Missense mutation function classification method and system based on multiple omics characteristics

Abstract

The invention relates to the technical field of biological gene mutation prediction, in particular to a multi-omics feature-based missense mutation function classification method and system. The classification method comprises the following steps: S1, collecting a multi-omics comprehensive feature set related to missense mutation; s2, constructing a heterogeneous graph; s3, learning a meta-path-based weight of each node in the heterogeneous graph; s4, calculating weights of different meta-paths by using attention of a semantic level; s5, performing weighted summation on the node weight and the meta-path weight to obtain final embedding representation of each node; s6, training to obtain a multi-omics feature-based missense mutation function classification model; and S7, outputting a prediction score based on the multi-omics feature missense mutation function classification model, and dividing missense mutation according to the prediction score. According to the method, the heterogeneous graph is constructed by utilizing multiple omics characteristics, and the relationship between biological entities is explicitly modeled, so that GOF and LOF mutations are distinguished more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of biological gene mutation prediction, and in particular to a missense mutation functional classification method and system based on multi-omics features. Background Art

[0002] Missense mutations can have diverse effects on the function of gene products, including affecting protein stability, interactions with binding partners, catalytic activity and many other properties. Studying the functional consequences of missense mutations is crucial for understanding their impact on the diversity of human disease phenotypes. The functional consequences of missense mutations are mainly divided into two categories: gain-of-function (GOF), which is manifested as enhanced protein function or the emergence of new functions; and loss-of-function (LOF), which is manifested as partial or complete loss of protein function. Gain-of-function and loss-of-function mutations in the same gene can produce completely different diseases through different molecular mechanisms. Therefore, studying the types of missense mutations is of great significance to the study of gene protein biology.

[0003] Most existing methods for functional classification of missense mutations are based on multi-omics features and combined with machine learning models for prediction. For example, VPatho integrates 138 mutation-level features, 262 protein-level features, and 103 genome-level features for classification using the XGBoost model; LoGoFunc uses information such as structure, conservation, and protein interaction networks at the gene and protein levels, combined with the LightGBM model, to predict the pathogenicity of missense mutations and distinguish their functional consequences. However, most of these methods treat input features as independent variables. For example, gene tolerance scores, AF2 structural features, and PPI network features are simply flattened into vector inputs, ignoring the biological significance of different omics, resulting in an inability to effectively model explicit associations between mutations, proteins, and domains. The effect of mutations on protein function may be transmitted through multi-hop pathways (e.g., "mutation → protein → protein interaction network → protein"). Modeling the relationship between mutations and other biological entities can more comprehensively capture the functional impact pathways of mutations, thereby effectively improving the accuracy of missense mutation functional classification.

[0004] Therefore, it is necessary to design a new functional classification method for missense mutations based on multi-omics features to solve the problems of flatness and lack of association in gene mutation prediction models in existing technologies. Summary of the Invention

[0005] The present invention provides a method and system for functional classification of missense mutations based on multi-omics features, which aims to solve the problems of flattening and lack of association in gene mutation prediction models in the prior art.

[0006] To achieve the above objectives, the present invention provides a method for functional classification of missense mutations based on multi-omics features, comprising the following steps:

[0007] S1. Collect comprehensive multi-omics feature sets related to missense mutations;

[0008] S2. Construct a heterogeneous graph based on the multi-omics comprehensive feature set collected in S1. During the construction process, missense mutation nodes are initially embedded using function prediction features, and the remaining nodes are initially embedded using features of the corresponding categories.

[0009] S3, based on the heterogeneous graph constructed in S2, uses node-level attention to learn the meta-path-based weight of each node in the heterogeneous graph;

[0010] S4, based on the heterogeneous graph constructed in S3, uses semantic-level attention to calculate the weights of different meta-paths;

[0011] S5: Weighted sum of the node weights in S3 and the meta-path weights in S4 to obtain the final embedding representation of each node;

[0012] S6. The final embedding representation and the initial embedding representation of each node are concatenated and input into a multi-layer perceptron to generate a prediction score. The multi-omics feature-based missense mutation functional classification model is trained.

