WO2023087277A1 - 序列变异分析方法、系统以及存储介质 - Google Patents

序列变异分析方法、系统以及存储介质 Download PDF

Info

Publication number
WO2023087277A1
WO2023087277A1 PCT/CN2021/131904 CN2021131904W WO2023087277A1 WO 2023087277 A1 WO2023087277 A1 WO 2023087277A1 CN 2021131904 W CN2021131904 W CN 2021131904W WO 2023087277 A1 WO2023087277 A1 WO 2023087277A1
Authority
WO
WIPO (PCT)
Prior art keywords
variation
data
sequence
site
prediction
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2021/131904
Other languages
English (en)
French (fr)
Inventor
熊云
刘水霞
曾权磊
张宇
韦懿
李红
方婷
柴相花
王梦杰
袁玉英
李宁
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
BGI Genomics Co Ltd
Original Assignee
BGI Genomics Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by BGI Genomics Co Ltd filed Critical BGI Genomics Co Ltd
Priority to CN202180103963.2A priority Critical patent/CN118302817A/zh
Priority to EP21964431.7A priority patent/EP4435791A4/en
Priority to AU2021474767A priority patent/AU2021474767B2/en
Priority to PCT/CN2021/131904 priority patent/WO2023087277A1/zh
Publication of WO2023087277A1 publication Critical patent/WO2023087277A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16B—BIOINFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR GENETIC OR PROTEIN-RELATED DATA PROCESSING IN COMPUTATIONAL MOLECULAR BIOLOGY
    • G16B20/00—ICT specially adapted for functional genomics or proteomics, e.g. genotype-phenotype associations
    • G16B20/20—Allele or variant detection, e.g. single nucleotide polymorphism [SNP] detection
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16B—BIOINFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR GENETIC OR PROTEIN-RELATED DATA PROCESSING IN COMPUTATIONAL MOLECULAR BIOLOGY
    • G16B20/00—ICT specially adapted for functional genomics or proteomics, e.g. genotype-phenotype associations
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16B—BIOINFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR GENETIC OR PROTEIN-RELATED DATA PROCESSING IN COMPUTATIONAL MOLECULAR BIOLOGY
    • G16B40/00—ICT specially adapted for biostatistics; ICT specially adapted for bioinformatics-related machine learning or data mining, e.g. knowledge discovery or pattern finding
    • G16B40/20—Supervised data analysis

