Method, system and device for analyzing pathogenicity of mutations based on target feature prediction

Abstract

The application relates to a target feature-based analysis method, system and device for predicting mutation pathogenicity. The method comprises the following steps: obtaining a mutation of a gene to be predicted; mapping the mutation of the gene to be predicted to a protein structure of the gene to be predicted to obtain a protein structure with the mutation mapped; extracting features of the protein structure with the mutation to obtain mutation protein structure features; selecting the mutation protein structure features to obtain a feature selection result; and inputting the feature selection result into a classification prediction model to obtain a classification result of the mutation being a pathogenic mutation or a benign mutation. The method aims to predict mutations based on target features in mutation protein structure features, and to explore high prediction capability and potential application value of the mutation pathogenicity.

Description

Technical Field

[0001] This invention relates to the field of gene data analysis, and more specifically, to a method, apparatus, system, computer-readable storage medium, and applications thereof for predicting the pathogenicity of mutations based on target features. Background Technology

[0002] The human genome contains approximately 3.16 billion base pairs, encoding about 30,000 genes, with a variation occurring on average once every 500-1000 base pairs. Within the human genome, protein-coding regions alone contain enormous differences between individuals; to date, 6.5 million missense variants have been observed. In fact, aside from the potentially lethal protein locations in the genomes of the 8 billion people living on Earth, every protein location is potentially susceptible to variation. Most clinically relevant variants have highly variable effects on protein structure and function, with only a small fraction being pathogenic.

[0003] The AlphaFold2 protein structure generation model expands the structural coverage of the human proteome from 17% to 98.5%, and this extensive and accurate structural information has unprecedented potential to help predict mutation effects. To assess the pathogenicity of gene mutations, many state-of-the-art mutation effect predictors have been developed, such as SIFT, REVEL, and EVE. Given the close relationship between protein structure and function, the structural background of mutations represents promising information independent of mutation frequency and evolutionary conservation. The degree of residue burial or exposure in the 3D structure is crucial for protein folding and stability. Therefore, specific structural features of mutant proteins are powerful predictors of mutation pathogenicity. Summary of the Invention

[0004] The method of this invention integrates the information of the gene to be predicted and the predicted protein structure information, and deeply mines the variation features in the gene data based on the target features, thereby predicting the pathogenicity of the mutation of the gene to be predicted, exploring its high predictive ability and potential application value for the pathogenicity of mutations, and solving related life science problems.

[0005] This application discloses a method for predicting the pathogenicity of mutations based on target features, including:

[0006] Obtain the mutation of the gene to be predicted, wherein the gene to be predicted includes wild-type genes;

[0007] The gene to be predicted is input into the protein structure generation model to obtain the protein structure of the gene to be predicted.

[0008] The mutation of the gene to be predicted is mapped onto the protein structure of the gene to be predicted, and the protein structure mapped with the mutation is obtained.

[0009] Feature extraction is performed on the mapped mutated protein structure to obtain the mutated protein structure features;

[0010] Feature selection was performed on the structural features of the mutant protein to obtain the feature selection results;

[0011] The feature selection results are input into the classification prediction model to obtain the classification results of whether the mutation is a pathogenic mutation or a benign mutation.

[0012] Furthermore, obtaining the mutation of the gene to be predicted includes preprocessing the mutation of the gene to be predicted; optionally, the preprocessing includes using an algorithm to predict the splicing effect of the mutation and excluding mutations with scores exceeding a threshold; preferably, the processing uses the SpliceAI algorithm to predict the splicing effect of the mutation and exclude mutations with scores greater than 0.5.

[0013] Optionally, the preprocessing may further include excluding mutations with a residue count exceeding a threshold;

[0014] Optionally, the preprocessing includes excluding mutations that cannot be mapped to the structure due to isomer inconsistencies;

[0015] Optionally, the pretreatment includes excluding mutations with clinical significance that are “conflicting in their pathogenicity explanations”.

[0016] Furthermore, the protein structure generation model includes one or more of the following models: AlphaFold, AlphaFold2, and ProteoGAN; preferably, the protein structure generation model is AlphaFold2.

[0017] Furthermore, the classification prediction model obtains the classification result of mutation as pathogenic or benign mutation by one or more of the target features among the following mutant protein structural features, which include: protein-level features, residue-level features, and mutation-level features.

[0018] Optionally, the protein-level features include thermodynamic features and protein volume features, the residue-level features include secondary structure-related features and RSA, and the mutation-level features include thermodynamic features and features derived from empirical rules for determining whether a mutation has a significant effect on protein structure.

[0019] Preferably, the thermodynamic characteristics include total unfolding energy and its van der Waals collisions, hydrogen bond energy, and side chain entropy;

[0020] Preferably, the protein volume characteristics include total volume, void volume, and van der Waals volume;

[0021] Optionally, the secondary structure-related features are generated using features assigned by the DSSP program: DSSP(H), DSSP(E), DSSP(G), DSSP(I), DSSP(T), DSSP(S), DSSP(C);

[0022] Optionally, the target features in the mutant protein structural features include total unfolding energy, van der Waals collision, hydrogen bond energy, side chain entropy, ionization energy, disulfide bond breakage, total protein volume, void volume, van der Waals volume, RSA, ΔΔG, ΔG, DSSP(H), DSSP(E), DSSP(G), DSSP(I), DSSP(T), DSSP(S), and DSSP(C).

[0023] Preferably, the target features in the mutant protein structural features include RSA, ΔΔG, side chain entropy, ionization energy, and van der Waals collisions.

