A scoring and ranking method and system for pathogenic mutations of monogenic genetic diseases

Abstract

The application discloses a single gene genetic disease pathogenic mutation scoring and sequencing method and system, relates to the biological medical technical field, and comprises the following steps: based on the pre-acquired phenotype data and the literature abstract, using a mixed strategy to extract the standardized phenotype terms from the pre-acquired symptom description, and scoring the correlation degree between the standardized phenotype terms and the pre-acquired gene mutation data by a similarity comparison method; annotating the pre-acquired gene variation database, and scoring the pre-acquired gene mutation data according to the gene variation annotation result; using a supervised learning model and a weighted summation method to perform multi-dimensional pathogenicity comprehensive scoring and evaluation, and obtaining a pathogenic mutation sequencing scoring result. The application recommends the most possible pathogenic mutation, so that the scoring and sequencing of the genetic disease pathogenic mutation has the advantages of high detection recall rate and high automation degree.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of biomedicine, in particular to a scoring and ranking method and system for pathogenic mutations of monogenic genetic diseases. BACKGROUND

[0002] Mendelian monogenic genetic diseases are caused by mutations in a single gene, have a clear pathogenic mechanism, and often involve rare or specific variations in individual patients. It is estimated that there are more than 8000 known monogenic genetic diseases, affecting about 1 / 20 of the global population. With the rapid development of second-generation high-throughput sequencing (NGS) technology, especially the widespread application of whole exome sequencing (WES) in clinical practice, molecular diagnosis of monogenic genetic diseases has become more efficient and feasible. Through WES technology, all the variation information in the exonic region of the patient can be obtained at one time, providing an important basis for the identification of pathogenic mutations.

[0003] In practical applications, WES can generate tens of thousands to hundreds of thousands of genetic variation sites, of which only a few are truly pathogenic mutations (Pathogenic), and the rest are mostly benign (Benign) or variants of uncertain significance (VUS). Therefore, how to quickly and accurately screen out the most likely pathogenic candidate mutations from the vast amount of variation data is one of the core challenges in the current molecular diagnosis of genetic diseases. The current mainstream analysis process usually relies on multiple databases (such as ClinVar, HGMD, COSMIC) and annotation tools (such as ANNOVAR, VEP, Transvar) to annotate the variations, but these methods often lack effective integration of patient phenotype information, resulting in low recall rate and easy omission of phenotype-related but not yet fully reported pathogenic mutations.

[0004] On the other hand, the clinical symptoms of patients are often recorded in a free text manner, which is not uniform in description and lacks standardization, and is difficult to be directly used for automatic analysis of computer programs. At present, the Human Phenotype Ontology (HPO) and the China Human Phenotype Ontology (CHPO) are gradually popularized as standardized terminology systems in the world, so as to standardize and structure the phenotype description, and better compare with the known gene-phenotype association database. Some data analysis institutions have introduced HPO into their systems, but it is difficult to link with the clinic due to lack of Chinese version. Non-clinical professionals are prone to bias in the process of converting HPO by secondary analysis of natural language phenotype information provided by doctors. In addition, it is still a challenge to automatically extract accurate HPO terms from unstructured text in practical application, and most processes still rely on manual annotation and review, which is time-consuming and prone to errors.

[0005] At present, some tools such as Phenolyzer, Phenomizer and Exomiser can screen candidate pathogenic genes on the basis of standardized phenotypes. These tools generally judge the matching degree of patient phenotypes and known diseases based on the semantic similarity calculation between HPO terms, and realize the similarity evaluation between HPOs through MICA (Maximum Information Common Ancestor) algorithm. However, the accuracy of such methods is limited by the hierarchical organization of HPO terms in anatomical structure, and cannot fully reflect the relevance of phenotypes in disease mechanism. For example, two phenotypes that are structurally similar (such as tricuspid valve anomaly HP:0001702 and tricuspid valve prolapse HP:0001704) may have a high semantic similarity, but their corresponding pathogenic genes are completely different, which may lead to an increase in false positives and false negatives.

[0006] In summary, the existing genetic disease diagnosis systems based on whole exome sequencing generally have the following problems: (1) The variant screening process relies on manual interpretation, and the degree of automation is low; (2) There is a lack of automatic analysis capability for patient unstructured clinical phenotype information; (3) The scoring system fails to comprehensively integrate the association of phenotype and gene, harmfulness of variation, frequency of occurrence, literature support and other multi-dimensional evidence. Therefore, there is an urgent need for a high-automation scoring and sorting method and system that can integrate and analyze clinical phenotypes and NGS variation data, which can combine phenotype matching, variation annotation, database information and other factors to automatically identify the most likely pathogenic mutations, so as to improve the diagnosis efficiency and accuracy of single gene genetic diseases.

[0007] In view of the problems in the related art, no effective solution has been proposed so far. SUMMARY

[0008] In view of the problems in the related art, the present application provides a scoring and ranking method and system for pathogenic mutations of monogenic genetic diseases to overcome the above technical problems existing in the prior art.

[0009] To this end, the present application adopts the following specific technical solutions:

[0010] According to an aspect of the present application, a scoring and ranking method for pathogenic mutations of monogenic genetic diseases is provided, which comprises the following steps:

[0011] S1, based on the pre-acquired phenotype data and literature abstracts, using a hybrid strategy to extract standardized phenotype terms from the pre-acquired symptom descriptions, and using a similarity comparison method to score the association degree between the standardized phenotype terms and the pre-acquired gene mutation data;

[0012] S2, annotating the pre-acquired gene variation database, and scoring the pre-acquired gene mutation data based on the gene variation annotation results;

[0013] S3, based on the scoring results of the association degree between the phenotype and the gene and the gene variation scoring results, using a supervised learning model and a weighted summation method to perform multi-dimensional pathogenicity comprehensive scoring and evaluation on the pre-acquired gene mutation data, and obtaining the pathogenic mutation ranking and scoring results.

[0014] Further, based on the pre-acquired phenotype data and literature abstracts, using a hybrid strategy to extract standardized phenotype terms from the pre-acquired symptom descriptions, and using a similarity comparison method to score the association degree between the standardized phenotype terms and the pre-acquired gene mutation data includes:

[0015] S11, based on the pre-acquired phenotype data and literature abstracts, using a dictionary method and a deep learning model to extract phenotype terms from the pre-acquired symptom descriptions, obtaining dictionary matching results and phenotype term recognition results;

[0016] S12, using a weighted summation method to integrate the dictionary matching results and the phenotype term recognition results to obtain standardized phenotype terms;

[0017] S13, using an OWL similarity semantic algorithm to compare the semantic similarity between the standardized phenotype terms and the pre-acquired gene mutation data, and scoring the association degree between the phenotype and the gene based on the similarity comparison results, to obtain the association degree scoring results between the phenotype and the gene.

[0018] Further, based on the pre-acquired phenotype data and literature abstracts, using a dictionary method and a deep learning model to extract phenotype terms from the pre-acquired symptom descriptions, obtaining dictionary matching results and phenotype term recognition results includes:

[0019] S111, based on the pre-acquired phenotype data, performing phenotype term matching on the clinical symptom description using a dictionary method to obtain a dictionary matching result;

[0020] S112, performing phenotype term matching on the pre-acquired literature abstract and the pre-acquired phenotype data based on positive and negative sample division, and constructing a supervised training set based on the positive and negative sample division result;

[0021] S113, training a deep learning model using the supervised training set, and performing phenotype named entity recognition on the clinical symptom description based on the trained deep learning model and a pre-set probability threshold to obtain a phenotype recognition result.