Definitions

  • the present disclosure relates to the technical field of gene detection, in particular to a sequence variation analysis method, system and storage medium.
  • NGS Next Generation Sequencing, Next Generation Sequencing
  • Gene detection procedures generally include sample processing, gene sequencing, variant identification, variant annotation, variant interpretation, variant verification, and test reports. At present, there are mainly the following problems in gene interpretation:
  • the results after manual or automatic interpretation may have manual subjective differences or interpretation errors, so the results need to be rechecked. Due to the inability to quickly and accurately locate the variation that needs to be reviewed, the workload of the review is heavy, resulting in unnecessary labor costs.
  • the disclosure proposes a sequence variation analysis method, system, and storage medium to quickly locate variations that need to be reviewed, reduce labor costs, improve review efficiency, and improve the accuracy of interpretation of genetic testing reports.
  • the present disclosure proposes a sequence variation analysis method, which includes the following steps: obtaining sequence variation data to be analyzed; performing feature extraction on the sequence variation data to be analyzed to obtain a first variation feature set and a second variation feature set ; Input the first variation feature set into the trained first phenotype relationship prediction model to obtain the first phenotype relationship prediction result, and input the second variation feature set into the trained second phenotype relationship prediction model , to obtain a second phenotypic relationship prediction result; taking the union of the first phenotypic relationship prediction result and the second phenotypic relationship prediction result to obtain a third phenotypic relationship prediction result.
  • the present disclosure provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above-mentioned sequence variation analysis method is realized.
  • the present disclosure proposes a sequence variation analysis system, including: an acquisition module for acquiring sequence variation data to be analyzed; a first analysis module for feature extraction of the sequence variation data to be analyzed to obtain the first A variation feature set, a second variation feature set, and inputting the first variation feature set into the trained first phenotype relationship prediction model to obtain the first phenotype relationship prediction result, and inputting the second variation feature set
  • the trained second phenotype relationship prediction model obtains the second phenotype relationship prediction result, and takes the union of the first phenotype relationship prediction result and the second phenotype relationship prediction result to obtain a third phenotype Relationship Prediction Results.
  • Fig. 1 is a flow chart of the sequence variation analysis method of the first embodiment of the present disclosure
  • FIG. 2 is a schematic diagram of the structure and application of a prediction model according to an embodiment of the present disclosure
  • Fig. 3 is a flow chart of the sequence variation analysis method of the second embodiment of the present disclosure.
  • Fig. 4 is a flow chart of the sequence variation analysis method of the third embodiment of the present disclosure.
  • Fig. 5 is a flow chart of the sequence variation analysis method of the fourth embodiment of the present disclosure.
  • Fig. 6 is a schematic diagram of the results of the variation type prediction analysis of an example of the present disclosure.
  • Fig. 7 is a schematic diagram of the prediction result of phenotypic relationship in an example of the present disclosure.
  • Fig. 8 is a structural block diagram of a sequence variation analysis system according to an embodiment of the present disclosure.
  • Fig. 1 is a flow chart of the sequence variation analysis method of the first embodiment of the present disclosure.
  • the sequence variation analysis method includes the following steps:
  • the unique identifier of each variation in the sequence variation data to be analyzed is used for data search and processing.
  • the unique identity has a definite and unique mapping relationship with the variation.
  • "chromosome number-genome coordinates-reference sequence-altered sequence” (chr-pos-ref-alt, referred to as cpra) is used as the unique identity of the variation; where, chr is the chromosome number, pos is the reference sequence physical position, ref is the reference sequence, and alt is the variant sequence.
  • the sequence variation data to be analyzed includes at least one variation site information.
  • the variation site information includes at least the variation site and the unique identifier of the variation site.
  • the variation site information can be obtained through biological information analysis, and the biological information analysis can include various routine steps, such as quality control, filtering, comparison, and the like.
  • the technical means for obtaining the data for example, specific sequence information can be obtained through various sequencing technologies.
  • each first variation feature value in the first variation feature set is the result of judging the phenotypic relationship of a public variation site
  • each second variation feature value in the second variation feature set is a variation site, etc. Allele frequency data as well as functional prediction data.
  • the aforementioned phenotypic relationship refers to the relationship between the variation and the clinical phenotype, and more precisely refers to the pathogenicity of the variation.
  • the first variation characteristic value can come from the collection information of various public databases, such as HGMD (http://www.hgmd.cf.ac.uk/ac/index.php), ClinVar (https://www.ncbi.nlm .nih.gov/clinvar/), LOVD (https://www.lovd.nl/3.0/search), etc.
  • the source of the first variation characteristic value is not limited, including specific databases and the number of databases.
  • the judgment results of phenotypic relationship recorded in each database are usually based on personalized judgment rules, and the judgment results are often described in personalized fields. Taking HGMD as an example, its phenotypic relationship determination results include DM (disease-causing mutation), DM?
  • the first variation feature can be extracted from the public database by using the unique identifier (such as cpra) of each variation in the sequence variation data to be analyzed.
  • the allele frequency data of the variant site in the second variant characteristic value can come from the collection information of various public databases, such as gnomAD (https://gnomad.broadinstitute.org/), EVS (http://evs.gs. washington.edu/EVS), 1000Genomes Project (http://browser.1000genomes.org), etc.
  • the source of the allele frequency data of the variant site in the second variant characteristic value is not limited, including specific databases and the number of databases.
  • the function prediction data in the second variation characteristic value can come from the prediction results of various existing function prediction software, and one or more specific indicators can be selected as the characteristic value from the prediction results.
  • the public database includes the function prediction data
  • the corresponding function prediction data can be directly extracted from the public database.
  • the source of the function prediction data in the second variable characteristic value is not limited, including specific software and the number of software, specific indicators and the number of indicators.
  • the prediction data of the function prediction software may include at least one of, but not limited to, protein conservation prediction data, nucleic acid conservation prediction data, shear hazard prediction data, and mutation site hazard prediction data.
  • the second variation feature can be extracted from the public database by using the unique identifier (such as cpra) of each variation in the sequence variation data to be analyzed.
  • step S4 it may also include after step S4:
  • the variation interpretation result refers to the determination result of the phenotypic relationship of each variation site in the sequence variation data to be analyzed, which is obtained by adopting a different method from the third phenotypic relationship prediction result.
  • Human interpretation results are a common type of variant interpretation results described.
  • conventional means in this field can be adopted to add custom marks to the variant interpretation results judged as credible and unreliable, respectively, for subsequent data processing.
  • This solution can quickly judge whether a large number of variation interpretation results are credible or unreliable, and deal with them by type in subsequent links. For example, a review may be performed on the variant sites that are judged to be unreliable by the variant interpretation results. It is usually confirmed by manual review. This solution can greatly reduce the workload of subsequent links, reduce labor costs, and improve interpretation efficiency.
  • the technical solution including steps S1 to S4 uses two relatively independent models (ie, the first phenotype relationship prediction model and the second phenotype relationship prediction model) according to different data types. model) to predict, and the two models are complementary.
  • the first phenotype relationship prediction model can make up for the false negatives that may be caused by the second phenotype relationship prediction model without functional data of variant sites, co-segregation data, and new data
  • the second phenotype relationship prediction model can make up for the first phenotype False negatives in the relationship prediction model caused by the judgment results of the phenotypic relationship of the variant site that is not disclosed.
  • Both the first phenotype relationship prediction model and the second phenotype relationship prediction model can be constructed based on machine learning models, and the construction method is as follows:
  • constructing a data set including a first data set and a second data set
  • the first data set is the first variation characteristic of the selected variation site and the credible phenotypic relationship determination result of the variation site
  • the second data set Determine the result of the second variation feature of the selected variation site and the credible phenotype relationship of the variation site.
  • the selected variation site refers to a plurality of variation sites including all phenotypic relationship determination result types.
  • the credible phenotypic relationship determination result may be the result of manual phenotypic relationship determination according to industry guidelines or consensus; the credible phenotypic relationship determination result may also be a plurality of public variants The results of point phenotype relationship judgment results were consistent.
  • the first machine learning model and the second machine learning model can be selected from at least one of logistic regression, naive Bayesian, support vector machine and artificial neural network.
  • the first phenotype relationship prediction model is a logistic regression model
  • the second phenotype relationship prediction model is a neural network model
  • the first phenotypic relationship prediction model is used to rapidly determine the pathogenicity of the variation recorded in the public database with a unified standard.
  • the construction process of the first phenotype relationship prediction model includes:
  • the first data set includes the pathogenicity determination results recorded in the public variation database and the pathogenicity determination results made by interpretation experts.
  • the public variation database may include at least one of ClinVar, HGMD, LOVD, UMD (http://umd-predictor.eu/) and other databases.
  • the variation annotation tools can be used to annotate the variation data collected from each public variation database to obtain the cpra information, and use this information as the unique identification of the variation.
  • the variation data of each public variation database is summarized.
  • the interpretation experts combined the population frequency database, prediction software, and medical literature to conduct a pathogenicity analysis of the same variation. Judgment, the judgment result is the credible phenotype relationship judgment result, and this result is used as the standard result of the training data.
  • the disclosure converts the character data in the first data set into numerical data, as shown in Table 2.
  • the first preset ratio may be 8:2.
  • the first machine learning model is at least one of machine learning models such as logistic regression, naive Bayesian, support vector machine and artificial neural network.
  • the function to be realized by the first phenotype relationship prediction model is to output a unique mutation pathogenicity judgment based on the mutation pathogenicity judgments of multiple different public mutation databases, and the pathogenicity judgment of mutations is a feature that is relatively complex.
  • machine learning models such as logistic regression, naive Bayesian, support vector machine and artificial neural network can be selected to realize this function.
  • a plurality of different machine learning models can be selected, and the selected machine learning models are independently trained through the first training set, and the performance of each machine learning model trained is evaluated through the first test set , the evaluation indicators can include accuracy rate, precision rate, recall rate, etc.
  • the performance evaluation results of each model a model with the best performance is selected as the first phenotype relationship prediction model used in this disclosure. For example, the model with the highest accuracy rate obtained from training can be used as the first phenotype relationship prediction model.
  • the first variation characteristic value of each variation can be extracted; input the first variation characteristic value into the first phenotypic relationship prediction model, and the first phenotypic relationship prediction can be obtained result.
  • the first phenotypic relationship prediction model outputs prediction results for the variation included in the public database and according to the preset phenotypic relationship.
  • the pathogenicity of human Mendelian genetic disease-related variants there are usually three types of preset phenotypic relationships (pathogenic or likely pathogenic, benign or likely benign, and unclear significance), five types ( pathogenic, possibly pathogenic, benign, possibly benign, and of unclear significance), without limitation.
  • the construction process of the second phenotypic relationship prediction model may include:
  • the second data set includes allele frequency data of variant sites, function prediction data, and credible phenotypic relationship determination results.
  • the allele frequency data can be derived from gnomAD, Thousands Database, ExAC database, etc.; the functional prediction data can be derived from the data of various prediction software, such as SIFT, Polyphen2, MutationTaster, GERP++, DANN, etc. It can be seen that, different from the first phenotypic relationship prediction model, the second phenotypic relationship prediction model is obtained based on the original data for judging the phenotypic relationship, while the first phenotypic relationship prediction model is based on the public phenotypic relationship The judgment result is obtained.
  • the variation sites in the second data set need to be consistent in the judgment results of the phenotypic relationship of multiple public variation locations.
  • the consistent judgment result is the credible phenotypic relationship judgment result, and this result is used as the standard result of the training data .
  • the second preset ratio may be 8:2.
  • the second machine learning model may choose machine learning models such as logistic regression, naive Bayesian, support vector machine and artificial neural network.
  • the optimal second machine learning model is used as the second phenotype relationship prediction model.
  • the second variation characteristic value of each variation can be extracted; input the second variation characteristic value into the second phenotypic relationship prediction model, and the second phenotypic relationship prediction can be obtained result.
  • the variation interpretation result can be obtained by manual interpretation based on the sequence variation data to be analyzed.
  • the variation interpretation result can be compared with the corresponding prediction result of the third phenotype relationship, and if the two results are consistent, it is determined that the variation interpretation result is credible, otherwise it is determined not to be credible. Unreliable variant interpretation results require further manual review.
  • the prediction result is the result of judging the phenotypic relationship of each variation site in the sequence variation data to be analyzed obtained by any different method, such as the result of manual interpretation by those skilled in the art.
  • sequence variation analysis method may further include:
  • the simple repeat region refers to a repeat region composed of 1-5 base repeat units, such as AAA, CAACAACAACAA.
  • a weak association with a phenotypic change can mean that the variant is located in a non-coding region or an intronic region, and the allele frequency of the variant is greater than 0.05.
  • the variant located in the non-coding region or the intron region can refer to the functional annotation of the variant as - (indicating no functional annotation), intron (indicating intron), 3′-UTR (indicating the untranslated region at the 3′ end), 5′ -UTR (represents the 5' untranslated region), nochange (no change compared to the reference mRNA sequence).
  • the first variable site also meets any of the following conditions:
  • Condition 1 the total number of detections of the first variant site is greater than the first preset value, and the number of low-quality detections is greater than the second preset value;
  • Condition 2 Located within the third preset value upstream and downstream of the physical position of the reference sequence of the variant site meeting condition 1.
  • the first variable site data set containing the first variable site can be pre-established, and its construction steps are as follows:
  • this step is the same as the aforementioned step S1. More specifically, sample sequencing data on a fixed sequencing chip (such as LCY171 chip) on a gene sequencer (such as MGISEQ-2000 sequencer) can be collected, and then the variation site and the corresponding unique identification of the variation can be obtained from the sample sequencing data.
  • a fixed sequencing chip such as LCY171 chip
  • a gene sequencer such as MGISEQ-2000 sequencer
  • the first variation site data set is continuously updated along with the accumulation of samples, so as to realize the filtering of sequence variation data to be analyzed.
  • variable site B1 judging whether each of the obtained variable sites is located in the simple repeat region of the reference sequence, and if so, recording the unique identifier of the variation corresponding to the variable site.
  • variable sites that meet the preset filtering conditions can be: functional annotation as - (indicating no functional annotation), intron (indicating intron), 3'-UTR (indicating the untranslated region at the 3' end), 5' - UTR (representing the untranslated region at the 5′ end), nochange, and a variation with a frequency (representing the probability of the variation in the population) greater than 0.05.
  • filtering can be performed based on the mutation function and mutation frequency, and the filtering conditions are: the function annotation is -, intron, 3′-UTR, 5′-UTR, nochange, and the frequency is greater than 0.05.
  • the setting of this filter condition is to select the variants that have a high probability of not causing phenotype changes, that is, high-frequency variants located in non-exon regions.
  • D1 count the number of low-quality detections and the total number of detections of each variant site in the results of C1.
  • the number of low quality detections, the number of high quality detections and the total number of detections are counted using the unique identifier of the variation as an index.
  • the total number of detections is the sum of the number of low-quality detections and the number of high-quality detections.
  • the statistical rules are: 1) The total number of detections ⁇ the first preset value (such as 8), the number of low-quality detections ⁇ the second preset value (such as 1), and non-single-base variation mutation sites; 2) Find the non-single-base variation variation site within the third preset value (such as 3 bp) upstream and downstream of the reference sequence physical position of the variation site in 1), which is still located in the simple repeat region. It should be noted that in the simple repeat region, the position of the variant site compared with the reference sequence will fluctuate. Different databases or analysis software have different recording methods, but the essence is still the same variant.
  • the reference sequence ATATAT, the variant site sequence AT, and the variant site sequence can be matched with the first, third and fifth positions of the reference sequence respectively. Therefore, the statistical rule 2) is set to ensure the comprehensive inclusion of the first variant site.
  • the first variable site data set is composed of the variable sites satisfying any one of the above statistical rules 1) and 2).
  • the variation unique identifiers of each variation sites in the sequence variation data to be analyzed can be firstly obtained, and filtering is performed based on the variation unique identifiers.
  • the analysis process may be: calling the first variation site data set, and traversing the first variation site data set based on the unique variation identifier of the sequence variation data to be analyzed. If there is a consistent unique identifier for the variation, it means that the variation site in the sequence variation data to be analyzed is the first variation site and needs to be filtered; otherwise, it needs to be retained.
  • the data used for phenotypic relationship prediction may be the data after the filtering operation in step S5, that is, steps S2-S4 may be performed after step S5, and the data filtered in step S5 may be used. Therefore, by reducing the data for prediction of phenotypic relationship, unnecessary prediction can be avoided and prediction efficiency can be improved.
  • sequence variation analysis method may further include:
  • the complex variation may include insertion and substitution composite events, deletion and deletion composite events, deletion and substitution composite events, and there is no limit to the number of each specific variation of each type of complex variation.
  • conventional means in the art can be adopted to add custom markers to sites judged to have complex variations and/or not exist complex variations for subsequent data processing.