[0024] Furthermore, the classification prediction model is implemented based on machine learning algorithms, and the processing includes feature selection, feature fusion, feature calculation, and model optimization; the machine learning algorithms include any one or more of the following algorithms: GBDT, GBM, SVM, RF, Adaboost, and Apriori algorithms; the model optimization includes one or more of the following methods: steepest descent method, Newton method, and quasi-Newton method.

[0025] Furthermore, the step of inputting the feature selection result into the classification prediction model to obtain the classification result of the mutation as a pathogenic mutation or a benign mutation also includes:

[0026] The feature selection results are input into the classification prediction model to obtain the classification results of whether the mutation is a pathogenic mutation or a benign mutation, and the first classification result is output.

[0027] The mutation of the gene to be predicted is input into the predictor to obtain the variation effect score of the gene to be predicted, and a second classification result of pathogenic mutation or benign mutation is obtained based on the variation effect score.

[0028] The first classification result and the second classification result are weighted and fused to obtain the fused classification result of whether the mutation is a pathogenic mutation or a benign mutation;

[0029] The predictor performs mutation effect prediction on the mutation of the gene to be predicted, and obtains the mutation effect score of the gene to be predicted.

[0030] Optionally, the predictor includes any one or more of the following software: Enformer, SIFT, PROVEAN, EVE, DEOGEN2, MUTPRED, Mutation Assessor, PolyPhen2, ClinPred, and REVEL.

[0031] Preferably, the predictor is PROVEAN, EVE, DEOGEN2, or MUTPRED;

[0032] Optionally, the variation effect score of the gene to be predicted includes one or more of the following scores: sequence conservation score, allele mutation frequency score, missense mutation score, frameshift mutation score, nonsense mutation score, and insertion / deletion mutation score.

[0033] Preferably, the variation effect score of the gene to be predicted includes sequence conservation score, allele mutation frequency score, and missense mutation score;

[0034] The weighted fusion adopts one or more of the following algorithms: Boosting, XGBoost, AdaBoost and Voting fusion algorithms.

[0035] A pathogenicity analysis device for predicting mutations based on target features, the device comprising:

[0036] Memory and processor;

[0037] The memory is used to store program instructions;

[0038] The processor is used to call program instructions, which, when executed, are used to perform the aforementioned pathogenicity analysis method based on target feature prediction of mutations.

[0039] A pathogenicity analysis system based on target feature prediction of mutations includes:

[0040] The acquisition module is used to acquire the gene mutations to be predicted;

[0041] The structure generation module maps the acquired mutations of the gene to be predicted onto the protein structure of the gene to be predicted, thereby obtaining the protein structure mapped with the mutations.

[0042] The structural feature extraction module extracts features from the protein structure mapped with mutations to obtain the structural features of the mutated protein;

[0043] The structural feature selection module is used to select structural features of the mutant protein to obtain feature selection results;

[0044] The classification module inputs the feature selection results into the classification prediction model to obtain the classification result of whether the mutation is a pathogenic mutation or a benign mutation;

[0045] The classification prediction model obtains the classification result of mutation as pathogenic or benign mutation by one or more of the target features in the following mutant protein structural features: protein level features, residue level features, and mutation level features.

[0046] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described pathogenicity analysis method based on predicting mutations using target features.

[0047] The association between the aforementioned structural features and the pathogenicity of variants was assessed based on genotype data types. For dichotomous features, contingency tables were constructed, and the chi-square test was used to determine whether an association existed between the two dichotomous variables. For continuous features, logistic regression analysis was used to examine the association between the features and the pathogenicity of variants. The strength of the association between features and pathogenicity of variants was quantified using the odds ratio (OR) and 95% confidence interval (CI). All statistical analyses and data visualizations were performed using R software packages: h2o, caret, pROC, forestplot, ggpubr, ggsci, viridis, and cutpointr, where a p-value <0.05 was considered statistically significant.

[0048] The above-mentioned equipment is used in the selection of auxiliary mutation pathogenicity analysis schemes; optionally, the scheme includes mutation pathogenicity analysis affected by changes in the type and number of the mutant protein structural features; optionally, the mutant protein structural features include mutant protein structural features at three levels: protein level features, residue level features, and mutation level features; preferably, the target features of the mutant protein structural features include RSA, ΔΔG, side chain entropy, ionization energy, and van der Waals collisions;

[0049] The above-mentioned equipment is used in predicting the pathogenicity of mutations; optionally, the application has a positive impact on research on diabetes, blood vessels, bones, brain function, and anti-aging.

[0050] The above-mentioned device is used in gene mutation classification or prediction of mutation attributes; optionally, gene mutation classification includes classification results of mutations as pathogenic or benign mutations; optionally, attributes include the gene family to which the mutation of the gene to be predicted belongs, the gene family includes ion channels and collagen, and when the gene family to which the mutation of the gene to be predicted belongs is an ion channel gene, the RSA in the structural features of the mutant protein is selected for prediction.

[0051] The above-mentioned system is used in the diagnosis of the occurrence and development of pathogenicity of mutations; optionally, the occurrence and development of pathogenicity of mutations are related to changes in the type and number of gene mutation characteristics;

[0052] The above-mentioned system is applied in predicting the pathogenicity of mutations; optionally, the application has a positive impact on research on diabetes, blood vessels, bones, brain function, and anti-aging.

[0053] The above system is applied to gene mutation classification or prediction of mutation attributes; optionally, gene mutation classification includes classification results of mutations as pathogenic or benign mutations; optionally, attributes include the gene family to which the mutation of the gene to be predicted belongs, the gene family includes ion channels and collagen, and when the gene family to which the mutation of the gene to be predicted belongs is an ion channel gene, the RSA in the structural features of the mutant protein is selected for prediction.

