Multi-modal annotation generated gene mutation prediction method

Abstract

The present application relates to the technical field of gene mutation prediction, and discloses a gene mutation prediction method generated by multi-mode annotation, and the specific process comprises the following steps: carrying out mutation type annotation on input single-base mutation position information to obtain mutation basic information containing mutation types, then using an ANNOVAR annotation tool, SpliceAI splicing effect prediction software and reference mutation information of a function effect database to carry out multi-dimensional feature annotation, using Bayesian PCA based on the obtained multi-dimensional feature mutation data set to fill in the annotation data, then using an automatic engineering feature list and a separated feature selection list to carry out feature combination and screening, and obtaining a gene mutation prediction score after gradient generation tree algorithm. The present application can be used for predicting all non-synonymous exon mutations, has good performance in classifying rare benign mutations, and can identify a small amount of mutations with high pathogenic probability from a large amount of candidate mutations.

Description

Technical Field

[0001] This invention relates to the field of gene mutation prediction technology, specifically a gene mutation prediction method based on multi-mode annotation generation. Background Technology

[0002] The rapid accumulation of whole-genome sequencing (WGS) and whole-exome sequencing (WES) data has led to the discovery of a large number of pathological and non-pathological genetic mutations. To aid in the assessment and understanding of these mutations, research institutions worldwide have established population databases, such as gnomAD, ExAC, and ChinaMap. Furthermore, genetic disease databases, such as ClinVar, OMIM, and HGMD, have also accumulated a wealth of information on known pathological or benign genetic mutations. These databases are widely used as references for the genetic diagnosis of Mendelian genetic diseases.

[0003] Pathological mutations leading to Mendelian genetic diseases are known to function through various biological mechanisms, resulting in extensive classification and research across different areas. For example, exon mutations based on protein sequence alterations are classified as synonymous mutations, missense mutations, termination mutations, termination deletions, frameshift mutations, etc. Synonymous mutations do not alter the protein sequence, while missense mutations result in encoding different amino acids. Since protein sequence changes associated with missense mutations can be pathological, various studies have focused on predicting the pathological effects of missense mutations. On the other hand, some mutations are pathological at the RNA level through splicing alterations; these mutations are typically located in splice donor, acceptor, and intron regions. Therefore, splicing mutations are also considered for pathological assessment of mutations. However, when using whole-exome sequencing data for practical genetic diagnosis, different types of mutations and mechanisms should be considered simultaneously to identify pathological mutations. With the development of machine learning (ML) and deep learning (DL), many computational methods using ML or DL ​​have been developed to predict mutational disruption or pathology.

[0004] Some algorithms in the aforementioned fields consider ensemble features obtained from multiple dimensions and build upon existing pathogenicity prediction methods. For example, MutPred2 (Mutation Predictor 2). [1] And REVEL (Rare ExomeVariant Ensemble Learner) [2] The input to MutPred2 is an amino acid sequence s, i.e., a wild-type protein sequence, and an amino acid substitution XiY, where X is the i-th amino acid in s that is replaced by Y. We refer to the mutant (mt) sequence as s. XiYMutPred2 used HGMD (Human Gene Mutation Database), SwissProt, and dbSNP as training sets. For a given sequence s and variant XiY, MutPred2 extracted 1345 features (including 20 optional features). These features were divided into six groups: (1) sequence-based features, (2) amino acid substitution-based features, (3) PSSM (Position-Specific Scoring Matrix)-based features, (4) conservation-based features, (5) homologous protein features (optional due to computational time requirements), and (6) features predicting changes in structural and functional properties. MutPred2 used a two-sample t-test for feature selection, retaining only features that returned a p-value < 0.01. To remove (approximately) collinear features, z-score normalization and principal component analysis were performed on the selected features, with the variance retained at 99%. An ensemble of 30 feedforward neural networks was then trained on the resulting feature matrix. Each network consists of a hidden layer with four neurons and one output neuron (both layers use the tanh activation function). Training is performed using a bagging method, with each network trained with replacement on balanced random samples from the original training set. To determine the required number of iterations for training, MutPred2 reserves 25% of the training data as a validation set. The final model is trained using the resilient propagation algorithm and stops when the optimal number of iterations is reached, 1000 epochs are completed, or 500 checkpoints are reached. The prediction score is then calculated as the average of the output scores from all 30 validation checks. The output of MutPred2 includes a pathogenicity score ranging from [0,1], and a list of molecular mechanisms that may be affected by XiY. A pathogenicity score of 1 indicates that the mutation is almost certainly pathogenic, while a score of 0 indicates that the mutation is almost certainly benign.

[0005] REVEL is a method proposed by Nilah et al. The training set for REVEL comes from HGMD, ESP (Exome Sequencing Project), and KGP (1000 Genomes Project). REVEL integrates 18 pathogenicity prediction scores from 13 tools as predictive features. These include 10 functional prediction scores (MutPred, PROVEAN, SIFT, PolyPhen-2 HVAR & HDIV, LRT, MutationTaster, MutationAssessor, FATHMM v2.3, and VEST 3.0) and 8 conservation scores (GERP++, SiPhy, phyloP and phastCons scores for primates, vertebrates, and mammals). REVEL imputes missing features using the k-nearest neighbor method implemented in the R package. For a given mutation, the missing feature value is assigned the average of the non-missing feature values ​​of its k nearest neighbor mutations; when more than 50% of the feature values ​​of a given mutation are missing, each missing feature value is assigned the overall average across all mutations. Finally, REVEL is trained using a random forest algorithm containing 1,000 binary classification trees.

[0006] Although existing mutation pathogenicity prediction algorithms are widely used and developed using state-of-the-art technologies up to their respective publication dates, most of them are only applicable to specific types of mutations or rely on scores from published mutation pathogenicity prediction tools as prior knowledge. In practical prediction tasks, the pathogenicity of certain types of genetic mutations cannot be predicted. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a gene mutation prediction method for multi-mode annotation generation, which can be used to predict all non-synonymous exon mutations.