  • variable site with complex variation may satisfy the following conditions:
  • At least two reference sequence coordinates of variant sites leading to amino acid changes are adjacent and affect the same encoded amino acid.
  • a variation detection tool such as GATK software
  • a substitution event or an insertion event is simply detected, and a compound event including insertion and substitution cannot be detected.
  • the present disclosure performs complex variation judgment on the variation annotations detected by the variation detection tool, specifically to determine whether the variation at adjacent positions of the same gene coding region affects the same amino acid residue coding. If there is a complex variation, it indicates that a review is required, such as by manual combined interpretation; and vice versa.
  • the HGVS naming rules are usually used to name the annotation results.
  • the variation with amino acid changes it is judged that there is a complex variation when any of the following conditions are met: 1) There are at least two variations of the same gene with overlapping reference sequence coordinates; 2) There are at least two variations of the same gene with reference sequence coordinates Adjacent and affect the same amino acid code.
  • condition 1) "coordinate” refers to the position of the reference sequence corresponding to the variation, and "overlap” refers to the reference sequence coordinate overlap of ⁇ 1 bp; this condition can screen complex variations as comprehensively as possible.
  • condition 2 "adjacent" refers to the range of 4 bp upstream and downstream of the reference sequence; this condition can accurately screen complex variations. Therefore, through the judgment of complex variation, the defects of existing variation detection tools can be made up for, and complex variation can be combined and interpreted, which can improve the efficiency of variation interpretation.
  • the sequence variation data to be analyzed in step S6 may be the sequence variation data to be analyzed in step S1, or it may be the sequence variation data to be analyzed after the third phenotypic relationship prediction result is compared with the corresponding variation interpretation result, with a judgment result The sequence variation data to be analyzed.
  • sequence variation analysis method may further include:
  • the third variation feature value in the third variation feature set may include variation support data (when the sequence variation data to be analyzed is obtained by sequencing means, the variation support data may include quality value, sequencing depth, reads ratio supporting variation) , allele frequency (eg ESP6500_MAF, G1000_AF).
  • variation support data when the sequence variation data to be analyzed is obtained by sequencing means, the variation support data may include quality value, sequencing depth, reads ratio supporting variation
  • allele frequency eg ESP6500_MAF, G1000_AF.
  • variation type may include homozygous variation, heterozygous variation, and no variation.
  • No variation can be the homozygous genotype of the reference sequence at this locus.
  • the method for determining the prediction result of the variation type is: compare the prediction probabilities of each variation type, and determine the variation type with the highest prediction probability; compare the maximum prediction probability with a preset threshold; if the prediction probability is greater than the preset threshold , then it is determined that the variation type corresponding to the maximum prediction probability is the variation type prediction result of the variation site.
  • the process of building a variant-type predictive model could include:
  • A4 construct the third data set, the third data set includes quality value, sequencing depth, reads ratio supporting variation, allele frequency and credible variation type determination results.
  • the credible determination result of the variation type may be the result obtained by using the consensus gold standard means in the industry, for example, the Sanger verification result.
  • the third preset ratio may be 8:2.
  • a variation type prediction model is constructed to calculate the prior probability and conditional probability required by the Bayesian classifier.
  • Model testing is performed on the variation type prediction model through the data in the third test set, the classification of the test set data is judged by the variation type prediction model, and the accuracy of the variation type prediction model is obtained according to the correct number of judgments.
  • the mutation sites verified by Sanger are selected as training samples, and their relevant data are collected to form the third training set.
  • These training samples are divided into K mutation types c 1 , c 2 , ..., c K ; the preprocessing obtains
  • the variation data of the variation information is obtained, which is denoted as xi ;
  • the prior probability of each characteristic value is calculated based on the selected training set, and the prior probability can be determined by the number of occurrences of various samples in the third training set Estimated, taking the eigenvalue x1 when the variation type is c1 as an example, its prior probability is:
  • P(x) is the evidence factor used for normalization, which is a fixed value for all categories.
  • the performance of the above machine learning model is evaluated through the third test set, and the evaluation indicators may include accuracy rate, precision rate, and recall rate.
  • the variation type prediction model constructed for the variation to be tested is predicted to obtain the probability P(c k
  • a quality control threshold is set, and if and only when P max (c k
  • the sequence variation data to be analyzed in step S8 may be the sequence variation data to be analyzed in step S1, or it may be the sequence variation data to be analyzed after the third phenotypic relationship prediction result is compared with the corresponding variation interpretation result, with a judgment result
  • the sequence variation data to be analyzed can also be the sequence variation data to be analyzed with complex variation results.
  • step S10 it may also include: when the variation type prediction result only includes the reference sequence genotype, then it is determined that the variation site is not credible; when the variation type prediction result contains at least one non-reference sequence genotype, it is determined that the variant site is credible.
  • conventional means in the art can be adopted to add custom marks to the variant sites judged to be unreliable and/or credible, so as to facilitate subsequent data processing. This solution can quickly determine whether a large number of mutation sites are credible or unreliable, and deal with them by type in subsequent links. For example, only the variant sites judged to be unreliable are verified.
  • variant sites associated with phenotypic changes that are judged to be unreliable are verified, where the variant sites associated with phenotypic changes may be pathogenic variants.
  • the gold standard methods recognized in the field are usually used for verification, such as generation sequencing verification. This solution can greatly reduce the workload of subsequent verification links, reduce labor costs, and improve interpretation efficiency.
  • the sequence variation analysis method of the present disclosure can quickly determine whether the variation needs to be rechecked, whether experimental verification is required, and the like. For the entire interpretation process, the cost of variation experiment verification is reduced to a certain extent, and the results of manual interpretation or automatic interpretation are analyzed, and the variation that needs to be reviewed is quickly output, which reduces the cost of manual review and ensures the accuracy of the results.
  • the filter conditions that can be set in the embodiment are: 1 the variation whose function annotation is "-”, “intron”, “3-UTR”, “5-UTR” and “nochange”; 2 the variation whose frequency is greater than 0.05, which can be specified by G1000, ESP6500, genomAD and other population databases were used for judgment.
  • the allele frequency of the variation recorded in the three databases G1000, ESP6500, and genomAD is greater than 0.05, so this variation is filtered.
  • the total number of detections is the sum of the number of low-quality detections and the number of high-quality detections.
  • variable sites that meet one of the following two conditions to construct the first variable site data set.
  • first variation site data set at least the cpra information of the included variation sites must be recorded.
  • complex variation determination can also process received data (for example, input data in VCF format, FASTQ format, etc.), and use bioinformatics software (such as GATK, ANNOVAR, Alamut, etc.) to perform annotation at the amino acid level, Determines whether the variation contains the result of the pHGVS point. For example, for the variation chr7-142458526-A-G, the annotation result at the amino acid level is p.Asn54Ser, which contains the pHGVS point.
  • bioinformatics software such as GATK, ANNOVAR, Alamut, etc.
  • the detected variation contains both chr7-142458526-A-G (c.161A>G, p.Asn54Ser) and chr7-142458527-C-G (c.162C>G, p.Asn54Lys), the former is cDNA position 161 and the latter is 162 bit, adjacent to each other.
  • chr7-142458526-A-G chr7-142458527-C-G
  • SNPs Single Nucleotide Polymorphism, single nucleotide polymorphism
  • variation annotation tool uses the variation annotation tool to annotate the training data to obtain its cpra information (genome version is hg19), and use this information as the unique identification of the variation.
  • variant cpra gene version is hg19
  • the variants and their pathogenicity determinations in the public variant database are sorted and summarized into a training summary database.
  • the interpretation experts will manually determine the pathogenicity of the mutations in the summary database in combination with the population frequency database, prediction software and medical literature, and use the determination results as the standard results of the training data.
  • Table 3 The structure and content of the preprocessed training summary database are shown in Table 3.
  • the character data in the summary database can be converted into numerical data according to the correspondence shown in Table 4 to Table 6.
  • Table 6 Complete the construction of the first training set.
  • the structure and content of the first training set after feature extraction are shown in Table 8.
  • HGMD pathogenicity determination (character type) HGMD pathogenicity determination (numerical type) Null (the database does not have the mutation record) 0 R 1 FP 2 DP 3 DFP 4 DM? 5 DM 6
  • the first training set is randomly split into a first training set and a first testing set.
  • the first training set is used for training by logistic regression model.
  • the logistic regression model is trained through the first training set, and its accuracy rate is 91.0% evaluated by the first test set, and the test performance data are shown in Table 8.
  • Collect variants to be predicted such as NM_000267.3(NF1):c.1722-2A>G, NM_000057.4(BLM):c.893C>T(p.Thr298Met) and NM_000244.3(MEN1):c.670- 6C>T, and its pathogenicity determination in the public variation database ClinVar and HGMD are used as the data to be predicted.
  • the variation to be predicted can be obtained from the variation site filtered by the above-mentioned first variation location.
  • the variant annotation tool uses the variant annotation tool to annotate the variant to be predicted, obtain its cpra information (genome version is hg19), and use this information as the unique identifier of the variant.
  • the variant cpra as the index, the pathogenicity determination of the variants to be predicted in the public variant databases ClinVar and HGMD is sorted and summarized into one database. Because the pathogenicity determination results of the public variation database are sequenced discrete data, the character data in the prediction summary database is converted into numerical data according to the conversion logic shown in Table 8, and the prediction data processing is completed. Table 9 shows the structure and content of the prediction summary database after preprocessing and feature extraction.
  • the variation data can be obtained from the ClinVar database.
  • the ClinVar database contains 46,585 clear pathogenic loci, screened in the HGMD database to be considered "DM", and the mutation in the dbNSFP (https://sites.google.com/site/jpopgen/dbNSFP) database There are a total of 21105 loci, and half of them, 10552 in total, are selected as the pathogenic variants in the second training set, and the other 10553 pathogenic loci constitute the first test subset of the second test set.
  • the ClinVar database contains 23892 benign loci, and a total of 4664 loci in the dbNSFP database are selected as benign variants in the second training set.
  • the processing of missing values in the dbNSFP database is to directly assign a value of 0.
  • the model is based on the allele frequency database 1000 Genomes Project, ESP6500, and multiple function prediction software data.
  • function prediction software can be divided into protein conservation prediction such as SIFT, Polyphen2, MutationTaster, nucleic acid conservation prediction such as GERP++, shear hazard prediction such as dbscSNV, and mutation site hazard prediction For example DANN.
  • score_rankscore FATHMM_converted_rankscore, PROVEAN_converted_rankscore, MetaSVM_rankscore, MetaLR_rankscore, REVEL_score, CADD_raw_rankscore, DANN_rankscore, GERP++_RS_rankscore, splicing_consensus_ada_score, splicing_consensus_rf_score, phyloP10 0way_vertebrate_rankscore, phyloP20way_mammalian_rankscore, phastCons100way_vertebrate_rankscore, phastCons20way_mammalian_rankscore, SiPhy_29way_logOdds_rankscore, 1000Gp3_AF, ESP6500_AA_AF.
  • a neural network model may be selected as the second machine learning model.
  • the constructed neural network model can contain three hidden layers, the number of nodes is 16, 128, and 16 respectively; the weight initialization function is uniform, the activation function is hard_sigmoid, the optimizer uses the default parameters of the Adadelta method, and the neuron inactivation probability dropout_rate is 0.05, the number of training iterations is 800, and the batch_size of samples selected for one-time training is 64.
  • Test subset 1 is the above-mentioned first test subset
  • test subset 2 can be the result of mutations interpreted by ACMG based on clinical interpretation and accumulation, and the screening is clearly considered to be Pathogenic (causal variant) and Benign (benign variant) Among them, for the site that has been read multiple times and the interpretation results are inconsistent, select the result with the most interpretation results.
  • dbNSFP database After screening the variants in the dbNSFP database and removing the variants in the training set, a total of 728 variants were obtained, including 618 pathogenic variants and 108 benign variants.
  • the performance of the trained second machine learning model was evaluated through the above two test subsets.
  • the results of ROC (Receiver Operating Characteristic, Receiver Operating Characteristic Curve) are shown in Figure 6.
  • the AUC can reach 0.99.
  • the prediction sub-model of the present disclosure can increase AUC by 4-5 percentage points compared with previous methods, and can assist genetic analysis for rapid pathogenicity review, and can be used for mutation pathogenicity quality control, which can achieve higher accuracy.
  • 3.2.b1 Collect the eigenvalues of the variation to be predicted, that is, collect the eigenvalues x n used in the training data for the variation to be predicted, specifically the same multiple eigenvalues as described in 3.2.1.2.
  • the variation to be predicted may be the variation site obtained from the filtering of the above-mentioned first variation location.
  • 4.1.1 Build the training set. Select 843 mutation sites from the historical sequencing data of the mutation results verified by Sanger, and randomly divide them into the third training set and the third test set according to the ratio of 4:1.
  • the third training set is used to train the model.
  • the third test set is used to test the model.
  • the specific data distribution of the training set with 674 variant sites is: 269 non-variant sites, 212 homozygous variant sites, and 193 heterozygous variant sites.
  • the prior probability P(c k ) and class conditional probability are calculated based on the third training set of 674 mutation sites.
  • the performance of the machine learning model obtained through the above training is evaluated through the third test set, and the evaluation indicators may include accuracy rate, precision rate, and recall rate.
  • the 169 mutation data selected above are used as the third test set.
  • the variation type that is greater than the preset threshold and corresponding to the maximum prediction probability is selected as the variation type prediction result of the variation site, and the preset threshold is set to 0.8.
  • the specific test results are shown in Table 11. It should be noted that the preset threshold of the variation type prediction model can be adjusted according to the actual situation.
  • the constructed variation type prediction model predicts the probability P(c k
  • the first phenotype prediction model and the second phenotype prediction model are judged as credible variant sites that do not require manual review.
  • the method for judging the prediction result of the variation type is specifically to select the variation type corresponding to the maximum prediction probability greater than the preset threshold of 0.8 as the prediction result of the variation type of the variation site.
  • the site where the variation interpretation result is a pathogenic variant as an example, when the predicted result of the variant type is homozygous or heterozygous, it is determined that the variant site is credible and no experimental verification is required; when the predicted result of the variant type is no If the mutation type is selected, it is determined that the mutation site is not credible and needs to be verified experimentally.
  • sequence variation analysis method of the embodiment of the present disclosure can quickly, accurately and comprehensively determine the variation that needs to be filtered, checked and verified. It only needs to review and maintain the interpretation results judged as unreliable, and verify the pathogenic variants judged as unreliable. As a result, labor costs are greatly reduced, the efficiency of review and verification of genetic variation sites is improved, the time for issuing genetic test reports is shortened, the accuracy of interpretation of genetic test reports is also improved, and the entire interpretation and analysis process is optimized.
  • the present disclosure proposes a computer-readable storage medium.
  • a computer program is stored on a computer-readable storage medium, and when the computer program is executed by a processor, the above-mentioned sequence variation analysis method is realized.
  • Fig. 8 is a structural block diagram of one of the specific implementations of the sequence variation analysis system of the present disclosure.
  • the sequence variation analysis system 100 includes: an acquisition module 110 and a first analysis module 120 .
  • the acquisition module 110 is used to obtain the sequence variation data to be analyzed;
  • the first analysis module 120 is used to perform feature extraction on the sequence variation data to be analyzed, obtain the first variation feature set, the second variation feature set, and convert the first variation feature set
  • Input the trained first phenotype relationship prediction model to obtain the first phenotype relationship prediction result, input the second variation feature set into the trained second phenotype relationship prediction model, obtain the second phenotype relationship prediction result, and take A union of the first phenotype relationship prediction result and the second phenotype relationship prediction result is obtained to obtain a third phenotype relationship prediction result.
  • the sequence variation analysis system 100 may further include: a second analysis module 130 .
  • the second analysis module 130 is used to filter the first variation site in the sequence variation data to be analyzed, the first variation site is located in the repeat region of the reference sequence, is weakly correlated with phenotypic changes, and is not a single base variation.
  • the sequence variation analysis system 100 may further include: a third analysis module 140 .
  • the third analysis module 140 is used to obtain the amino acid level annotation results of the sequence variation data to be analyzed, and judge whether complex variations exist in the sequence variation data to be analyzed according to the amino acid level annotation results.
  • the sequence variation analysis system 100 may further include: a fourth analysis module 150 .
  • the fourth analysis module 150 is used to perform feature extraction on the sequence variation data to be analyzed to obtain the third variation feature set, and input the third variation feature set into the trained variation type prediction model to obtain each variation in the sequence variation data to be analyzed.
  • the prediction probability of each variation type to which the site belongs, and the prediction result of the variation type of the corresponding variation site is determined according to the prediction probability.
  • sequence variation analysis system of the embodiment of the present disclosure can quickly, accurately and comprehensively determine the variation that needs to be filtered, checked and verified. Only need to review and maintain the unreliable interpretation results, and verify the unreliable pathogenic variants. As a result, labor costs are greatly reduced, the efficiency of review and verification of genetic variation sites is improved, the time for issuing genetic test reports is shortened, the accuracy of interpretation of genetic test reports is also improved, and the entire interpretation and analysis process is optimized.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Medical Informatics (AREA)
  • General Health & Medical Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Biophysics (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Biotechnology (AREA)
  • Proteomics, Peptides & Aminoacids (AREA)
  • Analytical Chemistry (AREA)
  • Chemical & Material Sciences (AREA)
  • Molecular Biology (AREA)
  • Genetics & Genomics (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioethics (AREA)
  • Artificial Intelligence (AREA)
  • Databases & Information Systems (AREA)
  • Epidemiology (AREA)
  • Evolutionary Computation (AREA)
  • Public Health (AREA)
  • Software Systems (AREA)
  • Measuring Or Testing Involving Enzymes Or Micro-Organisms (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种序列变异分析方法、系统以及存储介质，涉及基因检测技术领域。序列变异分析方法包括以下步骤：获取待分析序列变异数据；对待分析序列变异数据进行特征提取，得到第一变异特征集、第二变异特征集；将第一变异特征集输入训练好的第一表型关系预测模型，得到第一表型关系预测结果，并将第二变异特征集输入训练好的第二表型关系预测模型，得到第二表型关系预测结果；取第一表型关系预测结果和第二表型关系预测结果的并集，获得第三表型关系预测结果。