[0054] This invention trains high-quality mutation data with clinical significance using machine learning algorithms. Based on the structural features of mutant proteins, it identifies key regions susceptible to genetic variations and locks specific structural features of mutant proteins to better reflect the spatial clustering of pathogenic variations, which can significantly improve performance. It explores and determines the salient feature clustering effect in the structural features of pathogenic mutant proteins, which is highly innovative in the field of life sciences and will have a beneficial promoting effect on the pathogenicity analysis of gene data.

[0055] Advantages of this application:

[0056] 1. This application innovatively discloses a novel method for analyzing the pathogenicity of mutated genes. This method is based on the deep mining of target features to uncover the life laws hidden behind gene data. It maps the mutation of the gene to be predicted onto the protein structure to obtain the protein structure mapped with the mutation. Feature extraction is performed to obtain the structural features of the mutated protein. Five important features of the structural features of the mutated protein are selected, namely RSA, ΔΔG, side chain entropy, ionization energy, and van der Waals collision, to quantify the pathogenicity classification results, thereby improving the accuracy and depth of data analysis.

[0057] 2. This application innovatively discloses a method for obtaining feature selection results of the gene to be predicted based on key target features. The feature selection results of the mutation of the gene to be predicted are input into a classification prediction model to obtain classification results of whether the mutation is a pathogenic mutation or a benign mutation, and output the first classification result. Then, taking into account the variation effect score of the gene to be predicted, a second classification result of whether the mutation is a pathogenic mutation or a benign mutation is obtained based on the variation effect score. Finally, the first classification result and the second classification result are weighted and fused to obtain the fused classification result of whether the mutation is a pathogenic mutation or a benign mutation.

[0058] 3. This application creatively discloses a pathogenicity analysis device and system based on target feature prediction of mutations. By deeply interpreting gene data through key target features in the structural features of mutant proteins, and combining it with other VEPs to analyze the pathogenicity of the gene mutation to be predicted, it can better reflect the spatial clustering of pathogenic variants, significantly improve performance, and make this application more accurately applied to the auxiliary diagnosis of disease occurrence and development related to gene data, and the auxiliary selection of treatment plans. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a schematic flowchart illustrating the prediction of gene mutation pathogenicity based on target features, provided in an embodiment of the present invention.

[0061] Figure 2 This is a SIGMA score map for predicting mutation pathogenicity based on target features, provided in an embodiment of the present invention.

[0062] Figure 3 This is a protein structure feature map based on AlphaFold2 that maps to mutations, provided in an embodiment of the present invention.

[0063] Figure 4 This is an assessment diagram of the importance of target features in predicting the pathogenicity of variants, provided in an embodiment of the present invention.

[0064] Figure 5 This is a graph illustrating the SIGMA prediction of mutation pathogenicity based on target features, provided in an embodiment of the present invention.

[0065] Figure 6 This is a graph illustrating the SIGMA+ prediction of mutation pathogenicity based on target features, provided in this embodiment of the invention.

[0066] Figure 7 This is a SIGMA+ score map for predicting mutation pathogenicity based on target features, provided in an embodiment of the present invention.

[0067] Figure 8 This is a schematic diagram of a pathogenicity analysis device based on target feature prediction of mutations provided in an embodiment of the present invention. Detailed Implementation

[0068] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0069] In some processes described in the specification, claims, and accompanying drawings of this invention, multiple operations are included in a specific order. However, it should be clearly understood that these operations may not be performed in the order they appear herein, or they may be performed in parallel. The operation numbers, such as S101, S102, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be performed sequentially or in parallel.

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Figure 1 This is a schematic flowchart illustrating the prediction of gene mutation pathogenicity based on target features provided in an embodiment of the present invention. Specifically, the method includes the following steps:

[0072] S101: Obtain the mutation of the gene to be predicted.

[0073] In one embodiment, the gene to be predicted includes a wild-type gene, and obtaining mutations in the gene to be predicted includes publicly available genetic data. Optionally, the mutation information of the gene to be predicted also includes genetic information of the sample obtained by high-throughput sequencing. Optionally, the mutations in the gene to be predicted include one or more of the following datasets: dbSNP, dbVar, gnomAD, RefSeq, and ExAC, where dbSNP, dbVar, and RefSeq are all from the ClinVar database. A total of 27,165 benign and 22,957 pathogenic missense variants were retrieved from the gnomAD and ClinVar databases. For ClinVar mutations, single nucleotide missense mutations with an audit status of at least one star were retained (expert panel audit: provided standard, multiple submitters, no conflict; provided standard, single submitter).

[0074] Furthermore, step S101 also includes preprocessing the mutation data. Optionally, the preprocessing includes using an algorithm to predict the impact of mutations on splicing and excluding mutations with scores exceeding a threshold; preferably, the SpliceAI algorithm is used to predict the impact of mutations on splicing and excluding mutations with SpliceAI scores greater than 0.5.

[0075] SpliceAI predicts splicing changes caused by single nucleotide variations, accurately predicting splicing sites (location and probability of abnormal splicing) from any mRNA precursor sequence, thus enabling the prediction of hidden splicing caused by mutations in non-coding RNA regions.

[0076] Optionally, the pretreatment also includes excluding mutations with more than a threshold number of residues; for example, proteins with more than 2,700 residues are excluded.

[0077] Optionally, preprocessing includes excluding mutations that cannot be mapped to the structure due to isomer inconsistencies;

[0078] Optionally, pretreatment may include excluding mutations with clinical significance that are “conflicting in their pathogenicity explanations”.

[0079] S102: Map the mutation of the gene to be predicted onto the protein structure of the gene to be predicted, and obtain the protein structure mapped with the mutation.