[0022] Further, the standardized phenotype term and the pre-acquired gene mutation data are compared in semantic similarity by an OWL similarity semantic algorithm, and the association degree between the phenotype and the gene is scored based on the similarity comparison result to obtain a phenotype and gene association degree scoring result, including:

[0023] S131, constructing a semantic association network of the standardized phenotype term and the pre-acquired gene mutation data by an OWL similarity semantic algorithm to obtain a phenotype and gene association set;

[0024] S132, according to the phenotype and gene association set, the frequency of occurrence of each phenotype and gene is counted, and the most specific common ancestor is found based on the frequency of occurrence of each phenotype and gene;

[0025] S133, the Resnik similarity is combined with the most specific common ancestor to perform semantic similarity comparison, and the association degree between the phenotype and the gene is scored based on the similarity comparison result to obtain a phenotype and gene association degree scoring result.

[0026] Further, the pre-acquired gene variation database is annotated, and the pre-acquired gene mutation data is scored based on the gene variation annotation result, including:

[0027] S21, annotating the pre-acquired gene variation database to obtain a gene variation annotation result;

[0028] The pre-acquired gene variation database includes: a variation harmfulness database, a population frequency database, a human gene mutation database, a clinical variation database, and a variation classification system database;

[0029] The gene variation annotation result includes: a variation harmfulness annotation result, a gene frequency annotation result, and a variation classification annotation result;

[0030] S22. Based on the gene variation annotation results, use the weighted summation method and preset scoring rules to score the gene variation in the pre-acquired gene variation data and obtain the gene variation scoring results.

[0031] Furthermore, the gene mutation scoring results include: mutation harmfulness scoring results, mutation incidence scoring results, mutation reporting scoring results, and mutation pathogenicity scoring results.

[0032] Furthermore, based on the phenotypic-gene association score and the gene variation score, a supervised learning model and a weighted summation method are used to perform a multi-dimensional pathogenicity comprehensive scoring evaluation on the pre-acquired gene mutation data, resulting in a pathogenic mutation ranking score, including:

[0033] S31. Based on the scoring results of the degree of association between phenotype and gene and the scoring results of gene variation, the weights of the comprehensive scoring assessment of multi-dimensional pathogenicity are set.

[0034] S32. Utilize the weight training mechanism of ranking learning combined with pre-acquired gene mutation data to optimize the set weights and obtain the weight optimization results.

[0035] S33. Based on the weight optimization results, the pre-acquired gene mutation data is evaluated using a weighted summation method to achieve a multi-dimensional comprehensive pathogenicity score, and the pathogenic mutation ranking score is obtained.

[0036] Furthermore, the weights are optimized using a ranking learning-based weight training mechanism combined with pre-acquired gene mutation data. The optimized weights include:

[0037] S321. Training sample pairs are constructed using pre-acquired gene mutation data to obtain several positive and negative sample pairs with pathogenic sites.

[0038] S322. Construct a pairwise ranking loss function using positive and negative sample pairs with pathogenic sites, and use the pairwise ranking loss function as the training target of the weight training mechanism for ranking learning.

[0039] S323. Based on the pairwise sorting loss function, the optimizer is used to optimize and iteratively update the set weights to obtain the weight optimization results.

[0040] Furthermore, the expression for the pairwise sorting loss function is as follows:

[0041] ;

[0042] In the formula, This represents the pairwise sorting loss function; Indicates the first i Pathogenic variants in each sample; It is the firsti a combined score of pathogenic variants in the sample; is the score of benign variants paired with pathogenic variants in the sample; i is the score of benign variants paired with pathogenic variants in the sample; denotes benign variants paired with pathogenic variants; β denotes hyperparameters that control the steepness of the loss surface; N denotes the total number of positive and negative sample pairs.

[0043] According to another aspect of the present application, there is provided a scoring and ranking system for pathogenic mutations of monogenic genetic diseases, comprising a phenotype and gene association scoring module, a gene variant scoring module, and a pathogenicity comprehensive evaluation module.

[0044] The phenotype and gene association scoring module is configured to extract standardized phenotype terms from pre-acquired symptom descriptions based on pre-acquired phenotype data and literature abstracts using a hybrid strategy, and score the association degree between the standardized phenotype terms and pre-acquired gene mutation data using a similarity comparison method.

[0045] The gene variant scoring module is configured to annotate pre-acquired gene variant databases, and score pre-acquired gene mutation data based on the gene variant annotation results.

[0046] The pathogenicity comprehensive evaluation module is configured to use a supervised learning model and a weighted summation method to comprehensively score and evaluate the pathogenicity of pre-acquired gene mutation data in multiple dimensions based on the phenotype and gene association scoring results and the gene variant scoring results, to obtain a scoring and ranking result of pathogenic mutations.

[0047] The present application has the following advantages:

[0048] The present application extracts standardized phenotype terms (HPO, Human Phenotype Ontology) from pre-acquired symptom descriptions, combines multi-dimensional annotation and scoring of gene variants, and recommends the most likely pathogenic mutations from massive genetic disease detection variant data (i.e., pre-acquired gene mutation data), so that the scoring and ranking of genetic disease pathogenic mutations has the advantages of high detection recall rate and high automation, and solves the problems of single consideration factor, low automation, and poor recall rate in current clinical methods. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0050] Figure 1 This is a flowchart of a scoring and ranking method for pathogenic mutations in a single-gene hereditary disease according to an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of a scoring and ranking system for pathogenic mutations in a single-gene hereditary disease according to an embodiment of the present invention.

[0052] Figure 3 This is a flowchart of the deep learning training process for scoring the degree of association between phenotype and gene in a scoring and ranking method for pathogenic mutations of a single-gene hereditary disease according to an embodiment of the present invention.

[0053] Figure 4 This is a schematic diagram illustrating the identification of standardized HPO phenotype terms from unstructured clinical symptom descriptions in a scoring and ranking method for pathogenic mutations in single-gene hereditary diseases according to an embodiment of the present invention.

[0054] Figure 5 This is a schematic diagram of the scoring principle in a scoring and ranking method for pathogenic mutations in a single-gene hereditary disease according to an embodiment of the present invention.

[0055] In the picture:

[0056] 1. Phenotypic and gene association scoring module; 2. Gene variation scoring module; 3. Pathogenicity comprehensive assessment module. Detailed Implementation

[0057] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0058] According to embodiments of the present invention, a scoring and ranking method and system for pathogenic mutations in single-gene hereditary diseases are provided.

[0059] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a scoring and ranking method for pathogenic mutations in single-gene hereditary diseases is provided, the method comprising the following steps:

[0060] S1. Based on pre-acquired phenotypic data and literature abstracts, a hybrid strategy is used to extract standardized phenotypic terms from the pre-acquired symptom descriptions, and a similarity comparison method is used to score the degree of phenotypic-gene association between the standardized phenotypic terms and the pre-acquired gene mutation data.

