CNV pathogenicity prediction method based on ACMG guideline quantified features
Patent Information
- Application Number
- CN202611213749.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-10-30
AI Technical Summary
[0003]前述软件将指南中的研究证据拆解为具体评分条款，系统接收到变异的染色体编号、起止坐标以及变异类型等基础特征后，通过累计分值并比对固定阈值，输出致病性五分类状态，然而，在高通量处理工况下，由于最新文献、表型匹配以及家系共分离等核心条款无法通过固化程序自动获取，导致输入特征矩阵在遭遇条款缺项时产生大范围空缺，引发边界类别判定出现模糊偏倚；同时，知识库存在更新周期滞后问题，使得系统针对新发罕见变异缺乏动态证据支撑，而纯数据驱动的黑箱模型又因决策路径无法与临床得分条款建立因果映射，从而产生合规性壁垒与沉重的人工复核开销；针对此类缺陷，扩大知识库规模或者增大模型深度的线性改进路径无法攻克上述束缚，扩充数据源会加剧特征稀疏，而增加模型参数不仅无法建立透明映射通道，还会因样本类别分布失衡触发过拟合，使得系统面对复杂变异工况时丧失泛化边界稳定性
[0020]1、本发明将离散的美国医学遗传学与基因组学学会指南证据条款按照分值梯度归集至特定的风险分级，统计待测拷贝数变异在各分值风险分级内获得阳性评分的条款数目以构造结构化数值特征，配合多元逻辑回归模型处理拷贝数变异矩阵；此特征构造方式有效聚合零散微弱证据，在遭遇部分临床证据缺失而导致特征稀疏的工况时，通过特定风险包络线内的整体阳性计数稳定映射出变异区间致病强度的变化路径，避免因指标空缺引发的计算失真，实现低功耗下高密度遗传信息的稳健收敛。
Smart Images
Figure CN123017173A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of healthcare informatics technology, and more specifically, to a method for predicting CNV pathogenicity based on quantitative features of ACMG guidelines. Background Technology
[0002] In genetic data processing workflows such as chromosome microarray analysis and low-depth whole-genome sequencing, automatic interpretation of copy number variations is a key step in variation interpretation. The industry currently generally adopts rule-based automated software or supervised prediction models, which connect to public knowledge bases to obtain gene dose sensitivity, population allele frequency, and past case information covering the variant segments. This replaces the semi-quantitative scoring process of manual retrieval and meets the standardization requirements of high-throughput data flow.
[0003] The aforementioned software breaks down the research evidence in the guidelines into specific scoring clauses. After receiving basic features such as chromosome number, start and end coordinates, and variant type, the system accumulates scores and compares them with fixed thresholds to output a five-category pathogenicity status. However, under high-throughput processing conditions, core clauses such as the latest literature, phenotypic matching, and family co-segregation cannot be automatically obtained through a fixed program. This results in a large number of gaps in the input feature matrix when encountering missing clauses, leading to fuzzy bias in boundary category determination. At the same time, the knowledge base has a lag in its update cycle, which makes the system lack dynamic evidence support for newly emerging rare variants. The purely data-driven black box model cannot establish a causal mapping between the decision path and the clinical scoring clauses, thus creating compliance barriers and heavy manual review costs. To address these shortcomings, linear improvement paths such as expanding the size of the knowledge base or increasing the model depth cannot overcome the above constraints. Expanding the data source will exacerbate feature sparsity, while increasing model parameters will not only fail to establish a transparent mapping channel, but will also trigger overfitting due to the imbalance of sample class distribution, causing the system to lose generalization boundary stability when facing complex variant situations.
[0004] Therefore, how to construct a dual-dimensional fusion feature framework and integrate linear classification algorithms with additive interpretation mechanisms to achieve traceable and accurate five-class classification of pathogenicity of variants has become the technical problem to be solved by this invention. Summary of the Invention
[0005] This invention provides a method for predicting CNV pathogenicity based on ACMG guidelines-quantified features, comprising the following steps:
[0006] Step S1: Obtain the copy number variant sample to be tested, and extract the genomic variant coordinates, clinical phenotype text, and variant detection frequency of the copy number variant sample to be tested;
[0007] Step S2: Compare the genomic variant coordinates with the preset dose-sensitive region database and the pathogenic variant interval database, calculate the base overlap ratio of the dose-sensitive region covered by the copy number variant sample to be tested, and obtain the first quantitative score characterizing the dose-sensitive constraint of the region; retrieve the topological structure of the human phenotype ontology, calculate the multi-path directed acyclic graph semantic distance between the clinical phenotype text and the known pathogenic gene-associated phenotype to obtain the normalized semantic similarity, and obtain the second quantitative score characterizing the variant phenotype features;
[0008] Step S3: Integrate the first quantization score, the second quantization score, and the mutation detection frequency to construct a multidimensional feature quantization vector to characterize the copy number mutation sample to be tested; wherein, a weight adjustment rule is set in the multidimensional feature quantization vector, and when the first quantization score is lower than the preset safety boundary threshold, the weight coefficient ratio of the second quantization score in the multidimensional feature quantization vector is reduced.