[0080] In one embodiment, the protein structure of the gene to be predicted is obtained by inputting the gene to be predicted into a protein structure generation model. Preferably, the gene to be predicted is a wild-type gene. The protein structure generation model refers to a model or software capable of generating the protein structure of a gene or polypeptide based on its amino acid sequence. Optionally, the protein structure generation model includes one or more of the following models: VAE, AlphaFold, AlphaFold2, and ProteoGAN; preferably, the protein structure generation model is AlphaFold2.

[0081] VAE (Variational Autoencoder) is a generative model based on joint probabilities.

[0082] AlphaFold accelerated and enabled large-scale discoveries, including deciphering the structure of the nuclear pore complex. AlphaFold predictions, with very high confidence, provided a nine-helix topology, offering clues about protein function that are crucial for unraveling the causes of rare genetic diseases.

[0083] AlphaFold2 discovered a novel protein-gated mechanism, glucose-6-phosphatase, an enzyme that identifies binding sites for inhibiting enzymes, and a transmembrane protein, Wolframin, located in the ER. It improved the accuracy of predictions to the atomic level, more quickly and accurately identifying enzyme active sites, with 350,000 predicted sites.

[0084] ProteoGAN, a conditional generative adversarial network, outperforms classic and recent deep learning baselines in protein sequence generation. It primarily expands protein screening with candidates that are farther from the known sequence space than previously possible, but is more likely to have functionality with relatively novel candidates than other methods.

[0085] In one specific embodiment, preferably, the protein structure of the gene to be predicted is obtained by inputting the wild-type gene into the AlphaFold2 protein structure generation model. The protein structure dataset in AlphaFold2 includes the AlphaFold Protein Structure Database (AlphaFoldDB, https: / / alphafold.ebi.ac.uk / ), containing approximately 365,000 structure predictions. AlphaFold2 is used to provide these proteins with 1400 amino acid fragments; proteins exceeding 2700 residues are excluded. All mutations should be mapped to the predicted protein 3D structure; mutations that cannot be mapped to a structure due to isoform inconsistencies are excluded.

[0086] S103: Extract features from the protein structure mapped with mutations to obtain the structural features of the mutated protein.

[0087] In one embodiment, the structural features of the mutant protein include protein-level features, residue-level features, and mutation-level features; optionally, feature extraction is performed using FoldX.

[0088] FoldX can rapidly and quantitatively assess the importance of interactions to the stability of proteins and protein complexes using complete atomic descriptions of protein structures. Among various energy functions currently available, FoldX achieves a good balance between accuracy and speed, making it readily applicable to protein design algorithms and the prediction of protein structure and folding pathways where fast and accurate energy functions are required.

[0089] In one specific embodiment, the mutation of the gene to be predicted is mapped to the mutated protein structure by generating the wild-type protein structure based on AlphaFold2. Then, 57 mutant protein structural features are extracted from it, including three levels of features: protein level features, residue level features, and mutation level features. These 57 mutant protein structural features include thermodynamic features, protein volume features, secondary structure related features, relative solvent accessibility features of mutated residues, and empirical rule features that can determine whether the mutation has a significant impact on the protein structure. Among them, pathogenic mutations show greater changes in protein stability than benign mutations.

[0090] Optionally, protein-level features include 16 thermodynamic features and 3 protein volume features; the thermodynamic features include 15 components such as total unfolding energy and its van der Waals collisions, hydrogen bond energy, and side chain entropy; the protein volume features include total volume, interstitial volume, and van der Waals volume.

[0091] Optionally, residue-level features include eight secondary structure-related features and the relative solvent accessibility (RSA) of the mutated residue, used to characterize the structural background of the mutated residue. For each mutated residue, its secondary structure is assigned using the DSSP (protein secondary structure dictionary) program, generating features corresponding to eight secondary structure types: DSSP(H), DSSP(E), DSSP(G), DSSP(I), DSSP(T), DSSP(S), and DSSP(C).

[0092] Optionally, mutation level features include 16 thermodynamic features describing changes in protein stability after mutation and 13 features derived from empirical rules for determining whether a mutation has a significant effect on protein structure; the 16 thermodynamic features include the free energy difference (ΔΔG) between the mutant protein and its 15 components; the 13 features derived from empirical rules for determining whether a mutation has a significant effect on protein structure are a combination of the structural background and physicochemical properties of the mutant / mutant residues, including glycine bending, proline in the α-helix, substitution of residues in the α-helix by proline / glycine, cysteine ​​residues, energy charge loss, energy hydrophobicity replacement, outlier substitution, disulfide bond breakage, buried salt bridge breakage, buried proline, buried residues, energy displacement of glycine, protein unfolding free energy, and the unfolding free energy difference between mutant proteins.

[0093] Figure 3 This is a structural feature diagram of a mutant protein provided in an embodiment of the present invention, demonstrating the relationship between the structural features of the mutant protein and its pathogenicity, particularly the significant correlation between the relative solvent accessibility (RSA) and free energy difference (ΔΔG) between the mutant protein and its pathogenicity. Figure 3 As shown in A, the target features of pathogenic variants were significantly lower than those of benign variants (P<2.2e-16, Mann-Whitney U test), which is consistent with the fact that most proteins are less tolerant of hidden mutations than exposed mutations. Figure 3 As shown in B, ΔΔG measures the effect of a single amino acid substitution on protein stability; pathogenic variants showed greater changes in protein stability than benign variants (P < 2.2e-16, Mann-Whitney U test). Figure 3As shown in Figure C, as expected, disulfide bond breakage has the highest positive predictive value among all structural features (odds ratio [OR] = 93.8, 95% confidence interval [CI] = 44.5–198, P < 2.2e-16, Pearson chi-square test). Almost all (98.72%) missense mutations that break disulfide bonds are pathogenic, supporting the important role of disulfide bonds in protein function. Figure 3 As shown in D, the association between the type of secondary structure in which the mutation occurs and its pathogenicity is consistent with prior knowledge. Mutations in the circular or irregular extensions of proteins (Dictionary of Protein Secondary Structures [DSSP]-C) are often benign (OR = 0.32, 95% CI = 0.31–0.34, P < 2.2e-16, Pearson's chi-squared test). In contrast, mutations in regular secondary structures are often pathogenic, especially mutations in α-helices (DSSP-H; OR = 1.73, 95% CI = 1.66–1.79, P < 2.2e-16, Pearson chi-squared test) or β-sheets (DSSPE; OR = 1.97, 95% CI = 1.87–2.08, P < 2.2e-16, Pearson chi-squared test).