[0013] S7. Taking the missense mutations to be predicted with multi-omics features as input, the missense mutation functional classification model based on multi-omics features outputs a prediction score, and the missense mutations are divided according to the prediction score.

[0014] This application uses a dual weighted calculation of node weight and meta-path weight to divide missense mutations.

[0015] In the method provided in this application, the heterogeneous graph contains multiple meta-paths, and the model calculates the weight of each node in different meta-paths and the weight of each meta-path. Taking the gene node as an example, the heterogeneous graph constructed by this method has the following two meta-paths containing this node: (d) gene-transcript-missense mutation-protein and (e) tissue-gene-missense mutation-protein. The model calculates the embedding of the gene node in the meta-path (d) respectively, denoted as G d , and its embedding in the meta-path (e), denoted as G e At the same time, the model also calculates the weight of the meta-path (d), denoted as D, and the weight of the meta-path (e), denoted as E. Finally, through the weighted sum (G d ·D+G e E) Get the embedding of gene nodes.

[0016] The impact of gene mutations on protein function is complex, involving multiple levels of regulation and interactions. For example, mutations may alter protein function by altering amino acids (meta-pathway: missense mutation-amino acid-protein) or by altering protein-protein interactions (meta-pathway: missense mutation-protein-PPI-protein). Previous methods, which simply treat multi-omics features as simple feature vectors, fail to leverage this information. This method introduces meta-pathways to represent distinct biological relationships, enabling the model to capture these complex biological relationships. Furthermore, because the model dynamically learns weights, it automatically adjusts the importance of each meta-pathway and node based on the data, more accurately reflecting the role of different pathways in the propagation of mutation effects. This dynamic weight assignment not only improves the model's adaptability to diverse biological contexts but also enhances classification accuracy and interpretability, providing stronger support for functional research and clinical applications of gene mutations.

[0017] Preferably, the comprehensive feature set data sources in step S1 include: one or more of genome, protein, function prediction, transcript, protein interaction, tissue specificity, amino acid or post-translational modification related data.

[0018] Preferably, the genomic features include cell type markers from the chromHMM model and four types of conservation scores: phastCons, phyloP, GERP, and fitCons;

[0019] The protein features include secondary structure information from the AlphaFold model, predictions of solvent accessibility and protein disordered regions, as well as prediction scores based on sequence and three-dimensional structure, sequence conservation scores, MoRF and SCRIBER scores;

[0020] The functional prediction features include CADD, FATHMM, M-CAP, PROVEAN, SIFT, SiPhy 29-waylogOdds, and LIST-S2 scores, as well as frequency information from COSMIC, HGMD, and HapMap databases;

[0021] The transcriptome features include splicing scores from MaxEntScan, SpliceAI, dbscSNV, and MMSplice, the distances to transcription start and stop sites, as well as RNA motif interaction scores and transcript stability scores;

[0022] The protein interaction features include 64 features obtained by processing and analyzing protein-protein interaction data in the STRING database through a graph representation learning method;

[0023] The tissue-specific features include gene expression information in different tissues collected from the Genotype Tissue Expression (GTEx) database;

[0024] The amino acid features include substitution scores of BLOSUM62, PAM250 and EX matrices, as well as changes in physicochemical properties such as Grantham distance, charge, volume, hydrophobicity and polarity after amino acid substitution;

[0025] The post-translational modification features include phosphorylation, acetylation and methylation modification scores.