[0009] Step S4: Input the multidimensional feature quantization vector into the preset pathogenicity classification prediction model for regression convergence calculation, and output the pathogenicity prediction probability value of the copy number variant sample to be tested.
[0010] Preferably, before step S1, in order to balance the distribution of historical variant categories, a sample allocation step is also included: acquiring records of genes sensitive to insufficient haplo-dose and records of genes sensitive to triple-dose to supplement the pathogenic variant sample set; acquiring records of copy number variants with allele frequencies greater than 1% to supplement the benign variant sample set; and coordinating the distribution ratio of the pathogenic variant sample set and the benign variant sample set to construct a standardized training data environment for the preset pathogenicity classification prediction model.
[0011] Preferably, in step S2, the copy number variation sample to be tested is compared with the variation records in the preset document evidence database. Under the constraint that the variation range of the copy number variation sample to be tested completely includes the variation range of the variation records, multidimensional document evidence features reflecting the number of inconsistent variation types, the number of similar types of entries, and the number of benign records are extracted. The calculation of multidimensional document evidence features includes obtaining a document adjustment score, the formula for which is: ,in, Adjusting the score for the literature; The number of items of the same type; The number of entries with inconsistent variant types; The number of benign records; the literature adjustment score is input as an additional feature dimension into the preset pathogenicity classification prediction model, which is used to adjust the prediction classification boundary of the preset pathogenicity classification prediction model in step S4.
[0012] Preferably, the calculation of the first quantitative score in step S2 specifically includes the following sub-steps: Step S21, retrieve the scores of 14 automatically assessable clauses contained in the copy number variant sample to be tested, and count the number of clauses that obtained positive scores in each risk level of the copy number variant sample to be tested to generate a graded structured feature value; Step S22, fuse the graded structured feature value with the base overlap ratio of the dose-sensitive region to calculate and map the first quantitative score.
[0013] Preferably, the calculation of the second quantitative score in step S2 specifically includes the following sub-steps: Step S23, retrieving a literature knowledge base that stores known variation intervals, associated diseases, and clinical phenotypes; Step S24, retrieving the multi-path evolution trajectory of the clinical phenotype text in the topological structure of the human phenotype ontology, calculating the directed acyclic graph topological path distance between the clinical phenotype text and the gene-related phenotypes recorded in the literature knowledge base, and generating a normalized semantic similarity to map the second quantitative score.
[0014] Preferably, the pathogenicity prediction probability value corresponds to five categories of pathogenicity rating status, which consist of pathogenic, possibly pathogenic, clinically indeterminate, possibly benign, and benign.
[0015] Preferably, after step S4, an interpretability analysis step is included: Step S5, using the Shapley additive interpretability model to analyze the preset pathogenicity classification prediction model, calculate the marginal contribution of each input feature dimension to generate feature contribution distribution data, and map the feature contribution distribution data to the first quantitative score and the second quantitative score to generate traceable clinical compliance interpretability information.
[0016] Preferably, the pre-defined pathogenicity classification prediction model is constructed using a multivariate logistic regression model, and the linear weight coefficients of the multivariate logistic regression model are determined by training and fitting a standardized regression multidimensional parameter set matrix.
[0017] Preferably, the variant records stored in the pre-defined literature evidence database include gene symbols, chromosome numbers, genome start coordinates, genome end coordinates, variant length, copy number variant type, phenotypic information, disease names, inheritance patterns, and literature publication identifiers.
[0018] Preferably, in step S5, mapping the feature contribution distribution data to the first quantification score and the second quantification score includes: extracting the top three key core features with the highest marginal contribution in the feature contribution distribution data, highlighting the evidence-based clauses and related literature records corresponding to the key core features in the output visualization interactive page, and generating an auxiliary support carrier for clinical diagnostic decision-making.