[0008] To address the aforementioned technical problems, this invention provides a gene mutation prediction method using multi-mode annotation generation. The process includes: annotating the input single-base mutation location information with mutation type to obtain basic mutation information containing mutation types; then performing multi-dimensional feature annotation to obtain a mutation dataset containing multi-dimensional features of the annotation results; then using Bayesian PCA to fill in the annotation data of the mutation dataset containing the multi-dimensional features of the annotation results; then using an automatic engineered feature list to generate features from the annotation data filled in by Bayesian PCA; then using a separate feature selection list to filter the data; and finally, passing the obtained mutation dataset containing all filtered features through a gradient spanning tree algorithm to obtain the gene mutation prediction score.

[0009] As an improvement to the gene mutation prediction method generated by the multi-modal annotation of this invention:

[0010] The mutation type annotation is performed using the refGene database.

[0011] The aforementioned feature annotations include annotations using the ANNOVAR annotation tool, the SpliceAI splicing effect prediction software, and reference mutation information from the functional effects database.

[0012] As a further improvement to the gene mutation prediction method generated by multi-mode annotation in this invention:

[0013] The ANNOVAR annotation tool is used to annotate population-based features, amino acid biochemical changes, and conservation scores.

[0014] For population-based features, this includes retrieving allele frequencies in various populations: whole exons (AF), raw allele frequencies (AF_raw), Africans (AF_afr), Latinos / mixed-race Americans (AF_amr), Ashkenazi Jews (AF_asj), East Asians (AF_eas), Finns (AF_fin), non-Finnish Europeans (AF_nfe), and other populations (AF_oth), as well as obtaining allele frequencies for different sexes from annotation information;

[0015] For changes in the biochemical properties of amino acids, first check whether the mutation leads to changes in amino acids. If not, set all relevant features to 0.

[0016] The physicochemical properties of each amino acid are stored in a matrix. The corresponding properties of amino acids are obtained by querying the matrix, and the difference in properties before and after mutation is used as the feature of the mutation. When a mutation affects multiple amino acids, the average value before and after the change is calculated separately. Information obtained from the BLOSUM100 matrix is ​​also used as features.

[0017] Conserved fractional characteristics include the fractional characteristics of phastCons, phyloP, and SiPhy in primates, mammals, and vertebrates.

[0018] As a further improvement to the gene mutation prediction method generated by multi-mode annotation in this invention:

[0019] The SpliceAI splicing effect prediction software annotates splicing effect features to obtain the predicted effect of each mutant on splicing and information on splicing changes relative to the mutant position.

[0020] The probability of the mutant’s effect on splicing is represented by the mutant’s change score, including receptor gain (DS_AG), receptor loss (DS_AL), donor gain (DS_DG), and donor loss (DS_DL). Information on splicing changes relative to the location of the mutant includes the location difference of receptor gain (DP_AG), receptor loss (DP_AL), donor gain (DP_DG), and donor loss (DP_DL).

[0021] If SpliceAI does not annotate the mutant, it means that the mutant is near the end of the chromosome or the reference sequence is too long. Use 0 to fill the missing values ​​predicted by SpliceAI.

[0022] As a further improvement to the gene mutation prediction method generated by multi-mode annotation in this invention:

[0023] The functional effect database is used to annotate the mutation locations using the functional effect features of epigenomic characteristics and gene damage index (GDI), residual mutation intolerance score (RVIS), gene intolerance score based on loss of function tool (LoFtool) (loF_score), and OMIM.

[0024] The epigenomic features of each mutation were annotated in nine different cell lines using a 15-state ChromHMM model to capture the spatial context of interactions (chromatin states) between different chromatin markers.

[0025] Functional effects include Gene Damage Index (GDI), Residual Mutation Intolerance Score (RVIS), LoF-based gene intolerance score (loF_score), mutant type, and annotations from the OMIM database; mutation type is used as a feature, including missense mutations, termination mutations, initiation deletion mutations, frameshift mutations, non-frameshift mutations, and termination deletion mutations.

[0026] The mutation inheritance patterns were annotated using the OMIM database and classified into five different types, including autosomal recessive, autosomal dominant, X-linked recessive, X-linked dominant, and others.

[0027] Then, one-hot encoding was performed on epigenomic features, mutant type features, and mutant inheritance pattern features.

[0028] As a further improvement to the gene mutation prediction method generated by multi-mode annotation in this invention:

[0029] The gradient spanning tree algorithm includes the LGBM model. The LGBM model training process is as follows: a training set and a test set are constructed, and then after mutation type annotation, multi-dimensional feature annotation, and automatic feature engineering with separation feature selection, a training set and a test set containing all selected features are obtained. The LGBM model is trained in 5 rounds, with 5-fold cross-validation and stepwise parameter adjustment used in each round, thereby obtaining a trained LGBM model.

[0030] As a further improvement to the gene mutation prediction method generated by multi-mode annotation in this invention:

[0031] The automatic feature engineering process with separation feature selection is specifically as follows:

[0032] (1) Use the missing value estimation method based on Bayesian PCA to fill in the missing values ​​of features in the mutation dataset containing multidimensional features;

[0033] (2) Automated Feature Engineering:

[0034] The openFE software package is used to perform mathematical transformations on the original features, such as logarithmic, square root, and exponential operations, and the original features are grouped according to the feature values ​​to generate new classification features. Then, the features obtained by feature selection using the default method defined in openFE are saved as the automatic engineering feature list.

[0035] (3) Separation feature selection:

[0036] For datasets with features obtained through automated feature engineering, an LGBM model is used for no more than 50 rounds of feature selection to reduce the number of features to 200 or fewer. After each round of feature selection, the importance of each feature is evaluated, and features with a relative importance score below 1e are discarded. -3 The features are then divided into two categories after 50 rounds of screening: the core feature set and the additional feature set. The feature combinations output by the separate feature selection algorithm are saved as the separate feature selection list.

[0037] As a further improvement to the gene mutation prediction method generated by multi-mode annotation in this invention:

[0038] The specific process of the separation feature selection algorithm includes:

[0039] Input: Core feature set C = {c1, c2, ..., c n}, additional feature set A = {a1, a2, ..., a n}, the final number of features is m;

[0040]

[0041] Output: Feature combination {C, A} with optimal performance k}

[0042] As a further improvement to the gene mutation prediction method generated by multi-mode annotation in this invention:

[0043] The specific process for constructing the training and test sets is as follows:

[0044] The ClinVar database was retrieved and filtered to remove mutations with conflicting and unknown labels.