[0061] Specifically, based on the pre-acquired phenotype data and literature abstracts, the pre-acquired symptom descriptions are standardized phenotype term extraction using a hybrid strategy, and the standardized phenotype term and the pre-acquired gene mutation data are associated with the degree of scoring by similarity comparison method, including:

[0062] S11, based on the pre-acquired phenotype data and literature abstracts, the pre-acquired symptom descriptions are standardized phenotype term extraction using a hybrid strategy, and the standardized phenotype term and the pre-acquired gene mutation data are associated with the degree of scoring by similarity comparison method, including:

[0063] Specifically, based on the pre-acquired phenotype data and literature abstracts, the pre-acquired symptom descriptions are standardized phenotype term extraction using a hybrid strategy, and the standardized phenotype term and the pre-acquired gene mutation data are associated with the degree of scoring by similarity comparison method, including:

[0064] S111, based on the pre-acquired phenotype data, the clinical symptom descriptions are standardized phenotype term extraction using a hybrid strategy, and the standardized phenotype term and the pre-acquired gene mutation data are associated with the degree of scoring by similarity comparison method, including:

[0065] S112, the pre-acquired literature abstracts and the pre-acquired phenotype data are standardized phenotype term extraction using a hybrid strategy, and the standardized phenotype term and the pre-acquired gene mutation data are associated with the degree of scoring by similarity comparison method, including:

[0066] S113, the pre-acquired literature abstracts and the pre-acquired phenotype data are standardized phenotype term extraction using a hybrid strategy, and the standardized phenotype term and the pre-acquired gene mutation data are associated with the degree of scoring by similarity comparison method, including:

[0067] S12, the pre-acquired literature abstracts and the pre-acquired phenotype data are standardized phenotype term extraction using a hybrid strategy, and the standardized phenotype term and the pre-acquired gene mutation data are associated with the degree of scoring by similarity comparison method, including:

[0068] S13, the pre-acquired literature abstracts and the pre-acquired phenotype data are standardized phenotype term extraction using a hybrid strategy, and the standardized phenotype term and the pre-acquired gene mutation data are associated with the degree of scoring by similarity comparison method, including:

[0069] Specifically, the pre-acquired literature abstracts and the pre-acquired phenotype data are standardized phenotype term extraction using a hybrid strategy, and the standardized phenotype term and the pre-acquired gene mutation data are associated with the degree of scoring by similarity comparison method, including:

[0070] S131, the pre-acquired literature abstracts and the pre-acquired phenotype data are standardized phenotype term extraction using a hybrid strategy, and the standardized phenotype term and the pre-acquired gene mutation data are associated with the degree of scoring by similarity comparison method, including:

[0071] S132, statistics the frequency of each phenotype and gene according to the phenotype and gene association set, and finds the most specific common ancestor based on the frequency of each phenotype and gene;

[0072] S133, semantic similarity comparison is performed by using Resnik similarity combined with the most specific common ancestor, and scoring of the association degree of the phenotype and the gene is performed based on the comparison result of the similarity, to obtain the scoring result of the association degree of the phenotype and the gene.

[0073] Specifically, in actual application, the whole exome sequencing (WES) experimental process and library on-machine sequencing include the following steps: (a) sample collection and DNA extraction, specifically, collecting peripheral blood (EDTA tube), oral swab, amniotic fluid or tissue sample, using Eppendorf centrifuge (5810R and 5427R, Germany); centrifuging at 4°C for 10 min at 1600g, taking the supernatant, centrifuging at 16000g for 10 min, and then taking the supernatant, i.e. blood plasma. gDNA extraction, specifically, using TIANamp Blood DNA Kit (TIANGEN) or QIAamp DNA Blood Mini Kit (QIAGEN) for conventional extraction. (b) DNA quality and concentration detection. Concentration detection, specifically, using Qubit 3.0 fluorometer (Thermo, USA). Fragment size detection, specifically, using Nanodrop or Agilent TapeStation / Fragment Analyzer to detect integrity. (c) WES library construction, specifically, library construction kit, i.e. selecting Twist Bioscience Human Core Exome Kit. DNA starting amount, specifically, gDNA starting amount ≥ 100-200 ng. End repair & A tail addition (End Repair / A-Tailing); connecting adapter (Illumina double-end adapter); library purification using AMPure XP beads (Beckman Coulter); exon capture, specifically, hybridization capture after mixing the library with probes; target region enrichment after hybridization by magnetic beads; PCR amplification, specifically, amplifying the library to obtain sufficient product. Library quality control, specifically, KAPA Library Quant Kit (Roche) for qPCR quantification; library fragment size, specifically, Fragment Analyzer (Agilent) for detection. (d) on-machine sequencing platform, specifically, Illumina NovaSeq 6000.

[0074] Specifically, in actual application, the analysis of the data from the whole exome sequencing (WES) machine includes the following steps: (a) the BCL file from the sequencing machine is converted into a fastq file after being recognized by the bcl2fastq software; (b) the low-quality reads, the reads containing more than 5% of N bases, and the reads shorter than 50 bp in length are removed from the sequencing data by using the cutadapt software; (c) the reads are aligned to the hg19 reference genome by using the bwa software, and the PCR redundant sequences are removed; (d) the DNA fragments with low alignment quality, unaligned, and double-end reads that fail to perfectly pair are removed by using the samtools. The filtered DNA fragments are sorted according to the alignment positions; (e) the Recalibration algorithm of the GATK software package is used to correct the instrument bias and errors specific to the illumina sequencer; (f) INDEL region realignment: for the hotspot regions of frequent insertion (Insertion) and deletion (deletion) mutations, the debrujin graph is constructed, local assembly is performed, and false positive mutations are reduced by using the IndelRealigner algorithm of the GATK software package; (g) the SNV and INDEL mutation detection is performed by using the HaplotypeCaller algorithm of the GATK, and a VCF file is generated.

[0075] Specifically, in order to solve the problems that most of the current HPO-based phenotype recognition methods rely on: dictionary-based - although the precision is high, but the recall rate is low; or supervised machine learning model - a large amount of manually annotated data is needed, and it is difficult to expand, etc., the present application uses a hybrid approach to combine the dictionary method with an unsupervised / weakly supervised deep learning model, and constructs an automatic HPO concept recognition, which can realize efficient phenotype extraction without manual annotation of the training set.

[0076] Specifically, first, extract all HPO terms and their synonyms, definitions; filter ambiguous abbreviations (e.g. ASD can represent Atrial Septal Defect or Autism Spectrum Disorder). Second, build a distant supervision training set to match pubmed abstracts (i.e. literature abstracts) (totaling 27 million) with HPO terms to automatically generate positive samples; randomly sample n-grams that do not match as negative samples; use BioBERT for multi-classification task, input n-gram, output HPO ID. Use Trie tree structure to accelerate the search process. Third, use BioBERT model for training using pre-trained BioBERT model (based on PubMed training); classify each candidate phrase; set probability threshold to filter the final recognition result. Fourth, integrate dictionary and deep learning results merge the recognition results of both; introduce POS tag filter to remove meaningless phrases; handle overlapping concepts to improve recognition coverage.