[0026] Preferably, the heterogeneous graph constructed by the multi-omics comprehensive feature set in step S2 comprises the following entity nodes: one or more of a gene node, a protein node, a missense mutation node, a transcript node, a protein interaction node, a tissue node, an amino acid node or a post-translational modification node;

[0027] The heterogeneous graph constructed using the multi-omics comprehensive feature set includes the following edge relationships: (a) gene -> missense mutation (occurrence); (b) gene -> transcript (transcription); (c) protein -> PPI (involvement); (d) missense mutation -> amino acid (resulting in substitution); (e) missense mutation -> protein (influence); (f) transcript -> missense mutation (carrier); (g) PPI -> protein (network association); (h) tissue -> gene (regulates expression); (i) amino acid -> protein (belongs to); (j) PTM -> protein (modification). The defined metapaths include: (a) missense mutation-protein-PPI-protein; (b) missense mutation-protein-PTM-protein; (c) missense mutation-amino acid-protein; (d) gene-transcript-missense mutation-protein; and (e) tissue-gene-missense mutation-protein.

[0028] Preferably, the step S3 of using node-level attention to learn the meta-path-based weight of each node in the heterogeneous graph (i.e., the importance of each node's meta-path-based neighbors) specifically includes:

[0029] Let each category of nodes in the heterogeneous graph be φ i , node φ i There are eigenvectors of different dimensions, first through the transformation matrix M i Map them to the same dimension; let the feature vector before transformation be h i , the transformed eigenvector is h′ i , and the calculation method between them is:

[0030]

[0031] After obtaining the feature vectors of the same dimension, the weight of each node based on the meta-path is learned through multi-head attention; for a node pair (i, j) under a given meta-path Φ, the node attention value between them is Indicates the importance of node j to node i, calculated as:

[0032]

[0033] where att node represents node-level attention, σ is the sigmoid activation function, || is the concatenation operation, a Φ is the node-level attention vector; the weight coefficient of each feature vector is then obtained through the softmax operation:

[0034]

[0035] Finally, the final embedding of node i under a specific meta-path Φ is calculated as:

[0036]

[0037] For a given set of meta-paths {Φ1, Φ2, ..., Φ P}, after node-level attention, we get the semantically related embeddings of P groups, expressed as

[0038] Preferably, the calculation of the weights of different meta-paths using semantic-level attention in step S4 specifically includes:

[0039]

[0040] in, Represents the meta-path Φ i The attention weight, P represents the number of meta-paths, att sem Indicates semantic level attention; att sem The specific implementation is as follows: for the meta-path Φ i , first, calculate its importance The calculation formula is:

[0041]

[0042] Among them, q is the attention vector at the semantic level, W is the weight matrix, and b is the bias vector. The above data are all obtained through learning and are shared by all meta-paths. The attention weight of the meta-path is then calculated by the softmax function. The calculation formula is:

[0043]

[0044] Preferably, the specific formula for weighted summation of the node weights in S3 and the meta-path weights in S4 in step S5 is:

[0045]

[0046] Preferably, when the missense mutation functional classification model based on multi-omics features is obtained through training in step S6, the Adam optimizer is used to update the parameters through back propagation.

[0047] Preferably, the classification of missense mutations according to the prediction scores in step S7 specifically includes:

[0048] A threshold is set for the prediction score, and the patient is classified as gain-of-function or loss-of-function type according to the threshold interval of the prediction score.

[0049] Under the same technical concept, the present invention also provides a system for functional classification of missense mutations based on multi-omics features, including the following system including the following units:

[0050] Data collection unit: used to collect comprehensive multi-omics feature sets related to missense mutations;

[0051] Graph construction unit: used to construct a heterogeneous graph based on the collected multi-omics comprehensive feature set. During the construction process, missense mutation nodes are initially embedded using function prediction features, while the remaining nodes are initially embedded using features of the corresponding categories.

[0052] Node weight calculation unit: used to learn the meta-path-based weight of each node in the constructed heterogeneous graph using node-level attention;

[0053] Meta-path weight calculation unit: used to calculate the weights of different meta-paths based on the constructed heterogeneous graph using semantic-level attention;

[0054] Weighted calculation unit: used to weight the sum of node weights and meta-path weights to obtain the final embedding representation of each node;

[0055] Model building unit: used to concatenate the final embedding representation and initial embedding representation of each node, input them into the multi-layer perceptron, generate prediction scores, and train a missense mutation functional classification model based on multi-omics features;

[0056] Classification output unit: It is used to take the missense mutation of the multi-omics feature to be predicted as input, output the prediction score based on the missense mutation functional classification model of the multi-omics feature, and classify the missense mutation according to the prediction score.

