternative allele are assigned to the alternative allele read set.

3. The method for analyzing genome detection data based on deep learning according to claim 1, characterized in that, S1 further includes determining bilateral analysis windows on both sides of the candidate variant site based on the median length of the reads covering the candidate variant site, the reference allele length, and the alternative allele length; the bilateral analysis windows include a left window located to the left of the candidate variant site and a right window located to the right of the candidate variant site.

4. The method for analyzing genome detection data based on deep learning according to claim 3, characterized in that, In S2, the local uniqueness is obtained in the following way: The reference sequence in the left and right windows is split into continuous subsequences whose length is determined by the effective alignment length of the locally covered read. The occurrence frequency of the continuous subsequence covering the corresponding position in the reference genome sequence is counted, and the local uniqueness of the corresponding position is obtained based on the occurrence frequency. The more occurrences, the lower the local uniqueness of the corresponding position, and the fewer occurrences, the higher the local uniqueness of the corresponding position. The average uniqueness of the left side and the average uniqueness of the right side are obtained based on the local uniqueness of each position within the left and right windows, respectively. The side with higher average uniqueness is identified as the candidate anchoring side, and the side with lower average uniqueness is identified as the low comparable side. When the average uniqueness of the left side and the average uniqueness of the right side are the same, a bilateral uniqueness balance mark is generated, and the bilateral uniqueness balance mark is used as a regional context channel to participate in the construction of the subsequent unilateral anchored evidence tensor.

5. The method for analyzing genome detection data based on deep learning according to claim 4, characterized in that, In S3, the anchoring related read evidence includes the matching length of the locally covering read in the left window, the matching length in the right window, the soft cut length in the left window, the soft cut length in the right window, the insertion or deletion length in the bilateral analysis window, the base error probability at the candidate variant site, the alignment quality in the bilateral analysis window, and the supported allele type. The base error probability is calculated from the base quality fraction of the locally covered read at the candidate variant site; the supported allele types include reference allele types and alternative allele types, the reference allele type corresponds to the reference allele read set, and the alternative allele type corresponds to the alternative allele read set.

6. The method for analyzing genome detection data based on deep learning according to claim 5, characterized in that, In S3, the one-sided anchored evidence tensor is constructed as follows: the anchored relevant read evidence in the reference allele read set is arranged into a reference allele evidence matrix; the anchored relevant read evidence in the alternative allele read set is arranged into an alternative allele evidence matrix; and the left average uniqueness, right average uniqueness, and bilateral uniqueness balance markers are used as regional context channels, which are combined with the reference allele evidence matrix and the alternative allele evidence matrix to form the one-sided anchored evidence tensor; wherein, when bilateral uniqueness balance markers are not generated, the regional context channels include left average uniqueness and right average uniqueness.

7. The method for analyzing genome detection data based on deep learning according to claim 6, characterized in that, In S4, the dual-branch deep learning model includes a read segment evidence branch and a boundary context branch; The read evidence branch receives the reference allele evidence matrix and the alternative allele evidence matrix, and extracts the differences between the reference allele read and the alternative allele read in terms of left and right matching length, left and right soft splice length, insertion or deletion length, base error probability, and alignment quality; the boundary context branch receives the region context channel and extracts the difference between the left average uniqueness and the right average uniqueness; after fusing the outputs of the read evidence branch and the boundary context branch, the true mutation probability and the one-sided anchored pseudo-mutation probability are obtained.

8. The method for analyzing genome detection data based on deep learning according to claim 7, characterized in that, In S5, the evidence for interpreting the unilateral anchoring imbalance includes: evidence characterizing the set of alternative allele reads that depend on the candidate anchoring side for matching; evidence characterizing the soft splicing set of alternative allele reads on the low comparable side; evidence characterizing the insertion or deletion set of alternative allele reads on the low comparable side; evidence characterizing the directional difference between the alternative allele reads and the reference allele reads in the left and right matching lengths; and evidence characterizing the candidate variant site located at the boundary of a region with different average uniqueness on the left and right sides.

9. The method for analyzing genome detection data based on deep learning according to claim 8, characterized in that, In S5, the genomic detection data analysis results include candidate variant sites, variant types, reference alleles, substitute alleles, candidate anchoring sides, low-comparable opposite sides, true variant probability, one-sided anchoring pseudo-variable probability, classification results, and evidence explaining one-sided anchoring imbalance. Specifically, when the evidence explaining one-sided anchoring imbalance indicates that substitute allele reads are concentrated and match the candidate anchoring side, and exhibit a concentrated distribution of soft splicing or insertion / deletion on the low-comparable opposite side relative to the reference allele reads, the classification result is output as a suspected one-sided anchoring pseudo-variable. When the evidence explaining one-sided anchoring imbalance indicates that substitute allele reads match on both sides of the candidate variant site, and do not exhibit one-sided anchoring imbalance relative to the reference allele reads, the classification result is output as a true variant. When there is no single higher probability between the true variant probability and the one-sided anchoring pseudo-variable probability, or when the probability classification result determined by the higher of the true variant probability and the one-sided anchoring pseudo-variable probability differs from the evidence classification result corresponding to the one-sided anchoring imbalance evidence, the classification result is output as a variant awaiting verification.

10. A deep learning-based genome detection data analysis system, applied to the deep learning-based genome detection data analysis method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire the sequencing read alignment file, reference genome sequence and candidate variant set of the sample to be analyzed, extract local coverage reads from the candidate variant sites in the candidate variant set, and divide them into a reference allele read set and a substitute allele read set. The local uniqueness analysis module is used to determine the candidate anchoring side and the low comparability pair side based on the local uniqueness of the reference sequences on both sides of the candidate variant site. The evidence tensor construction module is used to extract anchor-related read evidence from the reference allele read set and the alternative allele read set according to the candidate anchoring side and the low-comparability pair side, and construct a one-sided anchoring evidence tensor in combination with the local uniqueness. The deep learning recognition module is used to input the one-sided anchored evidence tensor into the dual-branch deep learning model and output the true mutation probability and the one-sided anchored pseudo mutation probability. The results output module is used to generate one-sided anchoring imbalance explanation evidence corresponding to the candidate anchoring side and the low comparable pair based on the true mutation probability, the one-sided anchoring pseudo mutation probability, and the one-sided anchoring evidence tensor, and output the genome detection data analysis results based on the true mutation probability, the one-sided anchoring pseudo mutation probability, and the one-sided anchoring imbalance explanation evidence.

Images