[0077] wherein the degree of association between phenotype and gene is scored, the dictionary method and unsupervised / weakly supervised deep learning method are combined for efficient phenotype extraction, and the matching degree of patient phenotype and known disease is judged based on the semantic similarity calculation between HPO terms, and the similarity between HPOs is evaluated through MICA (maximum information common ancestor) algorithm. The calculation method of the dictionary-based method is as follows:

[0078] ;

[0079] wherein, h i represents a standardized phenotype term in HPO; Match( h i , T ) represents the success of exact or fuzzy matching of the term in the text; T represents the input text. The BioBERT-based deep learning method can be quantified as:

[0080] ;

[0081] wherein, w 1... w n is the word sequence in the text; is the probability of a certain HPO concept h j under the given context word sequence, then BioBERT is used to classify each candidate n-gram, and the probability score of each HPO concept is output. Finally, the combined method is for any two candidate HPO concepts ha and h b If they have overlap in the text (e.g. share a word), the following rules are adopted, specifically if h a and h b No overlap -> keep both; if h a and h b Share ID -> keep the highest scorer; if h a and h b Start and end positions are the same but ID is different -> keep the highest scorer; if h a and h b Start and end positions are different and ID is different -> keep both. The final integrated score can be quantified as:

[0082] ;

[0083] where, S dict ( h j ) is the matching score of the dictionary method for HPO concepts h j (0 or 1); S BioBERT ( h j ) is the softmax probability value output by the BioBERT model; α ∈ [0, 1] is a hyperparameter to balance the weights of the two models.

[0084] The detailed HPO standardized named entity (NER) extraction process diagram is shown in Figure 3 . Figure 3 The HPO dictionary is obtained from the HPO term name and synonym in the dictionary construction; the dictionary result is obtained by combining the patient clinical description dictionary construction term name and synonym; the phenotype named entity recognition (phenotype-NER) is performed by using the pubmed literature to train the BERT deep learning model and combining the patient clinical description, and the deep learning result is obtained; the standardized HPO is obtained by merging the HPO phenotype based on the dictionary result and the deep learning result.

[0085] In a real case, the original unstructured clinical description diagnosis provided by the doctor is used to extract the HPO, and the diagram is shown in Figure 4 . Figure 4Clinical description: Bilateral inherited retinal degeneration, both eyes. Visual acuity decreased 8 years ago without obvious inducement. Foveal reflex disappeared in the macula. Extensive retinal peripheral pigmentation. Complaints of leukopenia and thrombocytopenia for 7-8 years. Leukopenia and thrombocytopenia. Leukopenia and thrombocytopenia.

[0086] Among them, the extracted standardized phenotype terms are: retinal degeneration, HP:0000546; visual loss, HP:0000572; extensive retinal peripheral pigmentation, HP:0001106; leukopenia, HP:0001882 leukopenia, HP:0001882; thrombocytopenia, HP:0001873; thrombocytopenia, HP:0001873.

[0087] Finally, combined with the standardized HPO ID of the patient and the VCF mutation file, the correlation degree of any gene mutation and the clinical phenotype of the patient is calculated and weighted.

[0088] The HPO term list and the known HPO terms of each disease or gene in the database are compared in terms of semantic similarity, using the OWLsim semantic algorithm, which can be quantified as:

[0089] ;

[0090] In the formula, H patient HPO set of a given patient, HPO set associated with each gene-related disease (through OMIM / Orphanet) H gene Resnik similarity + MICA (Maximum Information Content Ancestor) as the core, for any two HPO terms h i ∈ H patient and h j ∈ H gene Their similarity is the information content of their most specific common ancestor, which is quantified as:

[0091] ;

[0092] In the formula, P ( h ) is the frequency of HPO term h in the entire disease database. The final calculated phenotype score ranges from 0 to 1, indicating that the higher the patient's phenotype matches the gene-related disease.

[0093] S2, gene variation annotation is performed on the pre-acquired gene variation database, and gene variation scoring is performed on the pre-acquired gene mutation data according to the gene variation annotation result.

[0094] Specifically, the gene variation annotation is performed on the pre-acquired gene variation database, and the gene variation scoring is performed on the pre-acquired gene mutation data according to the gene variation annotation result includes:

[0095] S21, gene variation annotation is performed on the pre-acquired gene variation database to obtain a gene variation annotation result;

[0096] The pre-acquired gene variation database includes a variation harmfulness database, a population frequency database, a human gene mutation database, a clinical variation database, and a variation classification system database.

[0097] The gene variation annotation result includes a variation harmfulness annotation result, a gene frequency annotation result, and a variation classification annotation result.

[0098] S22, gene variation scoring is performed on the pre-acquired gene mutation data according to the gene variation annotation result using a weighted summation method and a pre-set scoring rule to obtain a gene variation scoring result.

[0099] Specifically, the gene variation scoring result includes a variation harmfulness scoring result, a variation incidence scoring result, a variation report scoring result, and a variation pathogenicity scoring result.

[0100] Specifically, in the field of NGS (high-throughput sequencing), determining whether a mutation is deleterious or pathogenic is one of the key steps in variant annotation. Mutation-based deleteriousness prediction software mainly assesses the potential impact of mutations on protein function through sequence conservation, structural impact, evolutionary information, or machine learning methods. The following are commonly used deleteriousness prediction tools and databases widely used in scientific research and clinical practice: SIFT (Sorting Intolerant From Tolerant) determines whether a mutation affects protein function based on sequence homology and amino acid conservation. PolyPhen-2 (Polymorphism Phenotyping v2) combines sequence, structure, and protein functional domain information to predict the impact of amino acid substitution on protein function. MutationTaster integrates multiple dimensions of information such as sequence, conservation, protein function, and splice site changes for classification. CADD (Combined Annotation Dependent Depletion) is based on a machine learning model that integrates dozens of annotation features (conservation, epigenetic modification, regulatory information, etc.) to assess the deleteriousness of mutations. REVEL (Rare Exome Variant Ensemble Learner) integrates the results of multiple existing prediction tools (including SIFT, PolyPhen-2, MutationAssessor, etc.) and uses ensemble learning methods to train the final prediction. Alphamissense (developed by DeepMind) is based on AlphaFold protein structure + deep learning model to predict the impact of all possible missense mutations on protein function. MetaSVM / MetaLR is based on machine learning (SVM or logistic regression) and integrates the outputs of SIFT, PolyPhen, LRT, MutationTaster, etc. SpliceAI is based on deep neural networks and learns the upstream and downstream 10kbp sequences to predict whether a mutation affects normal splicing (gain / loss). MaxEntScan is based on the maximum entropy model and calculates the base combination probability around the standard donor / acceptor splice site (+ / - 3~6bp). dbscSNV (Database of Splicing Consensus SNVs) provides two machine learning model scores (AdaBoost, RandomForest) for each splice-related SNV.

[0101] The present application constructs an automatic weighted scoring by Ensemble-based pathogenicity prediction, combines the results of multiple existing variant harmfulness prediction tools, and comprehensively scores each candidate mutation, thereby improving the accuracy of identifying true pathogenic mutations. The Ensemble-based pathogenicity prediction can be used as an auxiliary decision-making tool in clinical diagnosis, genetic disease screening, and cancer mutation analysis. Specifically, in the present application, for the pathogenicity prediction of a single variant, a quantitative scoring system based on multiple mainstream variant function annotation tools is constructed. Specifically, the output results of multiple independent variant prediction algorithms or databases are introduced, and the results are discretized into two categories of "harmful" and "harmless". According to the preset threshold, each score is weighted and summarized, and finally the comprehensive harmfulness score of the variant is generated.