Description

序列变异分析方法、系统以及存储介质 技术领域
本公开涉及基因检测技术领域，尤其涉及一种序列变异分析方法、系统以及存储介质。
背景技术
基因缺陷是导致早期流产、死胎、围产儿死亡、婴幼儿死亡和先天残疾的主要原因。随着二代测序技术的应用和推广，人们对出生缺陷防控的重视程度越来越强，一次性对多种疾病进行快速、准确、低价的筛查或诊断成为可能。然而，如何客观、准确、高效地解读NGS(Next Generation Sequencing，下一代测序)测序结果以指导临床应用一直是需要优化的重要问题。
基因检测程序一般包含样本处理、基因测序、变异识别、变异注释、变异解读、变异验证、检测报告等。目前，基因解读主要存在以下问题：
1)基因检测数据需要依赖专业人员手工查询各种数据库和文献，以解读基因变异的致病性，这种解读往往耗费大量人力和时间；并且，基因解读对从业人员的要求高，需要从业人员有基因、变异、疾病等方面的知识储备，且由于解读是基于从业人员的经验逻辑，故因人而异解读可能会有不同的结果。
2)大多变异注释工具无法识别复杂变异，导致注释信息错误，继而可能会引起解读错漏。例如，对于变异的临床意义解读，通常采用常见变异类型(如SNV/InDel)的排序技术和致病性机器学习模型，该技术主要是利用公开数据库/软件数据信息，进行机器学习训练，或是结合一定的人工经验逻辑，从而得到变异临床意义排序和分类算法。然而，由于信息来源不够全面，且局限于训练量，该技术的准确性无法保证，对于部分变异的解读结果可能会有错漏。
3)检测获得的数据，经人工或自动化解读后的结果可能存在人工主观差异或解读错漏，故需对结果进行复核。由于不能快速准确的定位需重点复核的变异，导致复核工作量大，造成不必要的人工成本。
4)由于NGS测序技术均一性问题，部分基因变异检测结果需要进一步进行实验验证，由于变异位点缺乏有效的质量分类标准，故对检测结果的验证主要依赖于人工主观判断，对于解读致病的变异大都需要变异验证以确保解读结果的准确性。常见的变异验证是针对变异位点所在区段人工进行引物设计、PCR(Polymerase Chain Reaction，聚合酶链式反应)扩增、一代测序、结果分析，会耗费大量人力、时间和检测成本。
发明内容
本公开提出了一种序列变异分析方法、系统以及存储介质，以实现快速定位需复核的变异，降低人工成本，提升复核效率，提高基因检测报告解读的准确性。
第一方面，本公开提出了一种序列变异分析方法，包括以下步骤：获取待分析序列变异数据；对所述待分析序列变异数据进行特征提取，得到第一变异特征集、第二变异特征集；将所述第一变异特征集输入训练好的第一表型关系预测模型，得到第一表型关系预测结果，并将所述第二变异特征集输入训练好的第二表型关系预测模型，得到第二表型关系预测结果；取所述第一表型关系预测结果和所述第二表型关系预测结果的并集，获得第三表型关系预测结果。
第二方面，本公开提出了一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时，实现上述的序列变异分析方法。
第三方面，本公开提出了一种序列变异分析系统，包括：获取模块，用于获取待分析序列变异数据；第一分析模块，用于对所述待分析序列变异数据进行特征提取，得到第一变异特征集、第二变异特征集，并将所述第一变异特征集输入训练好的第一表型关系预测模型，得到第一表型关系预测结果，将所述第二变异特征集输入训练好的第二表型关系预测模型，得到第二表型关系预测结果，以及取所述第一表型关系预测结果和所述第二表型关系预测结果的并集，得到第三表型关系预测结果。
附图说明
通过阅读参照以下附图所作的对非限制性实施例所作的详细描述，本公开的其它特征、目的和优点将会变得更明显：
图1是本公开第一个实施例的序列变异分析方法的流程图；
图2是本公开一个实施例的预测模型的结构和应用示意图；
图3是本公开第二个实施例的序列变异分析方法的流程图；
图4是本公开第三个实施例的序列变异分析方法的流程图；
图5是本公开第四个实施例的序列变异分析方法的流程图；
图6是本公开一个示例的变异类型预测分析结果的示意图；
图7是本公开一个示例的表型关系预测结果的示意图；
图8是本公开实施例的序列变异分析系统的结构框图。
具体实施方式
为了使本公开的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本 公开进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本公开，并不用于限定本公开。为了说明本公开的技术方案，下面通过具体实施例来进行说明。
图1是本公开第一个实施例的序列变异分析方法的流程图。
如图1所示，序列变异分析方法包括以下步骤：
S1，获取待分析序列变异数据。
在本公开的实施例中，为了便于分析，使用待分析序列变异数据中的各变异唯一身份标识进行数据查找和处理。该唯一身份标识与变异具有确定的且唯一的映射关系。如本领域惯常使用的“染色体号-基因组坐标-参考序列-改变序列”(chr-pos-ref-alt，简称cpra)作为变异的唯一身份标识；其中，chr为染色体号，pos为参考序列物理位置，ref为参考序列，alt为变异序列。
所述待分析序列变异数据包括至少一个变异位点信息。变异位点信息至少包括变异位点和变异位点的唯一身份标识。具体地，该变异位点信息可通过生物信息分析获得，生物信息分析可包括多种常规步骤，如质控、过滤、比对等。对获取该数据的技术手段不做限制，例如具体的序列信息可通过各种测序技术获得。
S2，对待分析序列变异数据进行特征提取，得到第一变异特征集、第二变异特征集。
在一优选的实施方式中，第一变异特征集中的各第一变异特征值为公开的变异位点的表型关系判定结果，第二变异特征集中的各第二变异特征值为变异位点等位基因频率数据以及功能预测数据。
具体地，当本方案应用于人类孟德尔遗传病相关变异分析时，前述表型关系是指变异与临床表型的关系，更确切的是指变异的致病性。
第一变异特征值可来自于各公开数据库的收录信息，例如HGMD(http://www.hgmd.cf.ac.uk/ac/index.php)、ClinVar(https://www.ncbi.nlm.nih.gov/clinvar/)、LOVD(https://www.lovd.nl/3.0/search)等。在此，并不限制第一变异特征值的来源，包括具体的数据库及数据库的数目。各数据库所记录的表型关系判定结果通常基于个性化的判定细则，其判定结果也常采取个性化的字段进行描述。以HGMD为例，其表型关系判定结果包括DM(disease-causing mutation)、DM？(likely disease-causing mutation)、DP(disease-associated polymorphism)、FP(in vitro or in vivo functional polymorphism)、DFP(disease-associated polymorphism with additional functional evidence)、R(retired record)。在本公开的实施例中，可利用待分析序列变异数据中的各变异唯一身份标识(如cpra)，在公开数据库中提取第一变异特征。
第二变异特征值中变异位点等位基因频率数据，可来自于各公开数据库的收录信息，例如gnomAD(https://gnomad.broadinstitute.org/)、EVS(http://evs.gs.washington.edu/EVS)、1000Genomes Project(http://browser.1000genomes.org)等。在此，并不限制第二变异特征值中变异位点等位基因频率数据的来源，包括具体的数据库及数据库的数目。第二变异特征值中功能预测数据，可来自于现有多种功能预测软件的预测结果，从预测结果中可选取一个或多个具体指标作为特征值。此外，若公开数据库对功能预测数据进行了收录，则可直接从公开数据库中提取对应的功能预测数据。在此，并不限制第二变异特征值中功能预测数据的来源，包括具体的软件以及软件的数目、具体的指标以及指标的数目。所述的功能预测软件的预测数据可包括蛋白保守性预测数据、核酸保守性预测数据、剪切危害性预测数据、变异位点有害程度预测数据至少之一但不限于此。在本公开的实施例中，可利用待分析序列变异数据中的各变异唯一身份标识(如cpra)，在公开数据库中提取第二变异特征。
S3，将第一变异特征集输入训练好的第一表型关系预测模型，得到第一表型关系预测结果，并将第二变异特征集输入训练好的第二表型关系预测模型，得到第二表型关系预测结果。
S4，取第一表型关系预测结果和第二表型关系预测结果的并集，获得第三表型关系预测结果。
作为一优选实施方案，在步骤S4后还可以包括：
将所述第三表型关系预测结果与对应的变异解读结果进行比较；当所述第三表型关系预测结果与所述变异解读结果一致时，判定所述变异解读结果可信；当所述第三表型关系预测结果与所述变异解读结果不一致时，判定所述变异解读结果不可信。
其中，所述变异解读结果是指与所述第三表型关系预测结果采取不同方法获得的对所述待分析序列变异数据各变异位点的表型关系判定结果。人工解读结果是所述变异解读结果的常见类型。此外，可采取本领域的惯常手段，分别对判定为可信的变异解读结果、判定为不可信的变异解读结果添加自定义的标识，以便后续数据处理。本方案可以快速的对大量变异解读结果进行可信、不可信的判定，并在后续环节中按类型分别进行处理。例如，可对所述变异解读结果判定为不可信的变异位点进行复核。通常采取人工复核的方式进行确认。本方案可以极大的减少后续环节的工作量，降低人工成本，提高解读效率。
不同于现有技术中使用单一模型进行预测，包含步骤S1至S4的技术方案根据数据类型的不同，使用了两个相对独立的模型(即第一表型关系预测模型和第二表型关系预测模型)进行预测，两模型具有互补作用。第一表型关系预测模型能够弥补第二表型关系预测模型没有变异位点功能数据、共分离数据和新发数据等可能导致的假阴性，第二表型关系 预测模型能够弥补第一表型关系预测模型中因变异没有公开的变异位点表型关系判定结果导致的假阴性。
第一表型关系预测模型和第二表型关系预测模型，均可基于机器学习模型构建，构建方法为：
S31，构建数据集，包括第一数据集和第二数据集，第一数据集为选定变异位点的第一变异特征及该变异位点的可信表型关系判定结果，第二数据集为选定变异位点的第二变异特征以及该变异位点的可信表型关系判定结果。
其中，选定变异位点是指，包括全部表型关系判定结果类型的多个变异位点。在一优选实施方案中，所述可信表型关系判定结果可以是根据行业指南或共识人工对表型关系判定的结果；所述可信表型关系判定结果也可以是多个公开的变异位点表型关系判定结果均一致的结果。
S32，用所述第一数据集对第一机器学习模型进行训练和评估，用第二数据集对第二机器学习模型进行训练和评估。
在一优选实施方案中，第一机器学习模型、第二机器学习模型可选自逻辑回归、朴素贝叶斯、支持向量机和人工神经网络至少之一。
S33，根据所述第一机器学习模型的评估结果确定第一表型关系预测模型，根据第二机器学习模型的评估结果确定第二表型关系预测模型。
在一优选实施方案中，第一表型关系预测模型为逻辑回归模型，第二表型关系预测模型为神经网络模型。
当本方案应用于人类孟德尔遗传病相关变异分析时，具体过程如下：
在该实施例中，第一表型关系预测模型用于快速对公共数据库有记录的变异以统一标准实现致病性判定。第一表型关系预测模型的构建过程包括：
A2，构建第一数据集，第一数据集包括公共变异数据库中记录的致病性判定结果以及解读专家做出的致病性判定结果。
其中，公共变异数据库可包括ClinVar、HGMD、LOVD、UMD(http://umd-predictor.eu/)等数据库中的至少一个。
具体地，在从公共变异数据库收集变异数据之后，可利用变异注释工具对从各个公共变异数据库收集的变异数据进行注释，得到其cpra信息，并将该信息作为变异的唯一身份标识。以变异的cpra为索引，将各个公共变异数据库的变异数据进行汇总。
同时，由解读专家根据ACMG(The American College of Medical Genetics and Genomics，美国医学遗传学与基因组学学会)变异解读指南，结合人群频率数据库、预测软件和医学文献等信息，对相同变异进行致病性判定，该判定结果即为可信表型关系判定结果，并将 该结果作为训练数据的标准结果。
进一步地，需建立公共变异数据库中记录的致病性判定结果、解读专家做出的致病性判定结果与满足机器学习模型输入要求字符的唯一映射关系。
第一数据集的结构和内容如表1所示。
表1
Figure PCTCN2021131904-appb-000001
由于各个公共变异数据库和解读专家的致病性判定结果均为定序的离散型数据，因此本公开将第一数据集中的字符型数据转化为数值型数据，如表2所示。
表2
Figure PCTCN2021131904-appb-000002
B2，将第一数据集中的数据按照第一预设比例随机分为第一训练集和第一测试集。
其中，第一预设比例可以是8:2。
C2，利用第一训练集中的数据对第一机器学习模型进行训练，并利用第一测试集对训练得到的第一机器学习模型进行评估。
其中，第一机器学习模型以选择逻辑回归、朴素贝叶斯、支持向量机和人工神经网络等机器学习模型中的至少一个。
D2，根据评估结果从训练得到的多个第一机器学习模型中确定第一表型关系预测模型。
具体地，第一表型关系预测模型要实现的功能是根据多个不同的公共变异数据库的变异致病性判定输出唯一的变异致病性判定，而对变异进行致病性判定是一个特征较少的多分类问题，故可以选择逻辑回归、朴素贝叶斯、支持向量机和人工神经网络等机器学习模型来实现该功能。
作为一个实施方式，可选择多个不同的机器学习模型，并通过第一训练集对选择的各个机器学习模型各自独立进行训练，并通过第一测试集对训练的各个机器学习模型的性能进行评估，评估指标可包括准确率、精准率、召回率等。根据各个模型的性能评估结果，选择性能最优的一个模型作为本公开所使用的第一表型关系预测模型，例如，可将训练得 到的准确率最高的模型作为第一表型关系预测模型。
在利用第一表型关系预测模型预测表型关系时，可提取各变异的第一变异特征值；将第一变异特征值输入第一表型关系预测模型，即可得到第一表型关系预测结果。
需要说明的是，第一表型关系预测模型是对公共数据库有收录的变异、按预设的表型关系、输出预测结果。以人类孟德尔遗传病相关变异致病性判定为例，预设的表型关系通常有三类(致病的或可能致病的、良性的或可能良性的、意义不明确的)、五类(致病的、可能致病的、良性的、可能良性的、意义不明确的)的不同分法，对此并不做限制。
相应地，第二表型关系预测模型的构建过程可包括：
A3，构建第二数据集，第二数据集包括变异位点等位基因频率数据、功能预测数据以及可信表型关系判定结果。
具体地，等位基因频率数据可来源于gnomAD、千人数据库、ExAC数据库等；功能预测数据来源于各预测软件的数据，如SIFT、Polyphen2、MutationTaster、GERP++、DANN等。由此可见，与第一表型关系预测模型相区别的是，第二表型关系预测模型是基于判断表型关系的原始数据得到，而第一表型关系预测模型是基于公开的表型关系判定结果得到。
第二数据集中的变异位点需在多个公开的变异位点表型关系判定结果均一致，该一致的判定结果即为可信表型关系判定结果，并将该结果作为训练数据的标准结果。
进一步地，需建立等位基因频率数据、功能预测数据以及可信表型关系判定结果与满足机器学习模型输入要求字符的唯一映射关系。
B3，将第二数据集中的数据按照第二预设比例随机分为第二训练集和第二测试集。
其中，第二预设比例可以是8:2。
C3，利用第二训练集中的数据对第二机器学习模型进行训练，并利用第二测试集对训练得到的第二机器学习模型进行评估。
其中，第二机器学习模型可以选择逻辑回归、朴素贝叶斯、支持向量机和人工神经网络等机器学习模型。
D3，根据评估结果确定训练得到最优第二机器学习模型，并将最优第二机器学习模型作为第二表型关系预测模型。