[0094] S104: Feature selection is performed on the structural features of the mutant protein to obtain the feature selection results.

[0095] In one instance, the feature selection result refers to the target features in the structural features of the mutant protein, including total unfolding energy, van der Waals collision, hydrogen bond energy, side chain entropy, ionization energy, disulfide bond breakage, total protein volume, void volume, van der Waals volume, RSA, ΔΔG, ΔG, DSSP(H), DSSP(E), DSSP(G), DSSP(I), DSSP(T), DSSP(S), and DSSP(C).

[0096] like Figure 4 A map assessing the importance of target features in predicting the pathogenicity of variants is presented. Preferred target features in mutant protein structural features include RSA, ΔΔG, side chain entropy, ionization energy, and van der Waals collisions. Figure 4 A shows the ten most important features affecting the discriminative power of target features. Among them, the residue-level feature RSA contributes the most to the discriminative power, followed by two mutation-level features (ΔΔG and ΔVander). Seven of these ten most important features are protein-level features, demonstrating the important role of protein stability in predicting the pathogenicity of variants. Figure 4 B and Figure 4C represents the gene set enrichment analysis (GSEA) plots for the ion channel gene family (B) and the collagen gene family (C), respectively. Notably, RSA alone can predict the pathogenicity of mutations in 76 out of 200 genes with an accuracy of >0.8 (with at least 10 pathogenic variants and 10 benign variants used for evaluation), especially showing good predictive performance for ion channel gene mutations.

[0097] In addition, from Figure 4 The results show that RSA, relative solvent accessibility, has the highest accuracy in classifying ion channel genes (P = 8.67e-04); using the target feature alone is unsatisfactory for 11% of the genes (21 out of 200 genes, accuracy < 0.5), and 11% of the genes represent the collagen gene family (P = 1.99e-06). Figure 4 C). Where ΔG is the difference in unfolding free energy between the mutant protein and the mutant protein; ΔG is the unfolding free energy of the protein; VdW represents van der Waals.

[0098] S105: Input the feature selection results into the classification prediction model to obtain the classification results of whether the mutation is a pathogenic mutation or a benign mutation.

[0099] In a specific example, the classification prediction model obtains the classification result of whether the mutation is pathogenic or benign by using one or more of the target features from the following mutant protein structural features: protein-level features, residue-level features, and mutation-level features. For example... Figure 2 As shown, the SIGMA score map for predicting pathogenicity based on target features is presented, which shows the classification results of mutations as pathogenic or benign mutations when the classification prediction model is the SIGMA model.

[0100] Among them, the SIGMA (Structure-Informed Germline Missense mutation Assessor) model is developed based on the close relationship between protein structure and function. It mainly uses machine learning algorithms such as gradient enhancement machines to evaluate the impact of missense variants in the context of protein structure.

[0101] Furthermore, the classification prediction model is determined as follows:

[0102] Obtain the mutation dataset, which contains labeling information;

[0103] By mapping mutation data onto protein structures, we obtain protein structures mapped with mutations.

[0104] Feature extraction is performed on the protein structure mapped with mutations to obtain the structural features of the mutated protein;

[0105] Machine learning methods were used to select features from the structural feature data of mutant proteins. Target features in the structural features of mutant proteins were selected as the feature selection results. The target features in the structural features of mutant proteins include protein-level features, residue-level features, and mutation-level features.

[0106] A classification and prediction model is established by selecting target features from the structural features of mutant proteins.

[0107] Specifically, the construction process of the classification prediction model SIGMA is as follows: A mutation dataset containing labeling information is obtained; the mutation data is mapped onto protein structures to obtain mapped mutated protein structures; feature extraction is performed on the mapped mutated protein structures to obtain mutated protein structural features; machine learning algorithms are used to process the mutated protein structural features, and target features are selected from these features as the feature selection results to obtain the predicted classification results; based on the predicted classification results and the actual results, the machine learning algorithm is optimized to obtain the final mutation classification results of the genes to be predicted, and the trained SIGMA model is output.

[0108] Optionally, the mutation dataset includes one or more of the following datasets: gnomAD, HumVar, ExoVar, PredictSNP, VariBench, SwissVar, Humsavar, and ClinVar.

[0109] In a specific example, the SIGMA model is constructed using a final labeled dataset containing 27,165 benign (negative) and 22,957 pathogenic (positive) mutations as the "gold standard" dataset. This dataset is divided into 80% for training and 20% for testing. The gnomAD database retains 27,928 mutations for model evaluation. Training the SIGMA model can be as follows: obtain a mutation dataset containing labeling information (e.g., pathogenic / potentially pathogenic mutations are labeled positive, while benign / potentially benign mutations are labeled negative); repeat the protein structure generation and feature extraction processes described above to obtain a feature set; input the feature set into the SIGMA model to obtain the classification results of the samples; calculate the loss between the classification results and the true values; and optimize the machine learning algorithm through backpropagation, loss function, optimizer, etc., to obtain the trained SIGMA model.