[0045] Obtain the gnomAD database and select allele frequencies in the range of 1e -5 and 1e -3 Rare benign mutations were selected and screened: 5,000 mutations were randomly selected from each chromosome, and after being annotated with ANNOVAR, mutations that were missing in the mutation prediction features were filtered out. Then, 500 mutations from each chromosome were randomly selected and retained again. All mutations were retained from the Chr11 and ChrY chromosomes and qualified mutations from other chromosomes were randomly used to fill the gaps.

[0046] The mutations corresponding to the unique Reference SNP IDs were obtained and filtered to construct the orthogonal validation set SwissProt;

[0047] The mutations in the gnomAD library were divided into two subsets, gnomAD, at a ratio of 1.644:1. Clinvar and gnomAD SwissProt Then gnomAD Clinvar Merged with Clinvar to form ClinVar gnomAD , gnomAD SwissProt Merged with SwissProt to form SwissProt gnomAD Use a splitting algorithm to split ClinVar gnomAD The datasets are split into two sets, A and B, which are used as the training set and the test set, respectively.

[0048] As a further improvement to the gene mutation prediction method generated by multi-mode annotation in this invention:

[0049] The specific process of the splitting algorithm is as follows:

[0050] (1) Initialize two empty datasets A and B; two gene lists L A =[]、L B =[];The real-time proportion of A in D n A ;

[0051] (2) While true

[0052] (3) Do Ifn A ≥N

[0053] (4)、Then terminate and return{A,B}

[0054] (5) EndIf

[0055] (6) From L D Randomly select a gene g, n g The number of mutations corresponding to this gene in D.

[0056] (7) L A =L A ∪g,L D =L D -g,n A +=n g

[0057] (8) EndWhile

[0058] (9) L B =L D

[0059] (10) A = D[gene inL] A ],B=D[gene inL B ]

[0060] (11) Return A, B

[0061] Where D is the dataset to be partitioned, L A L B and L D Given a list of genes containing datasets A, B, and D, where N is the threshold for the proportion of mutations that should be added to A (0.9). A This represents the real-time proportion of A to D during the algorithm's execution.

[0062] The beneficial effects of this invention are mainly reflected in:

[0063] 1. This invention improves the model's performance in classifying rare benign mutations by adding more than 10,000 rare benign mutations from gnomAD (allele frequency (AF, mutation allele frequency) between 0.00001 and 0.0001) to the model training, which is crucial in reducing the false positive rate;

[0064] 2. The MAGPIE model of this invention extracts 6 feature modalities, including features with a low missing rate, thereby expanding the range of predictable mutation types. It also fills in the features in the dataset using a Bayes-based Principal Component Analysis (BPCA) method. It employs automatic feature engineering with separate feature selection to extract as much information as possible from the dataset based on the characteristics of the features and the training dataset.

[0065] 3. The MAGPIE model of this invention applies the Light GBM model (LGBM model, Light Gradient Boosting Machine) based on the gradient spanning tree algorithm, and achieves better performance compared with other tools that use machine learning or deep learning models. MAGPIE can predict multiple types of exon mutations and fill in 5-60% of the missing values ​​that were not applicable in previous methods. MAGPIE shows good performance on highly imbalanced validation datasets and mutations with low population allele frequencies, highlighting its advantage in interpreting unknown pathogenic mutations in clinically relevant applications for individual patients. It can be used to identify a small number of high-probability pathogenic mutations from a large number of candidate mutations. Attached Figure Description

[0066] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0067] Figure 1 This is a flowchart illustrating the MAGPIE model of the present invention;

[0068] Figure 2 is A schematic diagram illustrating the performance of the MAGPIE model of this invention on the test set;

[0069] Figure 3 This is a schematic diagram of the threshold test results when the number of pathogenic mutations exceeds that of benign mutations.

[0070] Figure 4 This is a schematic diagram illustrating the threshold test results when the number of benign mutations exceeds the number of pathogenic mutations.

[0071] Figure 5 A schematic diagram showing the prediction results of pathogenicity of mutations in genes related to Mendelian genetic diseases;

[0072] Figure 6 This is a schematic diagram illustrating the training process of the LGBM model of the present invention. Detailed Implementation

[0073] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto:

[0074] Example 1: The Multimodal Annotation-Generated Pathogenic Impact Evaluator (MAGPIE) is used to predict all non-synonymous exon mutations. To address various mutation types and pathological mechanisms, MAGPIE employs multimodal features for annotation, and then applies these generated high-dimensional features to a gradient boosting method for pathogenicity prediction. The specific methods for constructing and using MAGPIE online are as follows:

[0075] 1. Construct the training dataset

[0076] 1.1 Constructing ClinVargnomAD and SwissProtgnomAD

[0077] This invention uses a germline mutation database as the training dataset. Germline mutations refer to DNA changes inherited by an individual from egg and sperm cells during conception. The ClinVar database downloaded on June 24, 2022, includes non-synonymous single nucleotide mutations (missense mutations), termination mutations, initiation deletion mutations, frameshift mutations, non-frameshift mutations, and termination deletion mutations. Mutations are categorized as benign or pathogenic. Benign mutations are tagged with true negative labels, including Likely_benign, Benign, and Benign / Likely_benign tags; pathogenic mutations are tagged with true positive labels, including Likely_pathogenic, Pathogenic, and Pathogenic / Likely_pathogenic tags. All mutations with conflicting or unknown labels in ClinVar were filtered out. After filtering, 78,089 mutations were retained from the downloaded ClinVar database.

[0078] To improve the model's classification performance for rare benign mutations, these mutations need to be added to the training dataset. This helps identify the pathogenicity of rare mutations while preserving the importance of allele frequency information. Based on the gnomAD database, allele frequencies in the range of 1e... -5 and 1e -3Rare benign mutations were identified and screened. Specifically, for most chromosomes, approximately 500,000 mutations were available for subsequent use. 5,000 mutations were randomly selected from each chromosome, annotated with ANNOVAR (ANNOtate VARition), and then filtered to remove mutations with missing features used for mutation prediction. Then, 500 mutations were randomly selected and retained from each chromosome. However, after this screening process, chromosomes Chr11 and ChrY had fewer than 500 mutations. For these two chromosomes, all mutations were retained, and qualified mutations from other chromosomes were randomly used to fill the gaps. After this screening process, 11,998 mutations were introduced from gnomAD.