在利用第二表型关系预测模型预测表型关系时，可提取各变异的第二变异特征值；将第二变异特征值输入第二表型关系预测模型，即可得到第二表型关系预测结果。
在本公开的一个实施例中，实际应用中，基于待分析序列变异数据可通过人工解读的方式获得变异解读结果。同时，将相同变异位点的对应数据输入训练好的第一表型关系预测模型，获得第一表型关系预测结果；输入训练好的第二表型关系预测模型，获得第二表型关系预测结果；取第一表型关系预测结果与第二表型关系预测结果的并集作为第三表型 关系预测结果。可将变异解读结果与对应的第三表型关系预测结果进行比较，若两者结果一致则判定变异解读结果可信，否则判定不可信。不可信的变异解读结果需进一步人工复核。
需要说明的是，上述用于训练的可信表型关系判定结果是被认为准确的、少量的；用于与第三表型关系预测结果进行比较的解读结果可以是采取与第三表型关系预测结果任何不同方法获得的对待分析序列变异数据各变异位点的表型关系判定结果，如本领域技术人员的人工解读结果。
在本公开的一个实施例中，如图3所示，序列变异分析方法还可包括：
S5，过滤待分析序列变异数据中的第一变异位点，第一变异位点位于参考序列的简单重复区、与表型改变弱相关、且为非单碱基变异。
其中，简单重复区是指由1-5碱基重复单元构成的重复区域，例如AAA、CAACAACAACAA。与表型改变弱相关可以是指，变异位于非编码区或内含子区，且变异的等位基因频率大于0.05。变异位于非编码区或内含子区可以是指变异的功能注释为-(表示无功能注释)、intron(表示内含子)、3′-UTR(表示3′端的非翻译区)、5′-UTR(表示5′端的非翻译区)、nochange(与参考mRNA序列相比未发生改变)。
作为一个示例，第一变异位点还满足如下任一条件：
条件一：第一变异位点的总检出次数大于第一预设值，且低质量检出次数大于第二预设值；
条件二：位于满足条件一的变异位点参考序列物理位置上下游第三预设值内。
作为一个示例，可预先建立包含第一变异位点的第一变异位点数据集，其构建步骤如下：
A1，获取多个样本待分析序列变异数据。
具体地，该步骤与前述S1步骤相同。更具体地，可收集基因测序仪(如MGISEQ-2000测序仪)上固定测序芯片(如LCY171芯片)上的样本测序数据，进而从样本测序数据中得到变异位点和对应的变异唯一身份标识。
需要说明的是，实际使用中，第一变异位点数据集是随着样本的累计不断更新，实现对待分析序列变异数据的过滤。
B1，判断所得到的各变异位点是否位于参考序列的简单重复区，若是则记录该变异位点对应的变异唯一身份标识。
C1，对B1结果进行过滤，获得满足预设过滤条件的变异位点。
其中，所满足预设过滤条件的变异位点可为：功能注释为-(表示无功能注释)、intron(表示内含子)、3′-UTR(表示3′端的非翻译区)、5′-UTR(表示5′端的非翻译区)、nochange，且频率(表示变异在人群中的概率)大于0.05的变异。
具体地，可基于变异功能及变异频率进行过滤，过滤条件为：功能注释为-、intron、3′-UTR、5′-UTR、nochange，且频率大于0.05。该过滤条件的设置是为了挑选出大概率不会引起基因表型改变的变异，即位于非外显子区的高频变异。
D1，统计C1结果中各变异位点的低质量检出次数和总检出次数。
具体地，经步骤C1处理后，以变异唯一身份标识为索引统计低质量(low quality)检出次数，高质量(high quality)检出次数以及总(total)检出次数。其中，总检出次数为低质量检出次数与高质量检出次数的加和。
E1，利用总检出次数大于第一预设值，低质量检出次数大于第二预设值，且为非单碱基变异的变异，构建第一变异位点数据集。
具体地，统计规则为：1)总检出次数≥第一预设值(如8)，低质量检出次数≥第二预设值(如1)，非单碱基变异的变异位点；2)查找1)中变异位点参考序列物理位置上下游第三预设值(如3bp)内的、仍位于简单重复区的、非单碱基变异的变异位点。需要说明的是，在简单重复区，变异位点与参考序列比对的位置会出现波动，不同数据库或分析软件的记录方式不同，但实质仍为同一变异。例如，参考序列ATATAT，变异位点序列AT，变异位点序列可分别与参考序列的第一、三、五位进行匹配。因此设置了统计规则2)以确保第一变异位点的全面收录。由满足上述统计规则1)和2)中任一个的变异位点组成第一变异位点数据集。
进一步地，在过滤待分析序列变异数据中的第一变异位点时，可先得到待分析序列变异数据中各变异位点的变异唯一身份标识，并基于该变异唯一身份标识进行过滤。分析过程具体可为：调用第一变异位点数据集，并基于待分析序列变异数据的变异唯一身份标识遍历第一变异位点数据集。若存在一致的变异唯一身份标识，则说明待分析序列变异数据中的该变异位点为第一变异位点，需被过滤；否则需保留。
作为一个示例，用于表型关系预测的数据，可以是经步骤S5过滤操作之后的数据，即步骤S2-S4可在步骤S5之后执行，并利用步骤S5过滤后的数据。由此，通过减少表型关系预测的数据，可避免不必要的预测，提高预测效率。
在本公开的一个实施例中，如图4所示，序列变异分析方法，方法还可包括：
S6，获取待分析序列变异数据的氨基酸水平注释结果。
S7，根据氨基酸水平注释结果判断待分析序列变异数据是否存在复杂变异。
其中，复杂变异可包括插入和替换复合事件、缺失和缺失复合事件、缺失和替换复合事件，且各类复杂变异对各具体变异的数目无限制。此外，可采取本领域的惯常手段，对判定为存在复杂变异和/或不存在复杂变异的位点添加自定义的标识，以便后续数据处理。
作为一个示例，存在复杂变异的变异位点可满足如下条件：
至少两个导致氨基酸改变的变异位点参考序列坐标重叠；或者，
至少两个导致氨基酸改变的变异位点参考序列坐标相邻且影响同一个编码氨基酸。
在该实施例中，考虑到利用变异检测工具(如GATK软件)检测变异时，只会单纯的检测到替换事件或插入事件，不能检测到包含插入和替换的复合事件。为弥补该缺陷，本公开对变异检测工具检测到的变异注释，进行复杂变异判断，具体为判断相同基因编码区相邻位置的变异是否影响同一个氨基酸残基编码。若存在复杂变异，说明需要进行复核，例如通过人工进行合并解读；反之亦然。
具体而言，获取待分析序列变异数据的氨基酸水平注释结果，通常采取HGVS命名规则对注释结果进行命名。对存在氨基酸改变的变异，则在满足如下任意条件时，判定存在复杂变异：1)存在至少两个相同基因的变异发生参考序列坐标重叠；2)存在至少两个相同基因的变异发生参考序列坐标相邻且影响同一个氨基酸编码。条件1)中，“坐标”是指变异对应的参考序列的位置，“重叠”是指≥1bp的参考序列坐标重叠；该条件可以尽可能全面的筛选到复杂变异。条件2)中，“相邻”是指参考序列上下游4bp范围内；该条件可以准确的筛选到复杂变异。由此，通过复杂变异的判断能够弥补现有变异检测工具的缺陷，进而将复杂变异进行合并解读，可提升变异的解读效率。
作为一个示例，步骤S6中的待分析序列变异数据可以是步骤S1中的待分析序列变异数据，也可以是经第三表型关系预测结果与对应的变异解读结果进行比较后、带有判定结果的待分析序列变异数据。
在本公开的一个实施例中，如图5所示，序列变异分析方法还可包括：
S8，对待分析序列变异数据进行特征提取，得到第三变异特征集。
其中，第三变异特征集中的第三变异特征值可包括变异支持数据(当待分析序列变异数据是通过测序手段获得时，该变异支持数据可包括质量值、测序深度、支持变异的reads比值)、等位基因频率(如ESP6500_MAF、G1000_AF)。
S9，将第三变异特征集输入训练好的变异类型预测模型，得到待分析序列变异数据中各变异位点所属各变异类型的预测概率。
其中，变异类型可包括纯合变异、杂合变异、无变异。无变异可为参考序列在此位点的纯合基因型。
S10，根据预测概率确定对应变异位点的变异类型预测结果。
具体地，确定变异类型预测结果的方法为：将各变异类型的预测概率进行比较，并确定预测概率最大的变异类型；将最大的预测概率与预设阈值进行比较；如果预测概率大于预设阈值，则确定该最大的预测概率所对应的变异类型为该变异位点的变异类型预测结果。
作为一个示例，变异类型预测模型的构建过程可包括：
A4，构建第三数据集，第三数据集包括质量值、测序深度、支持变异的reads比值、等位基因频率以及可信的变异类型判定结果。
其中，可信的变异类型判定结果可以是使用行业内共识金标准手段验证所获得的结果，例如，sanger验证结果。
B4，将第三数据集中的数据按照第三预设比例随机分为第三训练集和第三测试集。
其中，第三预设比例可以是8:2。
C4，利用第三训练集中的数据对第三机器学习模型进行训练，并利用第三测试集对训练得到的第三机器学习模型进行评估。
具体地，以计算贝叶斯分类器所需的先验概率和条件概率，构建变异类型预测模型。
D4，根据评估结果从训练得到的多个第三机器学习模型中确定第三表型关系预测模型。
通过第三测试集中的数据对变异类型预测模型进行模型测试，通过变异类型预测模型判断测试集数据的所属分类，并且根据判断正确的数量来求得变异类型预测模型的准确度。
具体地，挑选经过sanger验证的变异位点作为训练样本，收集其相关数据组成第三训练集，这些训练样本一共被分为K个变异类型c 1、c 2、…、c K；预处理获得的变异数据，获取变异信息相关特征值，记为x i；基于选取的训练集统计出每个特征值的先验概率，该先验概率可通过第三训练集中的各类样本出现的次数来估计，以变异类型为c1时特征值x1为例，其先验概率为：
Figure PCTCN2021131904-appb-000003
对于每个类别c k，基于第三训练集计算先验概率P(c k)和类条件概率，利用贝叶斯公式估计给定训练特征x的条件概率P(c k|x)，即
Figure PCTCN2021131904-appb-000004
其中，P(x)是用于归一化的证据因子，对所有类别的取值为固定值。
通过第三测试集对以上机器学习模型的性能进行评估，评估指标可包括准确率、精准率，召回率。
在利用该模型时，将待测变异经过构建的变异类型预测模型进行预测，获得每个类别的概率P(c k|x)。取P max(c k|x)，判定数据x属于对应的类别c k。或者，设定质控阈值，当且仅当P max(c k|x)大于预设阈值时，判定数据x属于对应的类别c k。
作为一个示例，步骤S8中的待分析序列变异数据可以是步骤S1中的待分析序列变异数据，也可以是经第三表型关系预测结果与对应的变异解读结果进行比较后、带有判定结果的待分析序列变异数据，还可以是带有是否存在复杂变异结果的待分析序列变异数据。
作为一优选实施方案，在步骤S10后还可以包括：当所述变异类型预测结果仅包含参考序列基因型，则判定该变异位点不可信；当所述变异类型预测结果包含至少一个非参考序列基因型，则判定该变异位点可信。此外，可采取本领域的惯常手段，对判定为不可信和/或可信的变异位点添加自定义的标识，以便后续数据处理。本方案可以快速的对大量变异位点进行可信、不可信的判定，并在后续环节中按类型分别进行处理。例如，仅对判定为不可信的变异位点进行验证。又如，仅对判定为不可信的、与表型改变相关的变异位点进行验证，其中与表型改变相关的变异位点可为致病变异。通常采取本领域公认的金标准方式进行验证，例如一代测序验证。本方案可以极大的减少后续验证环节的工作量，降低人工成本，提高解读效率。
由此，本公开的序列变异分析方法，可快速判定该变异是否需要进行复核，是否需要实验验证等。对整个解读环节，一定程度上减少了变异实验验证成本，且对于人工解读或自动化解读结果进行分析，快速输出需复核的变异，减少了人工审核成本，确保结果的准确性。
下面结合一个具体实施方式描述本公开实施例的序列变异分析方法：
在该具体实施方式中，以人类孟德尔遗传病变异分析为例，从临床数据中选择599例样本的7937个变异解读数据，作为待分析序列变异数据，该临床数据需是合法合规获取的。具体实施如下：
1第一变异位点过滤
1.1构建第一变异位点数据集
1.1.1持续收集MGISEQ-2000测序仪上LCY171测序芯片上的样本测序数据，总共收集3038个样本的数据，将位于简单重复区的变异标记为yes。例如，chr1-241663902-T-TGAGAGAGA，参考序列存在“GA”重复，标记为yes。
1.1.2基于变异功能及等位基因频率对标记为yes的变异位点进行过滤。实施例中可设置的过滤条件为：①功能注释为“-”、“intron”、“3-UTR”、“5-UTR”、“nochange”的变异；②频率大于0.05的变异，具体可通过G1000、ESP6500、genomAD等人群数据库进行判定。
例如，变异chr11-108150207-CT-C，在G1000、ESP6500、genomAD三个数据库中记录 的变异的等位基因频率均大于0.05，因此该变异被过滤。
1.1.3对1.1.2过滤后保留的变异位点分别统计低质量(low quality)检出次数、高质量(high quality)检出次数以及总(total)检出次数。其中，总检出次数为低质量检出次数与高质量检出次数的加和。
1.1.4构造第一变异位点数据集
收录满足以下两条件之一的变异位点进行第一变异位点数据集的构建。第一变异位点数据集中至少需记录被收录变异位点的cpra信息。
①标签为yes，总检出次数≥8，低质量检出次数≥1，非单碱基变异(即ref碱基数≠alt碱基数)的变异；
②查找①中附近3bp内的ref碱基数≠alt碱基数，标签为yes的变异。
1.2查找及过滤第一变异位点
提取待分析数据的变异cpra信息，遍历第一变异位点数据集。若cpra与数据集中某一cpra一致，则输出“Yes”，说明存在第一变异位点，是需过滤的变异；否则输出“No”，说明不存在第一变异位点，是需保留的变异。
2复杂变异判定
具体实施如下：
2.1根据是否存在pHGVS点的结果判断变异是否存在氨基酸改变。
在一些实施方式中，复杂变异判定还可以处理接收到的数据(例如，以VCF格式、FASTQ格式等的输入数据)，利用生信软件(如GATK、ANNOVAR、Alamut等)进行氨基酸水平的注释，确定变异是否包含pHGVS点的结果。例如，变异chr7-142458526-A-G，氨基酸水平的注释结果为p.Asn54Ser，即包含pHGVS点。
2.2将存在氨基酸改变的变异进行位置判定，判断是否存在≥2个编码区相邻的同一基因变异。例如，检出变异同时包含chr7-142458526-A-G(c.161A>G，p.Asn54Ser)和chr7-142458527-C-G(c.162C>G，p.Asn54Lys)，前者为cDNA 161位后者为162位，两者相邻。
2.3根据相邻变异的变异类型及变异氨基酸改变位置判定是否存在复杂变异。以上两个变异(即chr7-142458526-A-G，chr7-142458527-C-G)都为SNP(Single Nucleotide Polymorphism，单核苷酸多态性)，氨基酸改变的位置按照pHGVS均导致第54位氨基酸Asn的改变，两者具有相同的氨基酸改变位置，判定为存在复杂变异。
对于判定为存在复杂变异的变异位点，使用自定义字符进行标识。后续可进行合并解读，从而可提高解读效率。
3表型关系预测
3.1第一表型关系预测模型
3.1.1搭建
3.1.1.1训练数据收集
从公共变异数据库ClinVar和HGMD获取总共3万多个变异及其致病性判定作为训练数据。
3.1.1.2训练数据预处理和特征提取
①训练数据预处理
利用变异注释工具，对训练数据进行注释，得到其cpra信息(基因组版本为hg19)，并将该信息作为变异的唯一身份标识。以变异的cpra为索引，将公共变异数据库的变异及其致病性判定整理汇总到一个训练汇总数据库中。由解读专家根据ACMG变异解读指南，结合人群频率数据库、预测软件和医学文献等，对汇总数据库中的变异人工进行致病性判定，并将判定结果作为训练数据的标准结果。经预处理后的训练汇总数据库结构和内容如表3所示。
表3
Figure PCTCN2021131904-appb-000005
Figure PCTCN2021131904-appb-000006
②训练数据特征提取
由于公共变异数据库和解读专家的致病性判定结果均为定序的离散型数据，故可以将汇总数据库中的字符型数据根据表4～表6所示的对应关系，转化为数值型数据，完成第一训练集的构建。经特征提取后的第一训练集的结构和内容如表8所示。