[0019] The embodiments of the present invention have at least the following beneficial effects:
[0020] 1. This invention aggregates discrete evidence clauses from the American Society for Medical Genetics and Genomics guidelines into specific risk tiers according to their score gradients. It counts the number of clauses that receive positive scores for the copy number variants under test within each risk tier to construct structured numerical features. This is then combined with a multivariate logistic regression model to process the copy number variant matrix. This feature construction method effectively aggregates scattered and weak evidence. In situations where some clinical evidence is missing, resulting in sparse features, the overall positive count within a specific risk envelope stably maps the path of change in pathogenicity of the variant interval, avoiding computational distortion caused by missing indicators and achieving robust convergence of high-density genetic information with low power consumption.
[0021] 2. Based on the guideline grading system, by comparing dynamic literature data containing variation intervals, diseases, and phenotypes, and under the constraint that the test interval completely includes the record interval, multidimensional quantitative features reflecting the number of inconsistent variation types, the number of similar types of entries, and the number of benign records are adaptively extracted, and the latest literature evidence is integrated into the five-category pathogenicity prediction model. This multi-mechanism linkage architecture enables the information processing loop to overcome the lag barrier of the static knowledge base, and uses multidimensional matching counts to adjust the classification boundary in reverse. When encountering challenging situations with rare structural variations, the calibration effect of the negative feature dimension suppresses the risk of false positives, and smoothly extends the data flow while ensuring the compliance of interpretation.
[0022] 3. This invention deeply integrates linear classification models with Shapley's additive interpretation mechanism. It uses a processor to calculate the marginal contribution value of each feature dimension in a single sample to generate feature contribution distribution data. After receiving basic variation data containing chromosome codes, start and end coordinates, variation types, and reference versions, it directly maps the calculated feature contributions to specific evidence-based clauses and related literature entries. This design constructs a technical closed loop of data perception and traceable output across the entire chain, transforming the unknown decision-making mechanism of the black-box prediction model into verifiable deterministic evidence, eliminating technical barriers in the auxiliary decision-making process, and enhancing human-machine trust while improving adaptability to complex working conditions. Attached Figure Description
[0023] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein:
[0024] Figure 1 This is a flowchart illustrating the pathogenicity determination process for variant samples in this invention.
[0025] Figure 2 This is a schematic diagram of the data flow in the variation prediction model of this invention. Detailed Implementation
[0026] The principles and spirit of the invention will now be described with reference to the accompanying drawings and several exemplary embodiments. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement the invention, and are not intended to limit the scope of the invention in any way. Rather, these embodiments are provided to make the invention more thorough and complete, and to fully convey the scope of the invention to those skilled in the art.
[0027] A method for predicting CNV pathogenicity based on ACMG guidelines-quantified features includes the following steps:
[0028] Step S1: Obtain the copy number variant sample to be tested, and extract the genomic variant coordinates, clinical phenotype text, and variant detection frequency of the copy number variant sample to be tested;
[0029] Step S2: Compare the genomic variant coordinates with the preset dose-sensitive region database and the pathogenic variant interval database, calculate the base overlap ratio of the dose-sensitive region covered by the copy number variant sample to be tested, and obtain the first quantitative score characterizing the dose-sensitive constraint of the region; retrieve the topological structure of the human phenotype ontology, calculate the multi-path directed acyclic graph semantic distance between the clinical phenotype text and the known pathogenic gene-associated phenotype to obtain the normalized semantic similarity, and obtain the second quantitative score characterizing the variant phenotype features;
[0030] Step S3: Integrate the first quantization score, the second quantization score, and the mutation detection frequency to construct a multidimensional feature quantization vector to characterize the copy number mutation sample to be tested; wherein, a weight adjustment rule is set in the multidimensional feature quantization vector, and when the first quantization score is lower than the preset safety boundary threshold, the weight coefficient ratio of the second quantization score in the multidimensional feature quantization vector is reduced.
[0031] Step S4: Input the multidimensional feature quantization vector into the preset pathogenicity classification prediction model for regression convergence calculation, and output the pathogenicity prediction probability value of the copy number variant sample to be tested.
[0032] Preferably, before step S1, in order to balance the distribution of historical variant categories, a sample allocation step is also included: acquiring records of genes sensitive to insufficient haplo-dose and records of genes sensitive to triple-dose to supplement the pathogenic variant sample set; acquiring records of copy number variants with allele frequencies greater than 1% to supplement the benign variant sample set; and coordinating the distribution ratio of the pathogenic variant sample set and the benign variant sample set to construct a standardized training data environment for the preset pathogenicity classification prediction model.