[0079] A SwissProt orthogonal validation set was constructed by downloading and filtering mutations corresponding to unique Reference SNP IDs from UniProt (downloaded on July 10, 2022) for further testing of the model's performance.

[0080] The ratio of mutations in ClinVar to those in SwissProt is 1.644:1. Based on this 1.644:1 ratio, the mutations in the gnomAD library are divided into two subsets: gnomAD. Clinvar and gnomAD SwissProt Then gnomAD Clinvar Merged with Clinvar to form ClinVar gnomAD , gnomAD SwissProt Merged with SwissProt to form SwissProt gnomAD This will be used for further training and evaluation. Enhancing the model's ability to handle rare benign mutations is key to reducing the false positive rate of allele frequencies in identifying pathogenic mutations in clinical scenarios.

[0081] 1.2 Constructing the training and test sets

[0082] Using the ClinVargnomAD constructed in step 1.1 as the dataset D to be split, a splitting algorithm is used to output the split datasets A and B. The splitting algorithm is as follows: Since this invention uses gene-level features, mutations in the same gene share the same gene-level feature scores, which may lead to some potential bias and label leakage. Therefore, dataset D is randomly split according to genes. The random state is set to 2^14 as the initial seed for data splitting, L... A L B and L D Given a list of genes containing datasets A, B, and D, where N is the threshold for the proportion of mutations that should be added to A (0.9). AThis represents the real-time proportion of A to D during the algorithm's execution.

[0083] Input: The dataset D to be split; the mutation rate threshold N; the list L of genes containing D. D .

[0084] Output: The split dataset {A,B}.

[0085] (1): Initialize two empty datasets A and B; two gene lists L A =[]、L B =[];The real-time proportion of A in D n A .

[0086] (2):While true

[0087] (3):do Ifn A ≥N

[0088] (4):Then terminate and return{A,B}

[0089] (5):EndIf

[0090] (6): From L D Randomly select a gene g, n g This represents the number of mutations corresponding to this gene in D.

[0091] (7):L A =L A ∪g,L D =L D -g,n A +=n g .

[0092] (8): EndWhile

[0093] (9):L B =L D .

[0094] (10): A = D[gene inL] A ],B=D[gene inL B ].

[0095] (11): Return A, B

[0096] After completing the above splitting steps, the resulting dataset is as follows: L A Includes from L D A randomly selected gene. L B These are the genes remaining after removing the selected genes from LD. Dataset A and Dataset B are respectively composed of genes belonging to category L. Aand L B Mutations that occur in the genes.

[0097] Dataset A and Dataset B are used as the training and testing sets, respectively. The training set contains 70,381 mutations originating from ClinVar and 6,675 mutations originating from gnomADClinVar, including 40,213 pathogenic mutations (true positive labels) and 36,843 benign mutations (true negative labels). The testing set contains 7,599 mutations originating from ClinVar and 1,067 mutations originating from gnomADClinVar, including 4,356 pathogenic mutations (true positive labels) and 4,310 benign mutations (true negative labels).

[0098] 2. MAGPIE Model Construction

[0099] MAGPIE model, for example Figure 1 As shown, the process includes four steps: mutation type annotation, multidimensional feature annotation, automatic feature engineering with separate feature selection, and mutation prediction based on the Light GBM (Light Gradient Boosting Machine, or LGBM) model.

[0100] 2.1. Mutation Type (Multi-type Variants) Annotation

[0101] The input is single-base mutation location information, including the chromosome where the mutation occurs, the mutation start position, the mutation end position, the base in the reference genome, and the base after the mutation. The input single-base mutation location information is annotated with the mutation type using the refGene database to obtain basic mutation information including the mutation type.

[0102] 2.2 Multimodal Feature Amotation Annotation

[0103] The input consists of basic mutation information including mutation types. Multidimensional feature annotation is performed using a functional effect related database, the SpliceAI open-source tool, and the ANNOVAR tool to obtain a mutation dataset containing multidimensional features from the annotation results of the above three tools. The multidimensional features cover six different modalities: (1) epigenomic features, (2) functional effect features, (3) splicing effect features, (4) population-based features, (5) biochemical property alteration features of amino acids, and (6) conservation score features.

[0104] (1) For epigenomic features and functional effect features, the mutation locations were annotated using functional effect related databases.

[0105] Epigenomic features of each mutation were annotated using a 15-state ChromHMM model across nine different cell lines to capture the spatial context of interactions (chromatin states) between different chromatin markers. Functional effects included the Gene Damage Index (GDI), Residual Mutation Intolerance Score (RVIS), a gene intolerance score based on the Loss-of-Function Tool (LoFtool) (loF_score), mutant type, and annotations from the OMIM database. Mutation type was used as a feature, including missense mutations, termination mutations, initiation deletion mutations, frameshift mutations, non-frameshift mutations, and termination deletion mutations. Mutation inheritance patterns were annotated using the OMIM database, categorized into five distinct types: autosomal recessive, autosomal dominant, X-linked recessive, X-linked dominant, and others. Epigenomic features, mutant type features, and mutation inheritance pattern features were one-hot encoded.

[0106] (2) For splicing effect features, the SpliceAI open-source tool downloaded from the official website was used for annotation;

[0107] Each mutant and its predicted effect on splicing are annotated by SpliceAI. The mutant change score represents the probability of the mutant's effect on splicing, including receptor gain (DS_AG), receptor loss (DS_AL), donor gain (DS_DG), and donor loss (DS_DL). Additionally, we obtained information on the splicing changes relative to the mutant's location, including the positional differences of receptor gain (DP_AG), receptor loss (DP_AL), donor gain (DP_DG), and donor loss (DP_DL). If SpliceAI does not annotate a mutant, it indicates that the mutant is close to the chromosome end or the reference sequence is too long. We use 0 to fill missing values ​​predicted by SpliceAI.

[0108] (3) For the remaining features (including population-based features, amino acid biochemical alteration features, and conservation score features), the latest version of ANNOVAR as of October 24, 2019 was used for annotation, where amino acid biochemical alteration features were represented by the difference before and after mutation.