[0057] The above solution of the present invention has the following beneficial effects:

[0058] (1) The present invention uses multi-omics features to construct a heterogeneous graph, explicitly modeling the relationships between biological entities such as mutations, genes, proteins, and transcripts, so that the model can fully utilize the association information between these biological entities; this method can not only utilize the information of multi-omics features, but also dynamically learn the interaction weights between mutations and nodes such as proteins, genes, and transcripts, automatically capturing the nonlinear propagation laws of mutation effects in heterogeneous biomedical relationships, thereby more accurately distinguishing good-of-fit and low-of-fit mutations.

[0059] Other beneficial effects of the present invention will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Schematic diagram of the overall process of the missense mutation functional classification method based on multi-omics features of the present invention;

[0061] Figure 2 Schematic diagram of the test results of this embodiment relative to other models during model verification in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] To make the technical problems, technical solutions, and advantages to be solved by the present invention more clear, the following is a detailed description with reference to the accompanying drawings and specific embodiments. It is obvious that the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0063] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0064] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood broadly. For example, they may refer to a locking connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0065] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0066] like Figure 1 This is a schematic diagram of the overall process of the missense mutation functional classification method based on multi-omics features of the present invention. It should be noted that the described embodiments are intended to facilitate understanding of the present invention and do not have any limiting effect on it.

[0067] Example 1:

[0068] The present embodiment provides a method for functional classification of missense mutations based on multi-omics features, comprising the following steps:

[0069] S1. Collect a comprehensive multi-omics feature set related to missense mutations, which integrates relevant data such as genome, protein, function prediction, transcript, protein interaction, tissue specificity, amino acid and post-translational modification.

[0070] Genomic features include cell type markers from the chromHMM model and four types of conservation scores: phastCons, phyloP, GERP, and fitCons;

[0071] Protein features include secondary structure information from the AlphaFold model, predictions of solvent accessibility and disordered regions of the protein, as well as prediction scores based on sequence and three-dimensional structure, sequence conservation scores, MoRF and SCRIBER scores;

[0072] Functional prediction features include CADD, FATHMM, M-CAP, PROVEAN, SIFT, SiPhy 29-way logOdds, and LIST-S2 scores, as well as frequency information from COSMIC, HGMD, and HapMap databases;

[0073] Transcriptome features include splicing scores from MaxEntScan, SpliceAI, dbscSNV, and MMSplice, the distances to transcription start and stop sites, as well as RNA motif interaction scores and transcript stability scores;

[0074] Protein interaction features include 64 features obtained by processing and analyzing protein-protein interaction data in the STRING database through graph representation learning methods;

[0075] Tissue-specific signatures include gene expression information in different tissues collected from the genotype-tissue expression database;

[0076] Amino acid features include substitution scores from BLOSUM62, PAM250, and EX matrices, as well as changes in physicochemical properties such as Grantham distance, charge, volume, hydrophobicity, and polarity after amino acid substitution;

[0077] Post-translational modification features include phosphorylation, acetylation, and methylation modification scores.

[0078] (1) We collected missense mutation samples from the GOF / LOF database based on the HGMD and Clinvar databases and then divided them into training and test sets. The training dataset contained a total of 1,104 GOF and 5,331 LOF variants. The test set contained 136 GOF and 552 LOF variants.

[0079] (2) Use VEP and ANNOVAR tools to extract multi-omics features from databases such as dbNSFP and Clinvar.

[0080] S2. Construct a heterogeneous graph based on the multi-omics features extracted in step S1.

[0081] (1) Initialize the entity node embedding of the heterogeneous graph. The missense mutation nodes are initially embedded using the function prediction features, and the remaining nodes are initially embedded using the features of the corresponding categories.