[0033] Preferably, in step S2, the copy number variation sample to be tested is compared with the variation records in the preset document evidence database. Under the constraint that the variation range of the copy number variation sample to be tested completely includes the variation range of the variation records, multidimensional document evidence features reflecting the number of inconsistent variation types, the number of similar types of entries, and the number of benign records are extracted. The calculation of multidimensional document evidence features includes obtaining a document adjustment score, the formula for which is: ,in, Adjusting the score for the literature; The number of items of the same type; The number of entries with inconsistent variant types; The number of benign records; the literature adjustment score is input as an additional feature dimension into the preset pathogenicity classification prediction model, which is used to adjust the prediction classification boundary of the preset pathogenicity classification prediction model in step S4.
[0034] Preferably, the calculation of the first quantitative score in step S2 specifically includes the following sub-steps: Step S21, retrieve the scores of 14 automatically assessable clauses contained in the copy number variant sample to be tested, and count the number of clauses that obtained positive scores in each risk level of the copy number variant sample to be tested to generate a graded structured feature value; Step S22, fuse the graded structured feature value with the base overlap ratio of the dose-sensitive region to calculate and map the first quantitative score.
[0035] Preferably, the calculation of the second quantitative score in step S2 specifically includes the following sub-steps: Step S23, retrieving a literature knowledge base that stores known variation intervals, associated diseases, and clinical phenotypes; Step S24, retrieving the multi-path evolution trajectory of the clinical phenotype text in the topological structure of the human phenotype ontology, calculating the directed acyclic graph topological path distance between the clinical phenotype text and the gene-related phenotypes recorded in the literature knowledge base, and generating a normalized semantic similarity to map the second quantitative score.
[0036] Preferably, the pathogenicity prediction probability value corresponds to five categories of pathogenicity rating status, which consist of pathogenic, possibly pathogenic, clinically indeterminate, possibly benign, and benign.
[0037] Preferably, after step S4, an interpretability analysis step is included: Step S5, using the Shapley additive interpretability model to analyze the preset pathogenicity classification prediction model, calculate the marginal contribution of each input feature dimension to generate feature contribution distribution data, and map the feature contribution distribution data to the first quantitative score and the second quantitative score to generate traceable clinical compliance interpretability information.
[0038] Preferably, the pre-defined pathogenicity classification prediction model is constructed using a multivariate logistic regression model, and the linear weight coefficients of the multivariate logistic regression model are determined by training and fitting a standardized regression multidimensional parameter set matrix.
[0039] Preferably, the variant records stored in the pre-defined literature evidence database include gene symbols, chromosome numbers, genome start coordinates, genome end coordinates, variant length, copy number variant type, phenotypic information, disease names, inheritance patterns, and literature publication identifiers.
[0040] Preferably, in step S5, mapping the feature contribution distribution data to the first quantification score and the second quantification score includes: extracting the top three key core features with the highest marginal contribution in the feature contribution distribution data, highlighting the evidence-based clauses and related literature records corresponding to the key core features in the output visualization interactive page, and generating an auxiliary support carrier for clinical diagnostic decision-making.
[0041] Example 1: This example combines Figures 1 to 2 This paper describes a method for predicting the pathogenicity of CNVs based on quantitative features from the ACMG guidelines. Figure 1As shown, step S1: Basic Feature Extraction extracts the genomic variant coordinates, clinical phenotype text, and variant detection frequency of the copy number variant sample to be tested, and transmits the extracted data to subsequent stages. In step S2: Regional Dose Sensitivity Constraint Calculation, the genomic variant coordinates are compared with the dose-sensitive region and pathogenic variant interval database to calculate the base overlap ratio, obtain the first quantization score, and output it to the safety boundary threshold determination. The safety boundary threshold determination determines whether the first quantization score is lower than the preset safety boundary threshold. If the determination is yes (lower than the safety boundary threshold), the signal is output to the weight adjustment rule to trigger the adjustment of the weight coefficient ratio of the second quantization score in the multidimensional feature quantization vector. If the determination is no (within the safety boundary), the process directly proceeds to step S3: Multidimensional Feature Quantization Vector Construction. At the same time, in step S2: Phenotype Text... In semantic similarity calculation, the topological structure of the human phenotype ontology is retrieved, and the semantic distance of the multi-path directed acyclic graph between the clinical phenotype text and the phenotype associated with known pathogenic genes is calculated. The normalized semantic similarity is obtained to obtain the second quantization score, which is then output to step S3: construction of multidimensional feature quantization vector. The data output by the safety boundary threshold judgment or weight adjustment rule, the second quantization score output by step S2: phenotype semantic similarity calculation, and the variant detection frequency output by step S1: basic feature extraction are all converged in step S3: construction of multidimensional feature quantization vector. The first quantization score, the second quantization score, and the variant detection frequency are fused to construct a multidimensional feature quantization vector to represent the copy number variant sample to be tested. Finally, the multidimensional feature quantization vector is input into step S4: classification model prediction for regression convergence calculation, and the pathogenicity prediction probability value of the copy number variant sample to be tested is output.