[0109] Population-based characteristics represent the incidence of an allele in a population, including nine different allele frequencies (AFs). We retrieved allele frequencies from various populations: whole exon (AF), raw allele frequency (AF_raw), African (AF_afr), Latino / mixed-race American (AF_amr), Ashkenazi Jews (AF_asj), East Asian (AF_eas), Finnish (AF_fin), non-Finnish Europeans (AF_nfe), and other populations (AF_oth). Additionally, allele frequencies for different sexes were obtained from annotation information.

[0110] Biochemical characteristics include the effects of amino acid changes before and after the mutation. We first check whether the mutation leads to an amino acid change; if not, we set all relevant features to 0. Amino acid change (AAchange) forms the basis of our subsequent annotations. We store the physicochemical properties of each amino acid in a matrix, where Boolean features are represented by 1 / 0. We retrieve the corresponding properties of amino acids by querying the matrix and use the difference in properties before and after the mutation as the mutation's feature. When a mutation affects multiple amino acids, we calculate the average before and after each change to ensure this feature is consistent across different mutation types. Information from the BLOSUM100 matrix is ​​also used as features to demonstrate the conservation and similarity of amino acid substitutions.

[0111] Conserved fractional characteristics include the fractional characteristics of phastCons, phyloP, and SiPhy in primates, mammals, and vertebrates.

[0112] 2.3 Automatic Feature Engineering with Separated Feature Selection

[0113] Automatic feature engineering with separation feature selection consists of three steps: data imputation using a Bayesian PCA-based missing value estimation method; feature generation using automatic feature engineering; and feature selection using separation feature selection.

[0114] (1) The missing value estimation method based on Bayesian PCA (BPCA, proposed by Oba et al.) was used to fill in the missing values ​​of the features in the mutation dataset containing multidimensional features obtained in step 2.2 as a data preprocessing with automatic feature engineering with separate feature selection.

[0115] (2) Automated feature engineering is performed using the openFE software package to generate features. First, new categorical features are generated based on existing numerical features (by performing mathematical transformations such as logarithmic, square root, and exponential operations on the original features and grouping them according to their values ​​to generate new categorical features). These new categorical features can capture nonlinear relationships that the original numerical features might not capture. Second, after generating these new categorical features, feature selection is performed using the default methods defined in openFE to remove redundant or irrelevant features. Nevertheless, each variant's dataset still contains more than 3,000 features, which may lead to overfitting. The features selected using the default methods are saved as an automatically engineered feature list generated during the construction of the MAGPIE model using the training set data. This invention describes the automated feature engineering operations when using the MAGPIE model online in subsequent applications.

[0116] (3) Separated feature selection (SFS) is a feature selection strategy that extracts the optimal combination of features from a large number of candidate features to input into the model. First, for a dataset with over 3000 features obtained from the previous automatic feature engineering step, an LGBM model is used for no more than 50 rounds of feature selection to reduce the number of features to 200 or less. After each round of feature selection, the importance of each feature is evaluated, and features with a relative importance score below 1e are discarded. -3 The features selected after 50 rounds of screening based on feature importance are divided into two categories: a core feature set and an additional feature set. The core feature set includes features related to mutation pathogenicity with validated evidence, such as population-based features, conservation score features, and functional effect features, which are retained during training. The additional feature set includes supplementary features, including features of altered amino acid biochemical properties, splicing effects, epigenomic features, and features generated by automated feature engineering. Analysis revealed that the core features are significantly more important than the entire feature set. To further improve model performance, we use the feature combination output from the separate feature selection algorithm (i.e., the mutation dataset containing all selected features) as input for training in the LGBM. The feature combination output from the separate feature selection algorithm is saved as a separate feature selection list generated during the construction of the MAGPIE model using the training set data, for use in subsequent online use of the MAGPIE model in this invention.

[0117] The process of the feature selection algorithm is as follows:

[0118] Input: Core feature set C = {c1, c2, ..., c n}, Additional feature set A =

[0119] {a1,a2,…,a n}, the final number of features is m.

[0120] Output: Feature combination {C, A} with optimal performance k}

[0121]

[0122]

[0123] The output is a mutation dataset containing all selected features.

[0124] 2.4 Mutation prediction based on LGBM

[0125] This invention uses the LGBM model to predict the pathogenicity of a mutation dataset with all selected features. Compared to other tree-based models, Light GBM uses gradient one-sided sampling (GOSS) and exclusive feature binding (EFB) algorithms to accelerate the training process and ensure accuracy, resulting in faster training speed and lower memory consumption. The training and test sets constructed in step 1 are annotated with mutation types, multi-dimensional features, and automatic feature engineering with separate feature selection to obtain training and test sets containing all selected features. These sets are used to train the LGBM model and are progressively adjusted to keep the process controllable and avoid overfitting. To better learn from the data and avoid overfitting, especially as a tree-based model, a 5-round training process is used, with each round employing 5-fold cross-validation and progressive parameter tuning. An interpretable gradient boosting model (Light GBM, LGBM) is trained on top of the ensemble tree to predict the pathogenicity probability of candidate mutations. The training process of the LGBM module is as follows: Figure 6 As shown.

[0126] First, the model's initialization parameters are set, with the boost type set to GBDT and the learning rate set to 0.1. The goal of the first round is to improve accuracy; therefore, in this round, the depth of the primary search tree is set to (3-8) and the maximum number of leaf nodes per tree to (5-100). Early-stopping is also set to 10 to minimize overfitting. The next four rounds of adjustments are to balance reducing overfitting with maintaining accuracy after the first round. The second round adjusts `max_bin`, which is the number of segments used to discretize feature values ​​in the histogram algorithm. The value of `min_data_in_leaf` depends on the sample trees and the number of leaf nodes in the training data. Setting it larger avoids generating overly deep trees. The third round focuses on `feature_fraction` and `bagging_fraction`, which specify a certain proportion of samples to be drawn from all data for training, reducing variance at the cost of increased bias. Therefore, in this step, the sampling proportion is minimized while ensuring accuracy. The fourth round of adjustments tweaked lambda_l1 and lambda_l2, representing L1 and L2 regularization terms respectively, to filter features and control their impact on the model, preventing any single feature from having an excessive influence on the overall model. Finally, min_split_gain was adjusted, meaning that node splits only occur when the gain exceeds a given threshold, significantly limiting tree growth. After training, no significant decrease in accuracy or classification performance metrics was observed. This indicates that the model generalizes well beyond the training dataset, mitigating common learning bias problems.