表4
HGMD致病性判定(字符型) HGMD致病性判定(数值型)
Null(该数据库无该变异记录) 0
R 1
FP 2
DP 3
DFP 4
DM？ 5
DM 6
表5
Figure PCTCN2021131904-appb-000007
Figure PCTCN2021131904-appb-000008
表6
Figure PCTCN2021131904-appb-000009
表7
Figure PCTCN2021131904-appb-000010
3.1.1.3模型选取和训练
以第一预设比例如8:2，将第一训练集随机拆分为第一训练集和第一测试集。通过逻辑回归模型，使用第一训练集进行训练。
3.1.1.4模型评估和确定
通过第一训练集对逻辑回归模型进行训练，经第一测试集评估其准确率为91.0％，测试性能数据如表8所示。
表8
Figure PCTCN2021131904-appb-000011
Figure PCTCN2021131904-appb-000012
3.1.2应用
3.1.2.1预测数据收集
收集待预测变异，如NM_000267.3(NF1):c.1722-2A>G，NM_000057.4(BLM):c.893C>T(p.Thr298Met)和NM_000244.3(MEN1):c.670-6C>T，以及其在公共变异数据库ClinVar和HGMD的致病性判定作为待预测数据。可选地，待预测变异可取自上述第一变异位点过滤后的变异位点。
需要说明的是，若待预测变异在这2个公共变异数据库均无结果，则该变异无法通过本公开实现致病性判定。
3.1.2.2训练数据预处理和特征提取
利用变异注释工具对待预测变异进行注释，得到其cpra信息(基因组版本为hg19)，并将该信息作为变异的唯一标识。以变异的cpra为索引，将待预测变异在公共变异数据库ClinVar和HGMD的致病性判定整理汇总到一个数据库中。因为公共变异数据库的致病性判定结果均为定序的离散型数据，根据表8所示转化逻辑将预测汇总数据库中的字符型数据转化为数值型数据，完成预测数据处理。经预处理和特征提取后的预测汇总数据库结构和内容如表9所示。
表9
Figure PCTCN2021131904-appb-000013
3.1.2.3模型结果输出
将预测汇总数据库中的数据转换成模型接受的输入格式，并将相应数据输入训练好的第一表型关系预测模型，即可得到上述3个待预测变异的变异致病性判定，即第一表型关系预测结果，结果如表10所示。通过表4和表5所示的数值型和字符型数据的转换可以得到 本模型。将NM_000267.3(NF1):c.1722-2A>G、NM_000057.4(BLM):c.893C>T(p.Thr298Met)和NM_000244.3(MEN1):c.670-6C>T分别预测为Pathogenic/Likely Pathogenic、Benign/Likely benign和VUS。
表10
Figure PCTCN2021131904-appb-000014
3.2第二表型关系预测模型
3.2.1搭建
3.2.1.1获取经专家解读具有明确致病性结果的变异数据。
具体地，变异数据可取自ClinVar数据库，ClinVar数据库(VCF_GRCH37，fileDate＝2017-10-29)中包含近30万个人类遗传变异，其涵盖了从致病变异到良性变异在内的各种类别的变异。该ClinVar数据库中包含明确致病位点46585个，筛选在HGMD数据库中为被认为是”DM”，且在dbNSFP(https://sites.google.com/site/jpopgen/dbNSFP)数据库中的变异位点共21105个，选择其中的一半共10552个作为第二训练集中的致病变异，另10553个致病位点组成第二测试集的第一测试子集。该ClinVar数据库包含良性位点23892个，选择在dbNSFP数据库中的位点共4664个作为第二训练集中的良性变异。对dbNSFP数据库中缺失值的处理，为直接赋值为0。
3.2.1.2预处理第二训练集和第一测试子集中的变异数据，以变异的cpra为变异的唯一标识及索引，获取用于训练的变异信息特征值x n。具体实施例中，该模型基于等位基因频率数据库1000 Genomes Project、ESP6500，以及多个功能预测软件数据。具体地，功能预测软件可分为，用于蛋白保守性预测例如SIFT、Polyphen2、MutationTaster，用于核酸保守性预测例如GERP++，用于剪切危害性预测例如dbscSNV，用于变异位点有害程度预测例如DANN。提取各功能预测软件的一个或多个技术指标作为特征值，共获取变异的多个特征值(计为x n)，具体特征值分别为：SIFT_converted_rankscore，Polyphen2_HDIV_rankscore，Polyphen2_HVAR_rankscore，LRT_converted_rankscore，LRT_Omega，MutationTaster_converted_rankscore，MutationAssessor_score_rankscore，FATHMM_converted_rankscore，PROVEAN_converted_rankscore，MetaSVM_rankscore， MetaLR_rankscore，REVEL_score，CADD_raw_rankscore，DANN_rankscore，GERP++_RS_rankscore，splicing_consensus_ada_score，splicing_consensus_rf_score，phyloP100way_vertebrate_rankscore，phyloP20way_mammalian_rankscore，phastCons100way_vertebrate_rankscore，phastCons20way_mammalian_rankscore，SiPhy_29way_logOdds_rankscore，1000Gp3_AF，ESP6500_AA_AF。可选的各具体特征的详细情况可参见https://drive.google.com/file/d/1Vse3b_qw_E46eDcsuLegF5HxpodAZ6Og/view、https://drive.google.com/file/d/12-sSAB_hP9fWNxvJZ6IpMZvGXCVYmfns/view。
3.2.1.3构造第二机器学习模型。基于变异信息相关特征值x n及变异的唯一致病性判定，可以选择逻辑回归、朴素贝叶斯、支持向量机和人工神经网络等机器学习模型作为第二机器学习模型来实现预测功能。
在本公开的一个具体实施例中，可选择神经网络模型作为第二机器学习模型。构建的神经网络模型可包含三个隐藏层，节点个数分别为16、128、16；权值初始化函数为uniform，激活函数为hard_sigmoid，优化器采用Adadelta方法的默认参数，神经元失活概率dropout_rate为0.05，训练迭代次数800，一次性训练所选取的样本数batch_size为64。
3.2.a4模型结果评估
可通过2个独立的测试子集对训练得到的第二机器学习模型进行性能评估，以测试及验证模型的准确性。测试子集1即为上述的第一测试子集，测试子集2可以是基于临床解读积累的通过ACMG解读出来的变异结果，筛选临床明确认为是Pathogenic(致病变异)和Benign(良性变异)的位点，其中，对于多次解读到并且解读结果不一致位点选择解读结果次数最多的结果。筛选在dbNSFP数据库中的变异，去除训练集中的变异，得到共728个变异，其中致病变异618个，良性变异108个。
通过以上2个测试子集对训练得到的第二机器学习模型的性能进行评估，ClinVar数据库中10553个致病位点，准确性为98.77％；临床728个位点，准确性为97.79％。ROC(Receiver Operating Characteristic，受试者工作特征曲线)结果如图6所示，对临床728个位点预测，AUC可达0.99。类似方法中REVEL测试最好效果为AUC＝0.957；MLP测试最好效果为AUC＝0.94；DANN测试最好效果为AUC＝0.9459。显然，本公开的预测子模型较以往方法AUC可提升4～5个百分点，可以辅助遗传分析进行快速致病性审核，用以进行变异致病性质控，可达较高的准确性。
3.2b应用
3.2.b1收集待预测变异的特征值，即收集待预测变异在训练数据所用的各特征值x n，具体地与3.2.1.2所述相同的多个特征值。
其中，待预测变异可以是取自上述第一变异位点过滤后的变异位点。
3.2.b2将变异的特征值输入第二表型关系预测模型，即可得到待预测变异的表型关系预测结果，即第二表型关系预测结果。
3.3取第一表型关系预测结果和第二表型关系预测结果的并集，获得第三表型关系预测结果。将第三表型关系预测结果与相应的变异解读结果进行比较，当所述第三表型关系预测结果与所述变异解读结果一致时，判定所述变异解读结果可信；当所述第三表型关系预测结果与所述变异解读结果不一致时，判定所述变异解读结果不可信。其中，变异解读结果不可信的变异需进一步人工复核。
4变异类型预测
4.1模型搭建
4.1.1搭建训练集。从经过sanger验证变异结果的历史测序数据中选取843个变异位点，将其按照4:1的比例随机分为第三训练集和第三测试集，第三训练集用来对模型进行训练，第三测试集用来对模型进行测试。具有674个变异位点的训练集具体数据分布为：269个无变异位点，212个纯合变异位点，193个杂合变异位点。
4.1.2预处理获得的674个变异位点数据，获取变异信息相关特征值x n。选取的特征值可包括x 1质量值、x 2测序深度、x 3支持变异的reads比值、x 4ESP6500_MAF(等位基因频率)、x 5G1000_AF(等位基因频率)这五维特征，例如，变异chr2-223086092-G-A，sanger验证为杂合变异，其特征值预处理为：x 1＝99；x 2＝250；x 3＝0.4480；x 4＝0.000077；x 5＝0.001597；
4.1.3针对变异信息相关特征值x n，基于选取的第三训练集统计出每个特征值类的先验概率，通过第三训练集中各类数据出现的次数估计。以变异类型为c1时特征值x1为例，其先验概率即：
Figure PCTCN2021131904-appb-000015
4.1.4训练。基于674个变异位点的第三训练集计算先验概率P(c k)和类条件概率。第三训练集中的样本一共被分为K个类别，如K＝3，类别为c 1,c 2,c 3。对于类别c k(k＝1、2、3)，可令c 1＝0代表无变异位点，c 2＝1代表纯合变异位点，c 3＝2代表杂合变异位点，利用贝叶斯公式估计在给定训练数据x时的条件概率P(c k|x)，即
Figure PCTCN2021131904-appb-000016
以变异chr2-223086092-G-A为例，训练计算的结果为：P(c 1|x)＝0.00190984495430601；P(c 2|x)＝3.63E-07；P(c 3|x)＝0.998089791607098。
4.1.5模型结果评估及质控阈值确定
通过第三测试集对上述训练得到的机器学习模型的性能进行评估，评估指标可包括准确率、精准率、召回率。利用上述选取的169个变异数据作为第三测试集。选择大于预设阈值且最大预测概率所对应的变异类型为该变异位点的变异类型预测结果，其中预设阈值设定为0.8。具体测试结果详见表11。需要说明的是，变异类型预测模型的预设阈值可根据实际情况调整。
表11
Figure PCTCN2021131904-appb-000017
4.2应用
4.2.1构建的变异类型预测模型预测待测变异位点属于每个类别的概率P(c k|x)，其中，待测变异位点可以是取自上述第一变异位点过滤后和经第一表型预测模型、第二表型预测模型判定为可信的不需人工复核的变异位点。
4.2.2变异类型预测结果判定方法具体为，选择大于预设阈值0.8且最大预测概率所对应的变异类型为该变异位点的变异类型预测结果。以变异解读结果为致病变异的位点为例，当变异类型预测结果为纯合变异或杂合变异时，则判定该变异位点可信，无需进行实验验证；当变异类型预测结果为无变异类型时，则判定该变异位点不可信，需进行实验验证。
本公开整体性能评估：
选取599例样本的7937个变异位点作为待分析序列变异数据。对此变异数据通过人工解读的方式获得解读结果，将人工解读结果与本公开的各环节结果进行比较。经第一变异位点过滤，过滤掉690个第一变异位点，进而可减少8.69％的复核工作量。经复杂变异判定，确定存在34个复杂变异位点，需要合并后进行解读，可弥补变异检测工具的缺陷。经第一表型关系预测模型和第二表型关系预测模型的分析结果(如图7所示)，相比于其他单纯的致病性预测模型，两者相互结合可获得解读结果中与两个模型不一致的变异位点2041个，确定7247个变异位点中有5206个的解读结果可信无需复核，可减少71.84％复核工作量，且得到较全面的人工复核量。经变异类型预测模型分析，可进一步确认76个致病变异中的61个无需进行变异验证，减少80.26％变异验证工作量。
综上，本公开实施例的序列变异分析方法，能够快速、准确、全面的确定需过滤、复核、验证的变异。只需对判定为不可信的解读结果进行审核和维护，对判定为不可信的致 病变异进行验证。从而大大降低了人工成本，提升了基因变异位点的复核、验证效率，缩短了基因检测报告的出具时间，同时也提高了基因检测报告解读的准确性，优化了整个解读分析过程。
基于上述的序列变异分析方法，本公开提出了一种计算机可读存储介质。
在该实施例中，计算机可读存储介质上存储有计算机程序，计算机程序被处理器执行时，实现上述的序列变异分析方法。
图8是本公开序列变异分析系统的具体实施方式之一的结构框图。
序列变异分析系统100包括：获取模块110和第一分析模块120。
其中，获取模块110用于获取待分析序列变异数据；第一分析模块120用于对待分析序列变异数据进行特征提取，得到第一变异特征集、第二变异特征集，并将第一变异特征集输入训练好的第一表型关系预测模型，得到第一表型关系预测结果，将第二变异特征集输入训练好的第二表型关系预测模型，得到第二表型关系预测结果，以及取第一表型关系预测结果和第二表型关系预测结果的并集，得到第三表型关系预测结果。
在本公开的一个实施例中，序列变异分析系统100还可包括：第二分析模块130。
其中，第二分析模块130用于过滤待分析序列变异数据中的第一变异位点，第一变异位点位于参考序列的重复区、与表型改变弱相关、且为非单碱基变异。
在本公开的一个实施例中，序列变异分析系统100还可包括：第三分析模块140。
其中，第三分析模块140用于获取待分析序列变异数据的氨基酸水平注释结果，并根据氨基酸水平注释结果判断待分析序列变异数据是否存在复杂变异。
在本公开的一个实施例中，序列变异分析系统100还可包括：第四分析模块150。
其中，第四分析模块150用于对待分析序列变异数据进行特征提取，得到第三变异特征集，并将第三变异特征集输入训练好的变异类型预测模型，得到待分析序列变异数据中各变异位点所属各变异类型的预测概率，以及根据预测概率确定对应变异位点的变异类型预测结果。
需要说明的是，本公开实施例的序列变异分析系统100其他具体实施方式，可参见本公开上述实施例的序列变异分析方法的具体实施方式。
综上，本公开实施例的序列变异分析系统，能够快速、准确、全面的确定需过滤、复核、验证的变异。只需对判定为不可信的解读结果进行审核和进行维护，对判定为不可信的致病变异进行验证。从而大大降低了人工成本，提升了基因变异位点的复核、验证效率，缩短了基因检测报告的出具时间，同时也提高了基因检测报告解读的准确性，优化了整个解读分析过程。
以上实施例仅用以说明本公开的技术方案，而非对其限制；尽管参照前述实施例对本 公开进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本公开各实施例技术方案的精神和范围，均应包含在本公开的保护范围之内。