[0102] Specifically, the variant harmfulness scoring is an Ensemble-based pathogenicity prediction, which builds an automatic weighted scoring system to combine the results of multiple existing variant harmfulness prediction tools to score each candidate mutation comprehensively, thereby improving the accuracy of identifying true pathogenic mutations. For example, the scoring rules of each tool are as follows: if the prediction result of SIFT is “Damaging”, it is scored as 1, and if it is “Tolerated”, it is scored as 0; if the prediction of PolyPhen-2 is “Probably Damaging”, it is scored as 1, and if it is “Possibly Damaging” or “Benign”, it is scored as 0; if the score of REVEL is less than 0.5, it is scored as 1, and if it is greater than or equal to 0.5, it is scored as 0; if the annotation of MutationTaster is “Disease causing”, it is scored as 1, and if it is “Polymorphism”, it is scored as 0; if CADD (phred-scaled score) is less than 15, it is scored as 1, and if it is greater than or equal to 15, it is scored as 0; if the score of AlphaMissense is greater than or equal to 0.8, it is scored as 1, and if it is less than 0.8, it is scored as 0; if the score of M-CAP is less than 0.025, it is scored as 1, and if it is greater than or equal to 0.025, it is scored as 0; if the score of MetaSVM and MetaLR is less than 0.5, it is scored as 1, and if it is greater than or equal to 0.5, it is scored as 0; if the score of SpliceAI is less than 0.2, it is scored as 1, and if it is greater than or equal to 0.2, it is scored as 0; if the model prediction of MaxEntScan indicates that the variant has a significant impact on the splicing signal, it is scored as 1, and if it has no significant impact, it is scored as 0; if the score of dbscSNV AdaBoost or RandomForest is greater than or equal to 0.6, it is scored as 1, and if it is less than 0.6, it is scored as 0. After the results of each tool are converted into binary scores under a unified standard, they are added together to serve as the basis for variant pathogenicity prediction, which can be further used to rank candidate mutation sites or combined with phenotype association scores for joint analysis. This method maintains biological interpretability while having strong scalability and compatibility. The calculation method of variant harmfulness scoring is as follows:

[0103] ;

[0104] wherein, k i is the weight of each tool (for example: SIFT: 1.0, REVEL: 1.2, AlphaMissense: 1.5); f i ( x) is binary output (0 or 1); the final score can be a floating point number, specifically Total_Score≥2.5: highly suspicious pathogenic mutation, Total_Score1.5~2.5: moderate risk mutation, Total_Score<1.5: probably harmless mutation. Therefore, it can be used to quantify the harmfulness of all detected mutations for later screening.

[0105] In NGS analysis, the population frequency of gene variation is one of the important bases for identifying potential pathogenic variations. The variation frequency analysis method used in this system is based on multiple authoritative public databases that integrate large-scale genomic data from different ethnic groups around the world, widely used in genetic disease research, cancer mutation analysis and clinical molecular diagnosis fields. The main population frequency database used in this invention and its functions include gnomAD (Genome Aggregation Database), which is maintained by Broad Institute and integrates more than 150,000 whole-exome and whole-genome sequencing data, covering multiple ethnic groups around the world. The 1000 Genomes Project is one of the earliest large-scale human genome databases, containing WGS data from about 2500 individuals from 26 human populations around the world. ExAC (Exome Aggregation Consortium) is the predecessor of gnomAD, mainly focusing on the frequency statistics of variations in the exon region. And the internal constructed normal MAF database, specifically including the MAF information of 3000 healthy subjects. The variation occurrence rate scoring is a harmfulness scoring function based on population frequency, used for automatic variation screening, where the lower the frequency → the more likely it is a pathogenic variation → the higher the score, the higher the frequency → the more likely it is a benign polymorphism → the lower the score. Specifically, this invention defines a "variation population occurrence rate harmfulness score" (Population Frequency Pathogenicity Score, PFPS), which is calculated as follows:

[0106] ;

[0107] In the formula, f : is the maximum population frequency of the variation in gnomAD, 1000G, etc. (such as GMAF: Global Minor Allele Frequency). f max : is a threshold value of 0.001 (0.1%), above which the variation is considered benign, otherwise pathogenic. ε : is, ε =1 e -6A very small constant to avoid division by zero error. The scoring mechanism of the present invention can effectively identify low-frequency or rare variants, thereby improving the recall and accuracy of pathogenic mutations.

[0108] Variant case report scoring (i.e., variant report scoring), the present invention further introduces internationally recognized variant pathogenicity databases, including Human Gene Mutation Database (HGMD), ClinVar, and Leiden Open Variation Database (LOVD), to assist in judging the pathogenicity of candidate variants. The databases provide different levels of pathogenicity classification labels, such as “DM”, “FP”, “R” in HGMD, “Pathogenic”, “Benign” in ClinVar, and “DM”, “DP” in LOVD, etc. The present invention adopts a weighted scoring mechanism to quantitatively score the variant pathogenicity labels provided by each database, and calculates the overall pathogenicity score of the variant through linear combination. Specifically, in the HGMD database, if the variant is labeled as “DM” or “FP”, it is identified as a clear pathogenic variant, and 1 point is added; if it is “DM” or “R”, 0.5 points and 0.2 points are added respectively; in the ClinVar database, if the variant is labeled as “Pathogenic” or “Likely Pathogenic”, 1 point and 0.8 points are added respectively; if it is “Uncertain Significance”, 0.2 points are added; and the rest are not added; in the LOVD database, if the variant is labeled as “DM”, 1 point is added; if it is “DP” or “DFP”, 0.1 points are added; and the rest are not added. Then, the scores of each database are weighted and summed according to the preset weights (HGMD: 0.4, ClinVar: 0.4, LOVD: 0.2) to obtain the comprehensive score of the pathogenicity of the variant (Multi-Database Pathogenicity Score, MDPS), which can be calculated as follows:

[0109] ;

[0110] In the formula, S HGMD , S ClinVar , S LOVD , respectively, represent the pathogenicity score of the variant in these three databases; a 1, a 2, a 3 is the weight set for each database, which can be adjusted according to its authority and data integrity (optionally, a 1=0.4, a2 = 0.4, a 3 = 0.2). This scoring mechanism can effectively improve the accuracy and recall rate of pathogenic mutation recognition, especially for clinical screening of single gene genetic diseases and cancer mutations.