[0127] After applying the LGBM model based on the gradient spanning tree algorithm, the predicted score for gene mutations ranges from 0 to 1. A score closer to 1 indicates that the model considers the mutation more likely to be pathogenic; a score closer to 0 indicates that the model considers the mutation more likely to be benign. By setting a threshold for the task, mutation pathogenicity can be classified; mutations above the threshold are considered pathogenic, and those below the threshold are considered benign. In this invention, the default threshold is set to 0.5.

[0128] 3. Use the MAGPIE model online.

[0129] The input single-base mutation location information is annotated with the refGene database to obtain basic mutation information containing mutation types. Then, multi-dimensional feature annotation is performed sequentially using the ANNOVAR annotation tool, SpliceAI splicing effect prediction software, functional effect related databases GDI, RVIS, and loF_score reference mutation information to obtain a mutation dataset containing multi-dimensional features of the annotation results. Next, Bayesian PCA is used to impute the annotation data of the mutation dataset containing multi-dimensional features of the annotation results. Then, the automatic engineering feature list obtained in step 2.3 is used to generate features from the annotation data imputed by Bayesian PCA. Then, the separation feature selection list obtained in step 2.3 is used to filter the data to obtain a mutation dataset containing all filtered features. Finally, the obtained mutation dataset containing all filtered features is passed through the Light GBM model (LGBM, Light Gradient Boosting Machine) based on the gradient spanning tree algorithm to obtain the gene mutation prediction score, ranging from 0 to 1.

[0130] experiment:

[0131] To quantitatively evaluate model performance, the MAGPIE model of this invention was compared with MutPred. [1] and REVEL [2] Two prediction tools were compared. The authors' recommended thresholds were used to distinguish between pathogenic and benign mutations: MutPred had a threshold of 0.79, and REVEL had a threshold of 0.5. In each tool's prediction results, mutations with scores greater than or equal to the threshold were considered pathogenic, while those with scores less than the threshold were considered benign.

[0132] Because the distribution of pathogenic and benign mutations varies across different test sets, several different metrics were used to evaluate the model's predictive performance in an attempt to mitigate the impact of class imbalance to some extent. These included accuracy, precision, recall, specificity, F1 score, G-means, and Matthew correlation coefficient (MCC). Compared to accuracy and F1 score, MCC considers all components of the confusion matrix, making it usable even with highly imbalanced datasets. Curves were plotted, and the area under the ROC curve (AUC) was calculated.

[0133]

[0134]

[0135]

[0136]

[0137]

[0138]

[0139]

[0140]

[0141] Where n is the total number of samples, n positive n is the number of positive samples. negative It is the number of negative samples, Rank i It represents the rank of the predicted value of the i-th sample among all samples. For samples with the same predicted value, their ranks are averaged.

[0142] 1. Comparative testing of the MAGPIE model, MutPred, and REVEL model.

[0143] Since most published models are designed to handle specific types of mutations or rely on certain key information, such as protein structure prediction, pathogenicity predictions cannot be obtained if mutations outside the model's scope exist in the test set or if the key information is unavailable. To more fairly evaluate the performance of these prediction tools, mutations for which no results were available for each tool were excluded, and then a confusion matrix was generated to calculate relevant metrics. Furthermore, the performance of these tools on unpredictable mutations was compared, classifying mutations without prediction results as benign.

[0144] To conduct benchmarking, MAGPIE was applied to the ClinVar training dataset, and its independent split dataset (hereinafter referred to as ClinVarTest) was used to evaluate the performance of the MAGPIE model, MutPred, and REVEL model.

[0145] The ClinVarTest dataset contains 4,356 pathogenic mutations and 4,310 benign mutations. Since we partitioned the dataset based on gene symbols, these mutations and their corresponding gene-level features were not visible to the model in the training dataset (2A, Table 1). MAGPIE outperformed all other classifiers on the benchmark, achieving the highest AUC score of 0.998 and AUPRC score of 0.998. Figure 2B (See Table 2). In comparison, the AUCs of REVEL and MutPred were 0.69 and 0.91, respectively. The above performance tests demonstrate that MAGPIE is a reliable and accurate pathogenicity prediction tool.

[0146] Table 1. Proportion of each type of mutation in the test set

[0147]

[0148]

[0149] Table 2. Performance of each model in the test set

[0150]

[0151] It is worth noting that most existing pathogenicity prediction tools can only predict a limited number of mutations, indicating that previous methods have insufficient generalization ability. For example, for published classifiers, REVEL and MutPred failed to predict 30% and 70% of candidate mutations, respectively. Figure 2D This resulted in a significant portion of the mutation outputs being missing. These missing values ​​were due to mutation types that the prediction tools could not predict. In contrast, MAGPIE successfully predicted the pathogenicity of all mutations in the dataset, filling in the missing values ​​that other methods could not predict. Figure 2C In other words, MAGPIE has a zero deletion rate for all evaluated exon mutations, including various mutation types such as nonsynonymous mutations. In summary, MAGPIE outperforms all previously published machine learning methods and state-of-the-art deep learning methods on this benchmark.

[0152] 2. Comparative testing of rare mutations

[0153] By screening for the frequency of mutations in the population, a rare mutation test set (ClinVarRare) was constructed to evaluate MAGPIE's ability to identify the pathogenicity of rare mutations, which is one of the major challenges in the practical analysis of genetic diseases.

[0154] The ClinVarRare test set includes rare pathogenic ClinVar mutations (AF < 0.01) and rare benign mutations from gnomAD (AF < 0.01), more closely resembling a real whole-exome sequencing dataset. For the rare mutation test set, MAGPIE achieved the best performance with an AUC of 0.995, followed by REVEL and MutPred. Previous computational tools predicted only a subset of mutations in the entire dataset, failing to predict the remainder because these tools are designed to classify specific mutation types. Therefore, we examined MAGPIE and found it predicted all 3897 mutations in the ClinVarRare test set, with the highest AUC of 0.977. Figure 2E In summary, MAGPIE outperforms the aforementioned deep learning and machine learning-based pathogen prediction models on the ClinVarTest and ClinVarRare test sets.