Claims (18)

  1. 一种序列变异分析方法，其特征在于，包括以下步骤：
    获取待分析序列变异数据；
    对所述待分析序列变异数据进行特征提取，得到第一变异特征集、第二变异特征集；
    将所述第一变异特征集输入训练好的第一表型关系预测模型，得到第一表型关系预测结果，并将所述第二变异特征集输入训练好的第二表型关系预测模型，得到第二表型关系预测结果；
    取所述第一表型关系预测结果和所述第二表型关系预测结果的并集，获得第三表型关系预测结果。
  2. 如权利要求1所述的序列变异分析方法，其特征在于，在所述获得第三表型关系预测结果后还包括：
    将所述第三表型关系预测结果与对应的变异解读结果进行比较；
    当所述第三表型关系预测结果与所述变异解读结果一致时，判定所述变异解读结果可信；
    当所述第三表型关系预测结果与所述变异解读结果不一致时，判定所述变异解读结果不可信。
  3. 如权利要求1所述的序列变异分析方法，其特征在于，所述方法还包括：
    过滤所述待分析序列变异数据中的第一变异位点，所述第一变异位点位于参考序列的简单重复区、与表型改变弱相关、且为非单碱基变异。
  4. 如权利要求3所述的序列变异分析方法，其特征在于，所述与表型改变弱相关是指，变异位于非编码区或内含子区，且所述变异的等位基因频率大于0.05。
  5. 如权利要求3所述的序列变异分析方法，其特征在于，所述第一变异位点还满足如下任一条件：
    条件一：总检出次数大于第一预设值，且低质量检出次数大于第二预设值；
    条件二：位于满足所述条件一的变异位点的参考序列物理位置上下游第三预设值内。
  6. 如权利要求1所述的序列变异分析方法，其特征在于，所述方法还包括：
    获取所述待分析序列变异数据的氨基酸水平注释结果；
    根据所述氨基酸水平注释结果判断所述待分析序列变异数据是否存在复杂变异。
  7. 如权利要求6所述的序列变异分析方法，其特征在于，存在复杂变异的变异位点满足如下条件：
    至少两个导致氨基酸改变的变异位点参考序列坐标重叠；或者，
    至少两个导致氨基酸改变的变异位点参考序列坐标相邻且影响同一个编码氨基酸。
  8. 如权利要求1所述的序列变异分析方法，其特征在于，所述方法还包括：
    对所述待分析序列变异数据进行特征提取，得到第三变异特征集；
    将所述第三变异特征集输入训练好的变异类型预测模型，得到所述待分析序列变异数据中各变异位点所属各变异类型的预测概率；
    根据所述预测概率确定对应变异位点的变异类型预测结果。
  9. 如权利要求8所述的序列变异分析方法，其特征在于，所述根据所述预测概率确定对应变异位点的变异类型预测结果，具体为：
    选择大于预设阈值且最大所述预测概率所对应的变异类型为该变异位点的变异类型预测结果。
  10. 如权利要求9所述的序列变异分析方法，其特征在于，确定对应变异位点的变异类型预测结果后还包括：
    当所述变异类型预测结果仅包含参考序列基因型，则判定该变异位点不可信；
    当所述变异类型预测结果包含至少一个非参考序列基因型，则判定该变异位点可信。
  11. 如权利要求8所述的序列变异分析方法，其特征在于，所述第三变异特征值包括变异支持数据以及等位基因频率，所述变异类型包括纯合变异、杂合变异、无变异。
  12. 如权利要求1所述的序列变异分析方法，其特征在于，所述第一变异特征集中的各第一变异特征值为公开的变异位点的表型关系判定结果，所述第二变异特征集中的各第二变异特征值为变异位点等位基因频率数据以及功能预测数据。
  13. 如权利要求12所述的序列变异分析方法，其特征在于，所述第二变异特征值中的所述功能预测数据包括蛋白保守性预测数据、核酸保守性预测数据、剪切危害性预测数据、变异位点有害程度预测数据至少之一。
  14. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，所述计算机程序被处理器执行时，实现如权利要求1-13中任一项所述的序列变异分析方法。
  15. 一种序列变异分析系统，其特征在于，包括：
    获取模块，用于获取待分析序列变异数据；
    第一分析模块，用于对所述待分析序列变异数据进行特征提取，得到第一变异特征集、第二变异特征集，并将所述第一变异特征集输入训练好的第一表型关系预测模型，得到第一表型关系预测结果，将所述第二变异特征集输入训练好的第二表型关系预测模型，得到第二表型关系预测结果，以及取所述第一表型关系预测结果和所述第二表型关系预测结果的并集，得到第三表型关系预测结果。
  16. 如权利要求15所述的序列变异分析系统，其特征在于，所述系统还包括：
    第二分析模块，用于过滤所述待分析序列变异数据中的第一变异位点，所述第一变异位点位于参考序列的重复区、与表型改变弱相关、且为非单碱基变异。
  17. 如权利要求15所述的序列变异分析系统，其特征在于，所述系统还包括：
    第三分析模块，用于获取所述待分析序列变异数据的氨基酸水平注释结果，并根据所述氨基酸水平注释结果判断所述待分析序列变异数据是否存在复杂变异。
  18. 如权利要求15所述的序列变异分析系统，其特征在于，所述系统还包括：
    第四分析模块，用于对所述待分析序列变异数据进行特征提取，得到第三变异特征集，并将所述第三变异特征集输入训练好的变异类型预测模型，得到所述待分析序列变异数据中各变异位点所属各变异类型的预测概率，以及根据所述预测概率确定对应变异位点的变异类型预测结果。
PCT/CN2021/131904 2021-11-19 2021-11-19 序列变异分析方法、系统以及存储介质 Ceased WO2023087277A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
CN202180103963.2A CN118302817A (zh) 2021-11-19 2021-11-19 序列变异分析方法、系统以及存储介质
EP21964431.7A EP4435791A4 (en) 2021-11-19 2021-11-19 SYSTEM AND METHOD FOR ANALYZING SEQUENCE VARIATION, AND STORAGE MEDIUM
AU2021474767A AU2021474767B2 (en) 2021-11-19 2021-11-19 Sequence variation analysis method and system, and storage medium
PCT/CN2021/131904 WO2023087277A1 (zh) 2021-11-19 2021-11-19 序列变异分析方法、系统以及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/131904 WO2023087277A1 (zh) 2021-11-19 2021-11-19 序列变异分析方法、系统以及存储介质