[0111] Variation pathogenicity scoring, the present application further introduces the internationally recognized gene variation pathogenicity classification standard - American College of Medical Genetics and Genomics (ACMG) variation classification system, which is used for automatic scoring and sorting of the pathogenicity of candidate variations. The ACMG standard provides multi-dimensional pathogenicity evidence items, including pathogenic support items (such as PM1, PM2, PP1, PP3) and benign support items (such as BA1, BS1, BP4, BP7), and different weights are given according to the evidence strength. Finally, ACMG forms a pathogenicity classification result, including: Pathogenic (P) - clear pathogenicity, this category indicates that the variation has been fully supported by sufficient evidence and has a clear causal relationship with a certain genetic disease. Likely Pathogenic (LP) - very likely pathogenic, this category indicates that the variation has a strong pathogenic possibility, but has not yet reached the standard of "clear pathogenicity". Uncertain Significance (VUS) - significance unknown, this category indicates that the available data is not enough to determine whether the variation is pathogenic or benign. May lack sufficient experimental data, uncertain population frequency information, missing family information, or inconsistent results from multiple prediction algorithms. Likely Benign (LB) - very likely benign, this category indicates that the variation is likely not pathogenic, but has not completely ruled out all pathogenic possibilities. Usually based on benign-related ACMG items (such as BS1, BP1-BP7), or based on higher frequency observations from large-scale population databases (such as gnomAD, ExAC). Although not absolutely benign, the pathogenic probability is very low. Benign (B) - clear benign, this category indicates that the variation has been confirmed by a large amount of evidence not to cause disease. Usually appears in very high population frequency (such as > 5%), widely exists in healthy population, and has been classified as benign polymorphism by multiple studies or expert groups. The present application proposes an automatic pathogenicity scoring method based on the American College of Medical Genetics and Genomics (ACMG) variation pathogenicity classification standard. By assigning different numerical weights to the ACMG classification labels (such as Pathogenic, Likely Pathogenic, Uncertain Significance, Likely Benign and Benign) corresponding to the candidate variation, a highly interpretable and highly standardized pathogenicity comprehensive score function is constructed. Specifically, if the variation Lis classified as Pathogenic, the highest pathogenicity score (e.g. 3) is assigned; if as Likely Pathogenic, the second highest score (e.g. 2) is assigned; if as Uncertain Significance (VUS), a neutral score (e.g. 1) is assigned; if as Likely Benign or Benign, a lower or zero score (e.g. 0.5 and 0) is assigned according to the strength of benign evidence, respectively. Through the above weighting mechanism, the present application can quickly and accurately sort the pathogenicity of the variants detected by large-scale NGS without human intervention, and realize multi-dimensional evaluation and priority sorting of the pathogenicity of the variants in combination with other dimensional annotation information (such as population frequency, splicing impact, structure prediction, etc.). This method significantly improves the automation level and diagnosis efficiency in clinical genetic testing, and is especially suitable for single gene disease, tumor mutation screening and personalized medicine guidance scenarios. Specifically, the present application defines a mapping function f ( L ) to convert each ACMG label into a real number score representing the pathogenicity of the variant. The calculation method of the variant pathogenicity scoring is as follows:

[0112] ;

[0113] In the formula, P: Pathogenic, LP: Likely Pathogenic, VUS: Variant of Uncertain Significance, LB: Likely Benign, B: Benign.

[0114] S3, according to the phenotype and gene association degree scoring result and the gene variant scoring result, using a supervised learning model and a weighted summation method to perform multi-dimensional pathogenicity comprehensive scoring evaluation on the pre-acquired gene mutation data, to obtain a pathogenic mutation ranking scoring result.

[0115] Specifically, according to the phenotype and gene association degree scoring result and the gene variant scoring result, using a supervised learning model and a weighted summation method to perform multi-dimensional pathogenicity comprehensive scoring evaluation on the pre-acquired gene mutation data, to obtain a pathogenic mutation ranking scoring result, including:

[0116] S31, based on the phenotype and gene association degree scoring result and the gene variant scoring result, setting the weight of the multi-dimensional pathogenicity comprehensive scoring evaluation;

[0117] S32, using the weight training mechanism of the ranking learning to combine the pre-acquired gene mutation data to optimize the weight, to obtain a weight optimization result.

[0118] Specifically, the weights are optimized using a ranking learning-based weight training mechanism combined with pre-acquired gene mutation data. The optimized weights include:

[0119] S321. Training sample pairs are constructed using pre-acquired gene mutation data to obtain several positive and negative sample pairs with pathogenic sites.

[0120] S322. Construct a pairwise ranking loss function using positive and negative sample pairs with pathogenic sites, and use the pairwise ranking loss function as the training target of the weight training mechanism for ranking learning.

[0121] S323. Based on the pairwise sorting loss function, the optimizer is used to optimize and iteratively update the set weights to obtain the weight optimization results.

[0122] Specifically, the expression for the pairwise sorting loss function is:

[0123] ;

[0124] In the formula, This represents the pairwise sorting loss function; Indicates the first i Pathogenic variants in each sample; It is the first i The overall score of pathogenic variants in each sample; It is the first i Benign variant scores paired with pathogenic variants in each sample; Indicates a benign variant paired with a pathogenic variant; β The hyperparameters that indicate the steepness of the loss surface; N This represents the total number of positive and negative sample pairs.

[0125] S33. Based on the weight optimization results, the pre-acquired gene mutation data is evaluated using a weighted summation method to achieve a multi-dimensional comprehensive pathogenicity score, and the pathogenic mutation ranking score is obtained.

[0126] Specifically, the pathogenic mutation ranking score output by the weighted score of multi-dimensional pathogenicity assessment (i.e., multi-dimensional pathogenicity comprehensive score assessment) is used to determine the phenotypic association score (i.e., the score of the degree of association between the phenotype and the gene). S phenotype ), the score for assessing the impact of variant function (i.e., the score for the harmfulness of the variant). S functional ), the variation population frequency analysis score (i.e., the obtained variation incidence score) S population ), multi-database pathogenicity score (i.e., the reported variant score) S database) and ACMG evidence item mapping score (i.e. the resulting variant pathogenicity scoring result S acmg ); phenotype association score, i.e. using hybrid natural language processing techniques such as BERT to extract standardized human phenotype ontology (HPO) terms from patient free-text descriptions and calculate the semantic matching score between them and known genetic disease phenotypes to quantify the clinical relevance between variants and phenotypes; variant functional impact assessment, i.e. based on multiple existing predictions (such as SIFT, PolyPhen-2, REVEL, AlphaMissense, etc.) to comprehensively judge the functional impact of candidate variants, using ensemble learning to weight and sum the output results of each algorithm to form a unified functional deleteriousness score; variant population frequency analysis, i.e. based on the variant frequency information provided by large-scale population genomic databases (such as gnomAD, 1000 Genomes, ExAC), a deleteriousness scoring function based on logarithmic transformation is designed, so that low-frequency or unreported variants get higher scores, and high-frequency benign variants get lower scores; multi-database pathogenicity score, i.e. integrating the annotation information about whether the variant is included as pathogenic in authoritative databases such as HGMD, ClinVar and LOVD, assigning different weights according to their inclusion levels and linearly combining them; ACMG evidence item mapping, i.e. converting the pathogenicity evidence items (such as PM, PP, BS, BP) specified by the American College of Medical Genetics and Genomics (ACMG) into quantitative scores, and introducing a weighting mechanism to reflect the evidence strength of different items. By combining multiple independent scores, covering multiple dimensions such as clinical phenotype matching, variant functional impact, population frequency distribution, database inclusion, and ACMG pathogenicity evidence, a highly interpretable and highly automated variant pathogenicity recognition system is constructed. The output results of the above five are fused by weighted linear combination, and finally a unified variant pathogenicity comprehensive score is obtained S pathogenicity :

[0127] ;

[0128] The comprehensive pathogenicity score can effectively improve the recognition accuracy of mutations related to single gene genetic diseases, and is especially suitable for clinical genetic diagnosis, personalized medication guidance and research-level variant screening tasks.