[0155] 3. Predictive ability for various mutations

[0156] MAGPIE demonstrates superior performance across a wide range of mutation types. MAGPIE can predict the clinical significance of marker mutations across various mutation types, whereas many existing methods are only applicable to specific mutation types. Figure 2G Furthermore, our method demonstrates robust performance across different mutation types, indicating its general applicability. Figure 2H ).

[0157] 4. MAGPIE's classification ability

[0158] We further analyzed the distribution of MAGPIE-predicted pathogenicity scores across different mutation types to assess MAGPIE's classification ability. To visualize the distribution of MAGPIE-predicted pathogenicity probabilities, we plotted the distribution of MAGPIE scores for different types of pathogenic and benign mutations. Figure 2H For pathogenic mutations, MAGPIE scores highly concentrated around 1, while for benign mutations, scores were close to 0 across all six mutation types. This indicates that our model can effectively distinguish between pathogenic and benign mutations. We further quantitatively analyzed MAGPIE's classification ability and evaluated its accuracy in identifying pathogenicity. The mean precision for all mutation types was 0.98, indicating that 98% of the predicted pathogenic mutations were consistent with their true labels. Particularly in the imbalanced dataset with termination loss and termination gain mutations, MAGPIE successfully identified both pathogenic and benign mutations. Figure 2G As shown, MAGPIE detected pathogenic mutations in both termination deletion and termination mutations, with a false positive rate of 0 (mean precision of 1.0). Therefore, MAGPIE can accurately distinguish between pathogenic and benign mutations in various mutation types.

[0159] 5. Threshold test

[0160] Different thresholds can affect model performance, especially with imbalanced data. In real-world clinical applications that use pathogenicity prediction tools to interpret unknown pathogenic mutations in individual patients, only a small fraction of the tens of thousands of candidate mutations invoked from loop exome sequencing (WES) data are defined as pathogenic. Therefore, pathogenicity prediction tasks may be assigned to models handling highly imbalanced datasets. Thus, we investigated the optimal threshold for MAGPIE under different data balance conditions. Our model's default parameter is 0.5 on balanced datasets such as the ClinVarTest dataset. We found that accuracy and F1 score are stable when the predicted probability is between 0.4 and 0.6. Figure 3 and Figure 4As shown. Furthermore, the MCC is relatively stable in the ClinVarTest dataset, confirming that 0.5 is a reliable threshold, as... Figure 3 and Figure 4 As shown.

[0161] To examine the impact of different ratios between the number of pathogenic mutations and the number of benign mutations, the ClinVarTest dataset was randomly divided into a series of subsets with different combinations. We randomly added different proportions of pathogenic mutations to all benign mutations in the ClinVarTest dataset to construct independent test subsets. In this case, the optimal threshold decreased from 0.95 to 0.45, as shown below. Figure 3 As shown. However, MCC, accuracy, and F1 score are relatively stable in these subsets, indicating that MAGPIE's performance is less affected by different thresholds, such as Figure 3 As shown. Furthermore, adding different proportions of benign mutations to all pathogenic mutations yielded similar results, such as... Figure 4 As shown.

[0162] 6. Pathogenic mutation testing for Mendelian genetic diseases

[0163] MAGPIE was applied to four genes (ATP7B, CFTR, FBN1, and LMNA), which contain a large number of known pathogenic mutations causing various Mendelian genetic disorders. For all four genes tested, namely ATP7B, CFTR, FBN1, and LMNA, MAGPIE detected the most known pathogenic mutations compared to the default parameters of other methods (REVEL and MutPred). Figure 5 As shown. On average, MAGPIE identified 96% of known pathogenic mutations. For example, for FBN1, the gene that causes Marfan syndrome, there are 1621 pathogenic mutations. MAGPIE predicted 1457 (90%) candidate mutations as pathogenic. Other tools correctly classified a lower proportion of mutations, below 70%. For CFTR (associated with cystic fibrosis), MAGPIE predicted 168 (95%) mutations as pathogenic. Furthermore, its performance on ATP7B and FBN1 was superior to other methods, demonstrating consistent performance in assessing pathogenicity despite the use of entirely different disease-associated genes such as... Figure 5 As shown. Therefore, MAGPIE has better performance in identifying pathogenic mutations that lead to Mendelian genetic diseases.

[0164] Finally, it should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

[0165] References

[0166] [1]Pejaver, V., Urresti, J., Lugo-Martinez, J., Pagel, K. A., Lin, G. N., Nam, H. J.,... & Radivojac, P. (2020). Inferring the molecular and phenotypic impact of amino acid variants with MutPred2. Nature communications, 11(1), 5918.

[0167] [2]Ioannidis, N. M., Rothstein, J. H., Pejaver, V., Middha, S., McDonnell, S. K., Baheti, S.,... & Sieh, W. (2016). REVEL: an ensemble method for predicting the pathogenicity of rare missense variants. The American Journal of Human Genetics, 99(4), 877 - 885。

Claims

1. A gene mutation prediction method generated by multi-modal annotation, characterized in that: The process includes annotating the input single-base mutation position information with mutation type to obtain basic mutation information containing mutation types, then performing multi-dimensional feature annotation to obtain a mutation dataset containing multi-dimensional features of the annotation results, then using Bayesian PCA to fill in the annotation data of the mutation dataset containing multi-dimensional features of the annotation results, then using an automatic feature engineering feature list to generate features from the annotation data filled in by Bayesian PCA, then using a separate feature selection list to filter the data, and finally using a gradient spanning tree algorithm to obtain the predicted score of gene mutations from the obtained mutation dataset containing all the filtered features. The gradient spanning tree algorithm includes the LGBM model. The LGBM model training process is as follows: a training set and a test set are constructed, and then after mutation type annotation, multi-dimensional feature annotation, and automatic feature engineering with separation feature selection, a training set and a test set containing all selected features are obtained. The LGBM model is trained in 5 rounds, with 5-fold cross-validation and stepwise parameter adjustment used in each round, thereby obtaining a trained LGBM model.