Publications (1)

Publication Number Publication Date
WO2023087277A1 true WO2023087277A1 (zh) 2023-05-25

Family

ID=86396019

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/131904 Ceased WO2023087277A1 (zh) 2021-11-19 2021-11-19 序列变异分析方法、系统以及存储介质

Country Status (4)

Country Link
EP (1) EP4435791A4 (zh)
CN (1) CN118302817A (zh)
AU (1) AU2021474767B2 (zh)
WO (1) WO2023087277A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117711487A (zh) * 2024-02-05 2024-03-15 广州嘉检医学检测有限公司 胚系SNV、InDel变异的鉴定方法、系统以及可读存储介质
CN118866092A (zh) * 2024-07-05 2024-10-29 华中农业大学 融合门控与线性注意力机制g×e互作的基因组预测方法
CN119601083A (zh) * 2024-11-16 2025-03-11 大连易康生物科技有限公司 一种基因测序数据处理方法及系统
CN119811479A (zh) * 2025-03-13 2025-04-11 北京市农林科学院信息技术研究中心 育种性状解析方法、装置、设备、介质及计算机程序产品
CN120470419A (zh) * 2025-07-15 2025-08-12 中南大学湘雅医院 变异富集区域的识别方法、装置及计算机设备
CN121483394A (zh) * 2026-01-08 2026-02-06 中南大学湘雅医院 面向新发变异的致病性预测方法

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119314553A (zh) * 2024-12-16 2025-01-14 上海第二工业大学 一种基于组合特征编码和dna结合位点的预测方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110832597A (zh) * 2018-04-12 2020-02-21 因美纳有限公司 基于深度神经网络的变体分类器
CN111063392A (zh) * 2019-12-17 2020-04-24 人和未来生物科技（长沙）有限公司 基于神经网络的基因突变致病性检测方法、系统及介质
US20200342955A1 (en) * 2017-10-27 2020-10-29 Apostle, Inc. Predicting cancer-related pathogenic impact of somatic mutations using deep learning-based methods
CN112687332A (zh) * 2021-03-12 2021-04-20 北京贝瑞和康生物技术有限公司 用于确定致病风险变异位点的方法、设备和存储介质

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8744982B2 (en) * 2011-05-12 2014-06-03 University Of Utah Research Foundation Gene-specific prediction

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200342955A1 (en) * 2017-10-27 2020-10-29 Apostle, Inc. Predicting cancer-related pathogenic impact of somatic mutations using deep learning-based methods
CN110832597A (zh) * 2018-04-12 2020-02-21 因美纳有限公司 基于深度神经网络的变体分类器
CN111063392A (zh) * 2019-12-17 2020-04-24 人和未来生物科技（长沙）有限公司 基于神经网络的基因突变致病性检测方法、系统及介质
CN112687332A (zh) * 2021-03-12 2021-04-20 北京贝瑞和康生物技术有限公司 用于确定致病风险变异位点的方法、设备和存储介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP4435791A4 *

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117711487A (zh) * 2024-02-05 2024-03-15 广州嘉检医学检测有限公司 胚系SNV、InDel变异的鉴定方法、系统以及可读存储介质
CN117711487B (zh) * 2024-02-05 2024-05-17 广州嘉检医学检测有限公司 胚系SNV、InDel变异的鉴定方法、系统以及可读存储介质
CN118866092A (zh) * 2024-07-05 2024-10-29 华中农业大学 融合门控与线性注意力机制g×e互作的基因组预测方法
CN119601083A (zh) * 2024-11-16 2025-03-11 大连易康生物科技有限公司 一种基因测序数据处理方法及系统
CN119811479A (zh) * 2025-03-13 2025-04-11 北京市农林科学院信息技术研究中心 育种性状解析方法、装置、设备、介质及计算机程序产品
CN120470419A (zh) * 2025-07-15 2025-08-12 中南大学湘雅医院 变异富集区域的识别方法、装置及计算机设备
CN121483394A (zh) * 2026-01-08 2026-02-06 中南大学湘雅医院 面向新发变异的致病性预测方法
CN121483394B (zh) * 2026-01-08 2026-03-27 中南大学湘雅医院 面向新发变异的致病性预测方法

Also Published As

Publication number Publication date
EP4435791A4 (en) 2025-09-10
AU2021474767B2 (en) 2026-02-05
AU2021474767A1 (en) 2024-06-06
EP4435791A1 (en) 2024-09-25
CN118302817A (zh) 2024-07-05

Similar Documents

Publication Publication Date Title
WO2023087277A1 (zh) 序列变异分析方法、系统以及存储介质
US10354747B1 (en) Deep learning analysis pipeline for next generation sequencing
US12242943B2 (en) Generating machine learning models using genetic data
Zhang et al. Genetic variants underlying differences in facial morphology in East Asian and European populations
CN111883210B (zh) 基于临床特征和序列变异的单基因病名称推荐方法及系统
CN115273970A (zh) 用于检测异常核型的方法和系统
US12272431B2 (en) Detecting false positive variant calls in next-generation sequencing
CN116486913A (zh) 基于单细胞测序从头预测调控突变的系统、设备和介质
CN116564406A (zh) 一种遗传变异自动化解读方法及设备
CN120164524B (zh) 一种遗传病基因检测的数据分析方法、系统及存储介质
WO2022123067A2 (en) Method and apparatus for classification and/or prioritization of genetic variants
US20250069702A1 (en) Population frequency modeling for quantitative variant pathogenicity estimation
Frazer et al. Large-scale clinical interpretation of genetic variants using evolutionary data and deep learning
AU2019480813A1 (en) Methods for determining chromosome aneuploidy and constructing classification model, and device
CN110211632A (zh) 一种基于神经网络的核苷酸单位点变异检测方法
CN117275577A (zh) 一种基于二代测序技术检测人线粒体遗传突变位点算法
US20240011105A1 (en) Analysis of microbial fragments in plasma
CN117497047B (zh) 基于外显子测序筛选肿瘤基因标志物的方法、设备和介质
CN114694752B (zh) 预测同源重组修复缺陷的方法、计算设备和介质
CN121506262B (zh) 一种胚胎植入前遗传学检测方法、系统、产品及设备
KR102532991B1 (ko) 태아의 염색체 이수성 검출방법
CN121483636A (zh) 一种基于机器学习的慢性淋巴细胞白血病预测模型的构建方法、构建系统及电子设备、存储介质
Chen et al. CoCoRV: a rare variant analysis framework using publicly available genotype summary counts to prioritize germline disease-predisposition genes
Nivashini et al. GENOME-WIDE IDENTIFICATION OF FUNCTIONAL GENETIC VARIANTS ASSOCIATED WITH PHENOTYPIC DIVERSITY
Fan et al. Stratifying variant deleteriousness and trait-modulating effect under human recent adaptation using the FIND model

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21964431

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 202180103963.2

Country of ref document: CN

WWE Wipo information: entry into national phase

Ref document number: 2021474767

Country of ref document: AU

Ref document number: AU2021474767

Country of ref document: AU

ENP Entry into the national phase

Ref document number: 2021474767

Country of ref document: AU

Date of ref document: 20211119

Kind code of ref document: A

WWE Wipo information: entry into national phase

Ref document number: 2021964431

Country of ref document: EP

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2021964431

Country of ref document: EP

Effective date: 20240619

WWE Wipo information: entry into national phase

Ref document number: 11202403283V

Country of ref document: SG