[0129] The weights in the pathogenic mutation ranking and scoring output of the multi-dimensional pathogenicity assessment weighted score w 1, w 2,…, w5Can be set according to the actual application scene, for example, using artificial experience assignment method or through supervised learning model optimization training. In particular, the weight training mechanism based on Learning to Rank (LTR) is provided, which is used to optimize each weight in the comprehensive pathogenicity scoring function (such as phenotype association score, variant functional impact score, variant population frequency score, multi-database pathogenicity score, ACMG evidence item mapping score, etc., that is, each weight w 1, w 2,…, w 5). The purpose is to make the known pathogenic variants be ranked first or as close to the front as possible in the candidate variant list through supervised learning. Specifically, the training process includes the following steps: 1) construct a training sample pair, that is, for each patient sample, extract all candidate variants thereof, and combine the pathogenic variant with the rest of the benign variants two by two to form multiple positive and negative sample pairs; 2) define the scoring function, that is, let the comprehensive score of the variant V be:

[0130] ;

[0131] 3) define the ranking loss function, that is, use Pairwise Ranking Loss as the training target. Specifically, now for each sample, there is 1 pathogenic site and multiple benign / non-pathogenic variants, for each pair (pathogenic site V p , benign site V q ), we hope: S ( V p )> S ( V q ), so the loss function can be constructed, and its formula is:

[0132] ;

[0133] In the formula, is the pathogenic variant in the i-th sample; i is the comprehensive score of the pathogenic variant in the i-th sample; is the score of the benign variant paired with the pathogenic variant in the i-th sample; i is the benign variant paired with it; is a hyperparameter that controls the steepness of the loss surface (in particular, it is 1 in specific applications); and i is a hyperparameter that controls the steepness of the loss surface (in particular, it is 1 in specific applications); and is a hyperparameter that controls the steepness of the loss surface (in particular, it is 1 in specific applications); and β is a hyperparameter that controls the steepness of the loss surface (in particular, it is 1 in specific applications); and N is a hyperparameter that controls the steepness of the loss surface (in particular, it is 1 in specific applications); andis the total number of positive and negative sample pairs. This loss function encourages pathogenic variant scores to be higher than benign variant scores. If the pathogenic variant score is greater than the benign variant, the loss approaches 0; otherwise, the loss increases. 4) Weight optimization and iterative update, i.e. using Adam optimizer, gradually adjusting the weights of each module by minimizing the above ranking loss function, and finally making pathogenic variants get higher ranking priority in the comprehensive score. This training mechanism does not require manual annotation of training data, but only relies on the existing pathogenic variants and benign variants in their context to complete model training, which has high automation and scalability. After loss function optimization, the optimal weight vector w =[0.30,0.25,0.15,0.15,0.15], so that the pathogenic variant recall rate reaches the expected effect.

[0134] Embodiment one:

[0135] Specifically, in the present application, 133 cases of known pathogenic sites and their rankings of hypercholesterolemia data, 56 cases of known pathogenic sites and their rankings of tuberous sclerosis data, and 99 cases of known pathogenic sites and their rankings of phenylketonuria data, i.e. test set 1 shown in Table 1.

[0136] Table 1 Basic information of clinical samples:

[0137]

[0138] So that the pathogenic variant recall rate of test set 1 reaches the expected effect as shown in Table 2.

[0139] Table 2 Comparison of recall rates of default weights and optimized weights of the present application:

[0140]

[0141] Among them, Recall@5 is the proportion of pathogenic variants ranked in the top 5 in each sample; Recall@10 is the proportion of pathogenic variants ranked in the top 10; Recall@20 is the proportion of pathogenic variants ranked in the top 20; 95% CI (confidence interval) is estimated by Wilson Score method, reflecting statistical uncertainty; the present application (optimized) represents the optimal weight combination obtained by the ranking learning method (i.e. the above PairwiseRanking Loss method). In particular, in this case w =[0.30,0.25,0.15,0.15,0.15]). The experimental results show that when using the optimized weight configuration w =[0.30,0.25,0.15,0.15,0.15]), the system achieves the best performance on all three indicators.

[0142] For example, in phenylketonuria, Recall@20 reached 97.6%, significantly higher than other settings.

[0143] The above results indicate that by setting weight parameters appropriately or through automated training, the accuracy and stability of pathogenicity identification of variants can be effectively improved, especially in large-scale NGS data analysis and clinical genetic diagnosis scenarios.

[0144] See the overall flowchart of the module of this invention. Figure 5 . Figure 5 The system extracts standardized phenotypic terms from clinical symptom descriptions, then performs a scoring module based on phenotype and genes; it annotates harmfulness using a variant harmfulness database, then performs a scoring module based on variant harmfulness; it annotates minor allele frequencies using a healthy population frequency database, then performs a scoring module based on variant incidence; it annotates minor allele frequencies using a diseased population frequency database, then performs a scoring module based on variant case reports; it annotates ACMG grading using an ACMG database, then performs a scoring module based on variant pathogenicity; and finally, it outputs a pathogenic mutation ranking score based on a multi-dimensional pathogenicity assessment weighted score from the above modules.

[0145] Example 2:

[0146] The clinical sample information used in the examples is shown in Table 3.

[0147] Table 3 Basic information of clinical samples:

[0148]

[0149] Test set 2 was used to compare the recall of the present invention with that of currently used clinical ranking and scoring software such as exomiser and phenolyzer (Recall@5 (95% CI) represents the recall of pathogenic sites within the top 5 and its 95% confidence interval).

[0150] Recall@10 (95% CI) represents the recall rate and its 95% confidence interval for pathogenic sites within the top 10, and Recall@20 (95% CI) represents the recall rate and its 95% confidence interval for pathogenic sites within the top 20. The comparison results are shown in Table 4.

[0151] Table 4 Comparison Results:

[0152]

[0153] As shown in Table 4, the overall recall rate of this invention is significantly better than that of currently used clinical ranking and scoring software such as Exomiser and Phenoyzer.

[0154] The comprehensive pathogenicity scoring system can effectively improve the recognition accuracy of mutations related to monogenic genetic diseases, and is especially suitable for clinical genetic diagnosis, personalized medication guidance and research-level variant screening tasks.

[0155] As shown in Figure 2 According to another embodiment of the present application, a scoring and ranking system for pathogenic mutations of monogenic genetic diseases is provided, which comprises a phenotype and gene association scoring module 1, a gene variation scoring module 2, and a comprehensive pathogenicity evaluation module 3.

[0156] The phenotype and gene association scoring module 1 is used to extract standardized phenotype terms from pre-acquired symptom descriptions based on pre-acquired phenotype data and literature abstracts using a hybrid strategy, and to score the association degree between standardized phenotype terms and pre-acquired gene mutation data through a similarity comparison method.

[0157] The gene variation scoring module 2 is used to annotate the pre-acquired gene variation database, and to score the pre-acquired gene mutation data based on the gene variation annotation results.

[0158] The comprehensive pathogenicity evaluation module 3 is used to perform multi-dimensional comprehensive scoring and evaluation of the pre-acquired gene mutation data based on the phenotype and gene association scoring results and the gene variation scoring results using a supervised learning model and a weighted summation method, to obtain the ranking and scoring results of pathogenic mutations.

[0159] In summary, the present application uses the combination of any one or more of the following six categories to automatically identify the most likely pathogenic mutation combination of a patient for the preparation of a product for detecting or assisting in the detection of Mendelian monogenic genetic diseases: (1) scoring based on the association degree between phenotype and gene; (2) scoring based on the harmfulness of the variation; (3) scoring based on the occurrence rate of the variation; (4) scoring based on the case report of the variation; (5) scoring based on the pathogenicity of the variation; (6) ranking and scoring output of pathogenic mutations based on the weighted score of multi-dimensional pathogenicity evaluation.