[0110] Furthermore, the labeling information was defined as follows: pathogenic / potentially pathogenic variants were labeled positive, while benign / potentially benign variants were labeled negative. For gnomAD mutations, preprocessing was performed using SpliceAI to predict the gnomAD database (for deep intron regions: distance from the exon-intron boundary >50 nt). Due to the difficulty of deep intron prediction, only 56% deletions were observed when the threshold was set to 0.8. Therefore, the impact of mutations on splicing was predicted, and mutations with SpliceAI scores greater than 0.5 were excluded. Mutations introducing mystery splicing sites primarily affected mRNA splicing rather than protein structure, and selected common missense mutations (the highest allele frequency in all populations with gnomAD > 0.05) were labeled negative. In addition, the GOF / LOF database labeled 193 gain-of-function (GoF) and 921 loss-of-function (LoF) pathogenic missense variants.

[0111] Furthermore, obtaining the mutation dataset also includes preprocessing the mutation dataset. Optionally, preprocessing includes using an algorithm to predict the impact of mutations on splicing and excluding mutations with scores exceeding a threshold; preferably, it involves using the SpliceAI algorithm to predict the impact of mutations on splicing and excluding mutations with SpliceAI scores greater than 0.5.

[0112] Optionally, the preprocessing also includes excluding mutations with more than 2700 residues.

[0113] Optionally, preprocessing includes excluding mutations that cannot be mapped to the structure due to isomer inconsistencies;

[0114] Optionally, pretreatment may include excluding mutations with clinical significance that are “conflicting in their pathogenicity explanations”.

[0115] Furthermore, the machine learning algorithms include any one or more of the following algorithms: GBDT, GBM, SVM, RF, Adaboost, and Apriori; the optimized machine learning algorithms include any one or more of the following methods: steepest descent method, Newton method, and quasi-Newton method.

[0116] In one embodiment, Figure 1 Step S105 further includes:

[0117] The feature selection results are input into the classification prediction model to obtain the classification result of whether the mutation is a pathogenic mutation or a benign mutation, which is used as the first classification result;

[0118] The mutation of the gene to be predicted is input into the predictor to obtain the variant effect score of the gene to be predicted. Based on the variant effect score, a second classification result is obtained as whether the mutation is a pathogenic mutation or a benign mutation.

[0119] The first classification result and the second classification result are weighted and fused to obtain the fused classification result of whether the mutation is a pathogenic mutation or a benign mutation;

[0120] Optionally, the weighted fusion may employ one or more of the following algorithms: Boosting, XGBoost, AdaBoost, and Voting ensemble algorithms.

[0121] Boosting is a mainstream and representative technique in ensemble learning in machine learning. It often beats deep neural networks' XGBoost in many data analysis competitions and is an efficient implementation of the GradientBoost algorithm in the Boosting family.

[0122] Voting ensemble algorithms are one of the most common combination strategies for classification problems in ensemble learning. The basic idea is to select the class that outputs the most among all machine learning algorithms.

[0123] Optionally, the predictor may include any one or more of the following software: Enformer, PROVEAN, EVE, DEOGEN2, MUTPRED, MutationAssessor, SIFT, ClinPred, PolyPhen2, REVEL; preferably, the predictor may be PROVEAN, EVE, DEOGEN2, or MUTPRED.

[0124] PROVEAN is a tool for predicting whether protein sequence variations affect protein function. Based on evolutionary conservation, neural network models, and the BLOSUM62 amino acid substitution scoring matrix, it identifies whether non-synonymous mutations or InDels affect protein biological function. Prediction scores range from -14 to 14, with a threshold of -2.5. Scores of -14 to 2.5 are predicted as Deleterious, and scores of -2.5 to 14 are predicted as Neutral. Lower scores indicate more harmful mutations, and vice versa.

[0125] EVE estimates the likelihood of each single amino acid variation being benign or pathogenic, understands the pathogenicity of human missense variations from the distribution of sequence variations across species, outperforms other computational prediction models in predicting clinical outcomes, and scores as high as or better than the current gold standard high-throughput experiments for testing the impact of mutations on biological function.

[0126] DEOGEN2 contains heterogeneous information about the molecular effects of mutations, the domains involved, gene relevance, and the interactions they participate in. This extensive contextual information is non-linearly mapped to a single harmfulness score for each mutation.

[0127] MUTPRED's function is to predict the pathogenicity and molecular mechanisms of amino acid substitutions, using Fasta-formatted amino acid sequences as the primary input. MUTPRED is a collection of machine learning tools that can predict the pathogenicity of protein-coding mutations to infer the molecular mechanisms of diseases.

[0128] REVEL is an ensemble method for predicting missense mutations. It combines the prediction results of multiple software programs (MutPred, FATHMM, VEST, PolyPhen, SIFT, PROVEAN, Mutation Assessor, Mutation Taster, LRT, GERP, SiPhy, phyloP, and phastCons) and uses a random forest algorithm, which shows good prediction results for rare missense mutations. The prediction score for a single mutation ranges from 0 to 1, and the threshold can be set according to the required sensitivity and specificity.

[0129] In one embodiment, the mutated gene comprised 20,047 pathogenic and 20,148 benign variants, and also included 27,928 mutations in six proteins characterized by the DMS experimental system, using 27 computational VEPs, such as EVE screening and DEOGEN2. The EVE score for each mutation was obtained from https: / / evemodel.org / . Predictions from other VEPs for each mutation were retrieved from the dbNSFP database (version 4.1a). These VEPs were divided into two categories: single predictors independent of other VEPs (n=16) and meta-predictors that integrated the results of other VEPs into input features (n=11).