[0042] like Figure 2As shown, the copy number variation sample to be tested is input into the basic feature data extraction. After processing, the basic feature data is split and output into multiple streams. Genomic variation coordinates are output to calculate the dose-sensitive constraint score of the region; variation interval data is output to adaptively extract multidimensional literature evidence features; clinical phenotype text is output to calculate the semantic distance of the phenotype topological path; and variation detection frequency is directly output to perform weight adjustment and construct a multidimensional feature quantization vector. Simultaneously, the data storage dose-sensitive region database and pathogenic variation interval database input base overlap ratio data to calculate the dose-sensitive constraint score of the region. The dose-sensitive constraint score of the region, combined with the genomic variation coordinates and base overlap ratio data, is used to calculate and output the first quantization score and hierarchical structured feature values to perform weight adjustment and construct a multidimensional feature quantization vector. At the same time, the data storage preset literature evidence database inputs literature interval inclusion constraint records to adaptively extract multidimensional literature evidence features. The adaptively extracted multidimensional literature evidence features, combined with the variation interval data and the literature interval inclusion constraint records, are processed and output. The process involves several steps: First, the literature-adjusted score is used to perform weight adjustment and construct a multi-dimensional feature quantization vector. Second, the human phenotype ontology topology is stored, and the semantic distance of the phenotype topological path is input into a directed acyclic graph (DAG) of the multi-path evolution trajectory. The semantic distance of the phenotype topological path is calculated by combining the clinical phenotype text with the DAG of the multi-path evolution trajectory, and the second quantization score and normalized semantic similarity are output to perform weight adjustment and construct a multi-dimensional feature quantization vector. Next, the received variant detection frequency, first quantization score, hierarchical structured feature values, literature-adjusted score, and second quantization score and normalized semantic similarity are fused and weighted. The resulting multi-dimensional feature quantization vector is then used for regression convergence calculation of the classification prediction model. Subsequently, the pathogenicity prediction probability value is output through the regression convergence calculation of the classification prediction model and processed by the marginal contribution analysis of the Shapley additive interpretation model. Finally, the marginal contribution analysis of the Shapley additive interpretation model outputs the pathogenicity rating status of five variants and traceable clinical compliance interpretation information to the clinical diagnostic decision support carrier.
[0043] The above description is merely an explanation of some preferred embodiments of the present invention and the technical principles employed. Those skilled in the art should understand that the scope of the invention as described in the embodiments of the present invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.
Claims
1. A method for predicting the pathogenicity of CNVs based on ACMG guidelines quantitative features, characterized in that, Includes the following steps: Step S1: Obtain the copy number variant sample to be tested, and extract the genomic variant coordinates, clinical phenotype text, and variant detection frequency of the copy number variant sample to be tested; Step S2: Compare the genomic variation coordinates with the preset dose-sensitive region database and the pathogenic variation interval database, calculate the base overlap ratio of the dose-sensitive region covered by the copy number variation sample to be tested, and obtain the first quantitative score characterizing the dose-sensitive constraint of the region. The topological structure of the human phenotype ontology is retrieved, and the semantic distance of the multipath directed acyclic graph between the clinical phenotype text and the phenotype associated with known pathogenic genes is calculated to obtain the normalized semantic similarity, thus obtaining the second quantitative score representing the characteristics of the variant phenotype. Step S3: Integrate the first quantization score, the second quantization score, and the mutation detection frequency to construct a multidimensional feature quantization vector to characterize the copy number mutation sample to be tested; wherein, a weight adjustment rule is set in the multidimensional feature quantization vector, and when the first quantization score is lower than the preset safety boundary threshold, the weight coefficient ratio of the second quantization score in the multidimensional feature quantization vector is reduced. Step S4: Input the multidimensional feature quantization vector into the preset pathogenicity classification prediction model for regression convergence calculation, and output the pathogenicity prediction probability value of the copy number variant sample to be tested.