[0160] In summary, by means of the above technical solutions of the present application, the present application extracts standardized phenotype terms (HPO, Human Phenotype Ontology) from pre-acquired symptom descriptions, combines multi-dimensional annotation and scoring of gene variations, and recommends the most likely pathogenic mutations from massive genetic disease detection variation data (i.e. pre-acquired gene mutation data), so that the scoring and ranking of genetic disease pathogenic mutations have the advantages of high detection recall rate and high automation, solving the problem of single consideration factor, low automation and poor recall rate in current clinical methods.

[0161] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method of scoring and ranking pathogenic mutations of monogenic genetic diseases, characterized in that, The method comprises: S1, based on the pre-acquired phenotype data and literature abstracts, using a hybrid strategy to extract standardized phenotype terms from the pre-acquired symptom descriptions, and scoring the association degree between the standardized phenotype terms and the pre-acquired gene mutation data through a similarity comparison method; The S1 comprises: S11, based on the pre-acquired phenotype data and literature abstracts, using a dictionary method and a deep learning model to extract phenotype terms from the pre-acquired symptom descriptions to obtain a dictionary matching result and a phenotype term recognition result; The S11 comprises: S111, based on the pre-acquired phenotype data, using a dictionary method to match phenotype terms from the clinical symptom descriptions to obtain a dictionary matching result; S112, dividing the pre-acquired literature abstracts and the pre-acquired phenotype data into positive and negative samples for phenotype term matching, and constructing a supervised training set based on the positive and negative sample division results; S113, training a deep learning model using the supervised training set, and based on the trained deep learning model, combining a pre-set probability threshold to perform phenotype named entity recognition on the clinical symptom descriptions to obtain a phenotype recognition result; S12, using a weighted summation method to integrate the dictionary matching result and the phenotype term recognition result to obtain standardized phenotype terms; S13, performing semantic similarity comparison on the standardized phenotype terms and the pre-acquired gene mutation data through an OWL similarity semantics algorithm, and scoring the association degree between the phenotype and the gene based on the similarity comparison result to obtain a phenotype and gene association degree scoring result; The S13 comprises: S131, constructing a semantic association network for the standardized phenotype terms and the pre-acquired gene mutation data through an OWL similarity semantics algorithm to obtain a phenotype and gene association set; S132, according to the phenotype and gene association set, calculating the frequency of occurrence of each phenotype and gene, and based on the frequency of occurrence of each phenotype and gene, finding the most specific common ancestor; S133, using a Resnik similarity combined with the most specific common ancestor to perform semantic similarity comparison, and based on the similarity comparison result, scoring the association degree between the phenotype and the gene to obtain a phenotype and gene association degree scoring result; S2, annotating the pre-acquired gene variation database, and scoring the pre-acquired gene mutation data based on the gene variation annotation result; S3, according to the phenotype and gene association degree scoring result and the gene variation scoring result, using a supervised learning model and a weighted summation method to perform multi-dimensional pathogenicity comprehensive scoring and evaluation on the pre-acquired gene mutation data to obtain a pathogenic mutation ranking scoring result.

2. The method of ranking pathogenic mutations of a single gene genetic disease according to claim 1, wherein, The method comprises: S21, annotating the pre-acquired gene variation database to obtain a gene variation annotation result; The pre-acquired gene variation database comprises: a variant harmfulness database, a population frequency database, a human gene mutation database, a clinical variation database, and a variation classification system database; The gene variation annotation result includes: a variant harmfulness annotation result, a gene frequency annotation result, and a variant classification annotation result. S22, according to the gene variation annotation result, using a weighted summation method and a preset scoring rule to score the pre-acquired gene mutation data to obtain a gene variation scoring result.

3. The method of ranking pathogenic mutations of a single gene genetic disease according to claim 2, wherein, The gene variation scoring result includes: a variant harmfulness scoring result, a variant occurrence rate scoring result, a variant report scoring result, and a variant pathogenicity scoring result.

4. The method of ranking pathogenic mutations of a single gene genetic disease according to claim 1, wherein, According to the phenotype and gene association degree scoring result and the gene variation scoring result, using a supervised learning model and a weighted summation method to perform multi-dimensional pathogenicity comprehensive scoring evaluation on the pre-acquired gene mutation data to obtain a pathogenic mutation ranking scoring result, including: S31, based on the phenotype and gene association degree scoring result and the gene variation scoring result, setting the weight of the multi-dimensional pathogenicity comprehensive scoring evaluation; S32, using a weight training mechanism of ranking learning to combine the pre-acquired gene mutation data to optimize the weight of the set weight to obtain a weight optimization result; S33, according to the weight optimization result, using a weighted summation method to perform multi-dimensional pathogenicity comprehensive scoring evaluation on the pre-acquired gene mutation data to obtain a pathogenic mutation ranking scoring result.

5. The method of ranking pathogenic mutations of a single gene genetic disease according to claim 4, wherein, The weight optimization result obtained by using the weight training mechanism of ranking learning to combine the pre-acquired gene mutation data to optimize the weight of the set weight includes: S321, constructing a training sample pair by using the pre-acquired gene mutation data to obtain a plurality of positive and negative sample pairs with pathogenic site division; S322, using the positive and negative sample pairs with pathogenic site division to construct a pair-wise ranking loss function, and taking the pair-wise ranking loss function as a training target of the weight training mechanism of ranking learning; S323, based on the pair-wise ranking loss function, using an optimizer to optimize and iteratively update the weight of the set weight to obtain a weight optimization result.

6. The method of ranking pathogenic mutations of a single gene genetic disease according to claim 5, wherein, The expression of the pair-wise ranking loss function is: ； wherein, represents the pair-wise ranking loss function; represents the pathogenic variant in the th sample; is the combined score of the pathogenic variant in the th sample; is the score of the benign variant paired with the pathogenic variant in the th sample; represents the benign variant paired with the pathogenic variant; β represents the hyperparameter that controls how steep the loss surface is; N represents the total number of positive and negative sample pairs.

7. A scoring and ranking system for pathogenic mutations of monogenic genetic diseases for implementing the method of scoring and ranking pathogenic mutations of monogenic genetic diseases according to any one of claims 1 to 6, characterized in that, The scoring and ranking system of pathogenic mutations of monogenic genetic diseases includes: a phenotype and gene association scoring module, a gene variation scoring module, and a pathogenicity comprehensive evaluation module; The phenotype and gene association scoring module is used for based on the pre-acquired phenotype data and literature abstracts, using a hybrid strategy to extract standardized phenotype terms from the pre-acquired symptom descriptions, and using a similarity comparison method to score the degree of association between the standardized phenotype terms and the pre-acquired gene mutation data; The gene variation scoring module is used for annotating the pre-acquired gene variation database, and scoring the pre-acquired gene mutation data according to the gene variation annotation result; The pathogenicity comprehensive evaluation module is used for according to the phenotype and gene association degree scoring result and the gene variation scoring result, using a supervised learning model and a weighted summation method to perform multi-dimensional pathogenicity comprehensive scoring evaluation on the pre-acquired gene mutation data to obtain a pathogenic mutation ranking scoring result.

Images