[0130] Figure 5 This is an analysis diagram of the predictive efficacy of the SIGMA model based on target feature prediction in predicting the pathogenicity of mutations, provided in an embodiment of the present invention. The SIGMA model uses a gradient boosting machine (GBM) for training a classifier to distinguish between pathogenic and benign mutations, and uses grid search for parameter tuning. Specifically, as shown... Figure 5 The performance evaluation plot shown uses 40,195 mutations with structural features as the training dataset and employs a SIGMA model based on gradient boosting machine (GBM) to predict the pathogenicity of missense mutations. Figure 5 As shown in Figure A, the SIGMA score distribution for 20,047 pathogenic and 20,148 benign variants in the training set, where the SIGMA scores (from 0 to 1) are quantitative calculations of variant potential pathogenicity using out-of-fold prediction, exhibits a bimodal distribution, indicating that SIGMA has strong discriminative power. For the training set, Figure 5The area under the ROC curve (AUC) of the receiver operating characteristic shown in B is 0.944 (95% CI = 0.942–0.946). Consistently, as... Figure 5 As shown in C, high prediction accuracy was achieved on the test set (AUC = 0.933, 95% CI = 0.928–0.938). Figure 5 As shown in D, when comparing the AUC of SIGMA and 16 individual variation effect predictors (VEPs) using the test set, SIGMA outperformed all individual VEPs, with AUCs ranging from 0.779 (FATHMM) to 0.929 (MutPred). Figure 5 E represents the correlation between VEP and Deep Mutation Scan (DMS) measurements, where Spearman's correlation is calculated between the functional score of the DMS experiment, the predicted score of SIGMA, and 16 individual VEPs. The closest competitor is DEOGEN2 (overall rho = 0.387, Spearman correlation analysis); Figure 5 E), which contains extensive heterogeneous information, saw its recently developed predictor, EVE, improve its performance ranking from sixth on the labeled dataset to third on the DMS dataset. Figure 5 E).

[0131] The above method is applied to the selection of auxiliary gene mutation pathogenicity analysis schemes. Based on the gene information of the mutation to be predicted, the classification result of whether the mutation is pathogenic can be predicted. Specifically, such as... Figure 6 This is a SIGMA+ prediction chart of mutation pathogenicity effect based on target features provided in an embodiment of the present invention. Figure 7 The SIGMA+ score map for predicting mutation pathogenicity based on target features provided in this embodiment of the invention has good classification performance. Here, SIGMA+ is a combination of all five predictors, namely SIGMA, DEOGEN2, EVE, PROVEAN, and MutPred. Figure 6 A demonstrates the potential applications of combining SIGMA with other predictors. For example, combining SIGMA with four individual VEPs with high performance, namely DEOGEN2, EVE, PROVEAN, and MutPred, yields a more comprehensive predictor of variant pathogenicity. Various combinations of these five predictors enhance predictive performance (AUC range 0.931–0.957). Figure 6 B specifically demonstrates the high predictive power of SIGMA+ for the pathogenicity of variants (AUC = 0.964), and can significantly distinguish between benign and pathogenic variants. Figure 6 C demonstrates the correlation between VEP and deep mutation scanning (DMS) measurements. From Figure 6The results show that SIGMA contributes the most to the combination, while DEOGEN2 contributes the least, indicating that SIGMA has strong discriminative power. Therefore, mining additional information sources may be more beneficial than iteratively building meta-predictors using more and more existing predictors.

[0132] This invention provides a pathogenicity analysis system based on target feature prediction of mutations, comprising:

[0133] The acquisition module is used to acquire mutations in the gene to be predicted.

[0134] The structure generation module is used to map the obtained mutations of the gene to be predicted onto the protein structure of the gene to be predicted, so as to obtain the protein structure mapped with the mutations.

[0135] The structural feature extraction module extracts features from the protein structure mapped with mutations to obtain the structural features of the mutated protein;

[0136] The structural feature selection module is used to select structural features of mutant proteins and obtain the feature selection results.

[0137] The classification module inputs the feature selection results into the classification prediction model to obtain the classification result of whether the mutation is a pathogenic mutation or a benign mutation;

[0138] The classification prediction model obtains the classification result of mutation as pathogenic or benign mutation by one or more of the target features in the following mutant protein structural features: protein level features, residue level features, and mutation level features.

[0139] Figure 8 This invention provides a pathogenicity analysis device based on target feature prediction of mutations, comprising: a memory and a processor; the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, they are used to perform the above-mentioned pathogenicity analysis method based on target feature prediction of mutations.

[0140] The device may also include input and output devices. The memory, processor, input and output devices can be connected via a bus or other means. Figure 8 The example shown is a bus connection.

[0141] The present invention provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described pathogenicity analysis method based on target feature prediction of mutations.

[0142] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0143] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for example, the division of modules is merely a logical functional division, and there may be other division methods in actual implementation; for example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0144] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0145] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The aforementioned integrated modules can be implemented in hardware or as software functional modules.

[0146] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc.

[0147] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0148] The computer device provided by the present invention has been described in detail above. For those skilled in the art, there will be changes in the specific implementation and application scope based on the ideas of the embodiments of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for pathogenicity analysis based on predicting mutations using target features, comprising: Obtain the mutation of the gene to be predicted, wherein the gene to be predicted includes wild-type genes; The gene to be predicted is input into the protein structure generation model to obtain the protein structure of the gene to be predicted. The mutation of the gene to be predicted is mapped onto the protein structure of the gene to be predicted, and the protein structure mapped with the mutation is obtained. Feature extraction is performed on the mapped mutated protein structure to obtain the mutated protein structure features; Feature selection was performed on the structural features of the mutant protein to obtain the feature selection results; The feature selection results are input into the classification prediction model to obtain the classification results of whether the mutation is a pathogenic mutation or a benign mutation; The classification prediction model obtains the classification result of mutation as pathogenic or benign mutation by one or more of the target features among the following mutant protein structural features: protein level features, residue level features, and mutation level features. The target features in the structural characteristics of the mutant protein include RSA, ΔΔG, side chain entropy, ionization energy, and van der Waals collision.