[0082] (2) Construct edge relationships of heterogeneous graphs, including: (a) gene->missense mutation (occurrence); (b) gene->transcript (transcription); (c) protein->PPI (involvement); (d) missense mutation->amino acid (cause substitution); (e) missense mutation->protein (influence); (f) transcript->missense mutation (carry); (g)PPI->protein (network association); (h) tissue->gene (regulate expression); (i) amino acid->protein (belong to); (j) PTM->protein (modification).

[0083] (3) Construct the metapaths of the heterogeneous graph, including: (a) missense mutation-protein-PPI-protein; (b) missense mutation-protein-PTM-protein; (c) missense mutation-amino acid-protein; (d) gene-transcript-missense mutation-protein; (e) tissue-gene-missense mutation-protein.

[0084] S3. Use the heterogeneous graph of the training set for training. For missense mutation functional classification, there are two labels: gain-of-function (GOF) and loss-of-function (LOF). We use gain-of-function as the positive sample label and loss-of-function as the negative sample label. The training steps are as follows:

[0085] (1) For the heterogeneous graph constructed in step S2, node-level attention is used to learn the importance of the meta-path-based neighbors of each node in the heterogeneous graph. Considering the node φ of each category in the heterogeneous graph i All have eigenvectors of different dimensions. First, we transform them by the matrix M i Map them to the same dimension. Let the feature vector before transformation be h i , the transformed eigenvector is h′ i , and the calculation method between them is:

[0086]

[0087] After obtaining the feature vectors of the same dimension, the weight of each node based on the meta-path is learned through multi-head attention. Specifically, for a node pair (i, j) under a given meta-path Φ, the node attention value between them is Indicates the importance of node j to node i, calculated as:

[0088]

[0089] where att node represents the node-level attention. σ is the sigmoid activation function, || is the concatenation operation, and a Φ is the node-level attention vector. Then the weight coefficient of each feature vector is obtained through the softmax operation:

[0090]

[0091] Finally, the final embedding of node i under a specific meta-path Φ is calculated as:

[0092]

[0093] For a given set of meta-paths {Φ1, Φ2, ..., Φ P}, after node-level attention, we get the semantically related embeddings of P groups, expressed as

[0094] (2) Use semantic-level attention to calculate the weights of different meta-paths. The specific calculation formula is:

[0095]

[0096] in Represents the meta-path Φ i The attention weight, P represents the number of meta-paths, att sem Indicates semantic level attention. att sem The specific implementation is as follows: for the meta-path Φ i, first, calculate its importance The calculation formula is:

[0097]

[0098] Where q is the semantic level attention vector, W is the weight matrix, and b is the bias vector, all of which are obtained through learning and are shared by all meta-paths. The attention weight of the meta-path is then calculated using the softmax function, as follows:

[0099]

[0100] (3) The node weight and meta-path weight are weighted and summed to obtain the final node embedding representation. The calculation formula is:

[0101]

[0102] The final embedding representation of each node is concatenated with the initial embedding representation and then input into a multi-layer perceptron to generate a prediction score. Binary cross entropy is used as the training loss. During training, the Adam optimizer is used to update the parameters through backpropagation, with a learning rate set to 0.001 and the number of heads of the multi-head attention mechanism set to 16.

[0103] S4. After training is completed, the heterogeneous graph information of the test set is input and predictions are made using the model. Missense mutations are classified as gain-of-function or loss-of-function based on the prediction results and a fixed threshold. For example, a prediction result greater than 0.5 is classified as gain-of-function, and a prediction result less than 0.5 is classified as loss-of-function. The method of the present invention provides a new exploration direction for the functional classification of missense mutations, providing a broader development space for the exploration of the pathogenic mechanisms of complex diseases and the practical application of precision medicine.

[0104] Model Validation:

[0105] Missense mutation samples based on the HGMD and Clinvar databases were collected from the GOF / LOF databases and then divided into training and test sets. The training dataset contained a total of 1,104 GOF and 5,331 LOF variants. The test set contained 136 GOF and 552 LOF. A second test set of 1,637 GOF mutations and 2,416 LOF mutations were collected from the gnomAD and GOFCard databases. A third test set of 51 LOF mutations and 30 GOF mutations was collected from Funncion's study.