2. The method for predicting CNV pathogenicity based on ACMG guideline quantitative features according to claim 1, characterized in that, Before step S1, in order to balance the distribution of historical variant categories, a sample allocation step is also included: obtaining records of haplo-dose-insensitive genes and triplo-dose-sensitive genes to supplement the pathogenic variant sample set; obtaining records of copy number variants with allele frequencies greater than 1% to supplement the benign variant sample set; and coordinating the distribution ratio of the pathogenic variant sample set and the benign variant sample set to construct a standardized training data environment for the preset pathogenicity classification prediction model.
3. The method for predicting CNV pathogenicity based on ACMG guideline quantitative features according to claim 1, characterized in that, In step S2, the copy number variation sample to be tested is compared with the variation records in the preset document evidence database. Under the constraint that the variation range of the copy number variation sample to be tested completely includes the variation range of the variation records, multidimensional document evidence features reflecting the number of inconsistent variation types, the number of similar types of entries, and the number of benign records are extracted. The calculation of multidimensional document evidence features includes obtaining a document adjustment score, the formula for which is: ,in, Adjusting the score for the literature; The number of items of the same type; The number of entries with inconsistent variant types; The number of benign records; the literature adjustment score is input as an additional feature dimension into the preset pathogenicity classification prediction model, which is used to adjust the prediction classification boundary of the preset pathogenicity classification prediction model in step S4.
4. The method for predicting CNV pathogenicity based on ACMG guideline quantitative features according to claim 1, characterized in that, The calculation of the first quantitative score in step S2 specifically includes the following sub-steps: Step S21, retrieve the scores of 14 automatically assessable clauses contained in the copy number variant sample to be tested, and count the number of clauses that obtained positive scores in each risk level of the copy number variant sample to be tested to generate hierarchical structured feature values. Step S22: The hierarchical structured feature values are fused with the base overlap ratio of the dose-sensitive region to calculate and map the first quantization score.
5. The method for predicting CNV pathogenicity based on ACMG guideline quantitative features according to claim 1, characterized in that, The calculation of the second quantitative score in step S2 specifically includes the following sub-steps: Step S23, retrieve the literature knowledge base that stores known variation intervals, associated diseases and clinical phenotypes; Step S24, retrieve the multi-path evolution trajectory of clinical phenotype text in the topological structure of human phenotype ontology, calculate the directed acyclic graph topological path distance between clinical phenotype text and gene-related phenotypes recorded in the literature knowledge base, and generate normalized semantic similarity to map the second quantitative score.
6. The method for predicting CNV pathogenicity based on ACMG guideline quantitative features according to claim 1, characterized in that, The pathogenicity prediction probability value corresponds to five categories of pathogenicity rating status, which consist of pathogenic, possibly pathogenic, clinically indeterminate, possibly benign, and benign.
7. The method for predicting CNV pathogenicity based on ACMG guideline quantitative features according to claim 1, characterized in that, Following step S4, an interpretability analysis step is also included: Step S5, using the Shapley additive interpretability model to analyze the preset pathogenicity classification prediction model, calculate the marginal contribution of each input feature dimension to generate feature contribution distribution data, and map the feature contribution distribution data to the first quantitative score and the second quantitative score to generate traceable clinical compliance interpretability information.
8. The method for predicting CNV pathogenicity based on ACMG guideline quantitative features according to claim 1, characterized in that, The pre-defined pathogenicity classification prediction model is constructed using a multivariate logistic regression model. The linear weight coefficients of the multivariate logistic regression model are determined by training and fitting a standardized regression multidimensional parameter set matrix.
9. The method for predicting CNV pathogenicity based on ACMG guideline quantitative features according to claim 3, characterized in that, The variant records stored in the pre-defined literature evidence database include gene symbols, chromosome numbers, genome start coordinates, genome end coordinates, variant length, copy number, variant type, phenotypic information, disease name, inheritance pattern, and literature publication identifier.
10. A method for predicting CNV pathogenicity based on ACMG guideline quantitative features according to claim 7, characterized in that, In step S5, mapping the feature contribution distribution data to the first quantification score and the second quantification score includes: extracting the top 3 key core features with the highest marginal contribution in the feature contribution distribution data, highlighting the evidence-based clauses and related literature records corresponding to the key core features in the output visualization interactive page, and generating an auxiliary support carrier for clinical diagnostic decision-making.