2. The pathogenicity analysis method based on target feature prediction of mutations according to claim 1, characterized in that, The protein-level features include thermodynamic features and protein volume features; the residue-level features include secondary structure-related features and RSA; and the mutation-level features include thermodynamic features and features derived from empirical rules for determining whether a mutation has a significant effect on protein structure.

3. The pathogenicity analysis method based on target feature prediction of mutations according to claim 2, characterized in that, The thermodynamic characteristics include total unfolding energy and its van der Waals collisions, hydrogen bond energy, and side chain entropy.

4. The pathogenicity analysis method based on target feature prediction of mutations according to claim 2, characterized in that, The protein volume characteristics include total volume, interstitial volume, and van der Waals volume.

5. The pathogenicity analysis method based on target feature prediction of mutations according to claim 2, characterized in that, The secondary structure-related features are generated using the DSSP program: DSSP(H), DSSP(E), DSSP(G), DSSP(I), DSSP(T), DSSP(S), DSSP(C).

6. The pathogenicity analysis method based on target feature prediction of mutations according to claim 1, characterized in that, The target features in the structural characteristics of the mutant protein include total unfolding energy, van der Waals collision, hydrogen bond energy, side chain entropy, ionization energy, disulfide bond breakage, total volume of wild-type protein, void volume, van der Waals volume, RSA, ΔΔG, ΔG, DSSP (H), DSSP (E), DSSP (G), DSSP (I), DSSP (T), DSSP (S), and DSSP (C).

7. The pathogenicity analysis method based on target feature prediction of mutations according to claim 1, characterized in that, The step of inputting the feature selection result into the classification prediction model to obtain the classification result of the mutation as a pathogenic mutation or a benign mutation also includes: The feature selection results are input into the classification prediction model to obtain the classification results of whether the mutation is a pathogenic mutation or a benign mutation, and the first classification result is output. The mutation of the gene to be predicted is input into the predictor to obtain the mutation effect score of the gene to be predicted, and a second classification result of whether the mutation is a pathogenic mutation or a benign mutation is obtained based on the mutation effect score. The first classification result and the second classification result are weighted and fused to obtain the fused classification result of mutation as pathogenic mutation or benign mutation.

8. The pathogenicity analysis method based on target feature prediction of mutations according to claim 7, characterized in that, The predictor performs mutation effect prediction on the gene to be predicted and obtains the mutation effect score of the gene to be predicted; the predictor includes any one or more of the following software: Enformer, PROVEAN, EVE, DEOGEN2, MUTPRED, Mutation Assessor, SIFT, ClinPred, PolyPhen2, REVEL.

9. The pathogenicity analysis method based on target feature prediction of mutations according to claim 8, characterized in that, The predictors include PROVEAN, EVE, DEOGEN2, and MUTPRED.

10. The pathogenicity analysis method based on target feature prediction of mutations according to claim 7, characterized in that, The variation effect score of the gene to be predicted includes one or more of the following scores: sequence conservation score, allele mutation frequency score, missense mutation score, frameshift mutation score, nonsense mutation score, and insertion / deletion mutation score.

11. The pathogenicity analysis method based on target feature prediction of mutations according to claim 7, characterized in that, The variation effect score of the gene to be predicted includes sequence conservation score, allele mutation frequency score, and missense mutation score.

12. The pathogenicity analysis method based on target feature prediction of mutations according to claim 7, characterized in that, The weighted fusion adopts one or more of the following algorithms: Boosting, XGBoost, AdaBoost and Voting fusion algorithms.

13. The pathogenicity analysis method based on target feature prediction of mutations according to claim 1, characterized in that, The protein structure generation model includes any one or more of the following generation models: AlphaFold, AlphaFold2, and ProteoGAN.

14. The pathogenicity analysis method based on target feature prediction of mutations according to claim 1, characterized in that, The protein structure generation model is AlphaFold2.

15. A pathogenicity analysis device based on target feature prediction of mutations, characterized in that, The device includes: a memory and a processor; the memory is used to store program instructions; the processor is used to invoke the program instructions, and when the program instructions are executed, they are used to perform the pathogenicity analysis method based on target feature prediction of any one of claims 1-14.

16. A pathogenicity analysis system for predicting mutations based on target features, characterized in that, The system includes: The acquisition module is used to acquire mutations in the gene to be predicted. The structure generation module maps the acquired mutations of the gene to be predicted onto the protein structure of the gene to be predicted, thereby obtaining the protein structure mapped with the mutations. The structural feature extraction module extracts features from the protein structure mapped with mutations to obtain the structural features of the mutated protein; The structural feature selection module selects features from the structural features of the mutant protein to obtain the feature selection results; The classification module inputs the feature selection results into the classification prediction model to obtain the classification result of whether the mutation is a pathogenic mutation or a benign mutation; The classification prediction model obtains the classification result of mutation as pathogenic or benign mutation by one or more of the target features in the following mutant protein structural features: protein level features, residue level features, and mutation level features; the target features in the mutant protein structural features include RSA, ΔΔG, side chain entropy, ionization energy, and van der Waals collision.

17. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the pathogenicity analysis method based on target feature prediction of any one of claims 1-14.

Images