2. The gene mutation prediction method generated by multi-modal annotation according to claim 1, characterized in that: The mutation type annotation is performed using the refGene database. The multidimensional feature annotation includes annotation using the ANNOVAR annotation tool, SpliceAI splicing effect prediction software, and reference mutation information from the functional effects database.

3. The gene mutation prediction method generated by multi-modal annotation according to claim 2, characterized in that: The ANNOVAR annotation tool is used to annotate population-based features, amino acid biochemical changes, and conservation scores. For population-based features, this includes retrieving allele frequencies in various populations: whole exons (AF), raw allele frequencies (AF_raw), Africans (AF_afr), Latinos / mixed-race Americans (AF_amr), Ashkenazi Jews (AF_asj), East Asians (AF_eas), Finns (AF_fin), non-Finnish Europeans (AF_nfe), and other populations (AF_oth), as well as obtaining allele frequencies for different sexes from annotation information; For changes in the biochemical properties of amino acids, first check whether the mutation leads to changes in amino acids. If not, set all relevant features to 0. The physicochemical properties of each amino acid are stored in a matrix. The corresponding properties of amino acids are obtained by querying the matrix, and the difference in properties before and after mutation is used as the feature of the mutation. When a mutation affects multiple amino acids, the average value before and after the change is calculated separately. Information obtained from the BLOSUM100 matrix is ​​also used as features. Conserved fractional characteristics include the fractional characteristics of phastCons, phyloP, and SiPhy in primates, mammals, and vertebrates.

4. The gene mutation prediction method generated by multi-modal annotation according to claim 3, characterized in that: The SpliceAI splicing effect prediction software annotates splicing effect features to obtain the predicted effect of each mutant on splicing and information on splicing changes relative to the mutant position. The probability of the mutant’s effect on splicing is represented by the mutant’s change score, including receptor gain (DS_AG), receptor loss (DS_AL), donor gain (DS_DG), and donor loss (DS_DL). Information on the splicing change relative to the mutant’s location includes the location difference of receptor gain (DP_AG), receptor loss (DP_AL), donor gain (DP_DG), and donor loss (DP_DL). If SpliceAI does not annotate the mutant, it means that the mutant is near the end of the chromosome or the reference sequence is too long. Use 0 to fill the missing values ​​predicted by SpliceAI.

5. The gene mutation prediction method generated by multi-modal annotation according to claim 4, characterized in that: The functional effect database is used to annotate the mutation locations using the functional effect database for epigenomic features and gene damage index (GDI), residual mutation intolerance score (RVIS), gene intolerance score based on loss of function tool (LoFtool) (loF_score), and OMIM. The epigenomic features of each mutation were annotated in nine different cell lines using a 15-state ChromHMM model to capture the spatial context of interactions (chromatin states) between different chromatin markers. Functional effects include Gene Damage Index (GDI), Residual Mutation Intolerance Score (RVIS), Gene Intolerance Score based on Loss of Function Tool (LoFtool) (loF_score), mutant type, and annotations from the OMIM database; mutation type is used as a feature, including missense mutations, termination mutations, initiation deletion mutations, frameshift mutations, non-frameshift mutations, and termination deletion mutations; The mutation inheritance patterns were annotated using the OMIM database and classified into five different types, including autosomal recessive, autosomal dominant, X-linked recessive, X-linked dominant, and others. Then, one-hot encoding was performed on epigenomic features, mutant type features, and mutant inheritance pattern features.

6. The gene mutation prediction method generated by multi-modal annotation according to claim 5, characterized in that: The automatic feature engineering process with separation feature selection is specifically as follows: (1) Use the missing value estimation method based on Bayesian PCA to fill in the missing values ​​of features in a mutation dataset containing multidimensional features; (2) Automatic feature engineering: The openFE software package is used to perform mathematical transformations on the original features, such as logarithmic, square root, and exponential operations, and the original features are grouped according to the feature values ​​to generate new classification features. Then, the features obtained by feature selection using the default method defined in openFE are saved as the automatic engineering feature list. (3) Separation feature selection: For datasets with features obtained through automated feature engineering, an LGBM model is used for no more than 50 rounds of feature selection to reduce the number of features to 200 or fewer. After each round of feature selection, the importance of each feature is evaluated, and features with a relative importance score below a certain threshold are discarded. The features are then divided into two categories after 50 rounds of screening: the core feature set and the additional feature set. The feature combinations output by the separation feature selection algorithm are saved as the separation feature selection list.

7. The gene mutation prediction method generated by multi-modal annotation according to claim 6, characterized in that: The specific process of the separation feature selection algorithm includes: Input: Core feature set Additional feature set Final number of features ; ； Output: Feature combination with optimal performance .

8. The gene mutation prediction method generated by multi-modal annotation according to claim 7, characterized in that: The specific process for constructing the training and test sets is as follows: The ClinVar database was retrieved and filtered to remove mutations with conflicting and unknown labels. Obtain the gnomAD database and select allele frequencies. and Rare benign mutations were selected and screened: 5,000 mutations were randomly selected from each chromosome, and after being annotated with ANNOVAR, mutations that were missing in the mutation prediction features were filtered out. Then, 500 mutations from each chromosome were randomly selected and retained again. All mutations were retained from the Chr11 and ChrY chromosomes and qualified mutations from other chromosomes were randomly used to fill the gaps. The mutations corresponding to the unique Reference SNP IDs were obtained and filtered to construct the orthogonal validation set SwissProt; The mutations in the gnomAD library were divided into two subsets, gnomAD, at a ratio of 1.644:

1. Clinvar and gnomAD SwissProt Then gnomAD Clinvar Merged with Clinvar to form ClinVar gnomAD , gnomAD SwissProt Merged with SwissProt to form SwissProt gnomAD Use a splitting algorithm to split ClinVar gnomAD The datasets are split into two parts: dataset A and dataset B. and dataset They were used as the training set and the test set, respectively.

9. The gene mutation prediction method generated by multi-modal annotation according to claim 8, characterized in that: The specific process of the splitting algorithm is as follows: ； Where D is the dataset to be partitioned. , and For containing datasets , and The gene list, To be added The mutation rate threshold is 0.

9. During the algorithm execution occupy The real-time ratio.

Images