[0106] During training, good-of-fit (GOF) samples were considered positive, and low-of-fit (LOF) samples were considered negative. Binary cross-entropy was used as the training loss. The Adam optimizer was used to update parameters via backpropagation during training, with a learning rate of 0.001 and 16 heads for the multi-head attention mechanism. The dimension of the initial entity embedding vector was set to 128. To avoid overfitting, early stopping was used. 10% of the samples in the training set were randomly selected as the validation set. The loss on the validation set was calculated after each iteration, and training was stopped early if the loss on the validation set did not decrease after 20 iterations. The model and training algorithm were implemented using the PyTorch deep learning framework.

[0107] The proposed method is compared with the LoGoFunc method on three test sets. It is also compared with the Funncion method on the Funncion test set. The AUROC curves and AUPRC curves on the three test sets are shown in Figure 2. Figure 2 shown.

[0108] Observation shows that this method performs better than LoGoFunc on the HGMD and Clinvar test sets, and on the GOFCard and gnomAD test sets. On the Funncion test set, this method outperforms not only LoGoFunc but also Funncion. It is worth noting that the Funncion test set only contains missense mutations in voltage-gated sodium and calcium channels, while the Funncion method is specifically designed to study the functional consequences of point mutations in such specific genes. However, this method still performs better than Funncion, a result that fully demonstrates its greater adaptability and generalization capabilities. Through its heterogeneous graph structure and attention mechanism, this method is able to better process complex biological information and perform well even on datasets in specific domains.

[0109] The present invention also provides a system for functional classification of missense mutations based on multi-omics features, comprising the following system comprising the following units:

[0110] Data collection unit: used to collect comprehensive multi-omics feature sets related to missense mutations;

[0111] Graph construction unit: used to construct a heterogeneous graph based on the collected multi-omics comprehensive feature set. During the construction process, missense mutation nodes are initially embedded using function prediction features, while the remaining nodes are initially embedded using features of the corresponding categories.

[0112] Node weight calculation unit: used to learn the meta-path-based weight of each node in the constructed heterogeneous graph using node-level attention;

[0113] Meta-path weight calculation unit: used to calculate the weights of different meta-paths based on the constructed heterogeneous graph using semantic-level attention;

[0114] Weighted calculation unit: used to weight the sum of node weights and meta-path weights to obtain the final embedding representation of each node;

[0115] Model building unit: used to concatenate the final embedding representation and initial embedding representation of each node, input them into the multi-layer perceptron, generate prediction scores, and train a missense mutation functional classification model based on multi-omics features;

[0116] Classification output unit: It is used to take the missense mutation of the multi-omics feature to be predicted as input, output the prediction score based on the missense mutation functional classification model of the multi-omics feature, and classify the missense mutation according to the prediction score.

[0117] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for functional classification of missense mutations based on multi-omics features, characterized in that: The following steps are involved: S1. Collect comprehensive multi-omics feature sets related to missense mutations; S2. Construct a heterogeneous graph based on the multi-omics comprehensive feature set collected in S1. During the construction process, missense mutation nodes are initially embedded using function prediction features, and the remaining nodes are initially embedded using features of the corresponding categories. S3, based on the heterogeneous graph constructed in S2, uses node-level attention to learn the meta-path-based weight of each node in the heterogeneous graph; S4, based on the heterogeneous graph constructed in S3, uses semantic-level attention to calculate the weights of different meta-paths; S5: Weighted sum of the node weights in S3 and the meta-path weights in S4 to obtain the final embedding representation of each node; S6. The final embedding representation and the initial embedding representation of each node are concatenated and input into a multi-layer perceptron to generate a prediction score. The multi-omics feature-based missense mutation functional classification model is trained. S7. Taking the missense mutations to be predicted with multi-omics features as input, the missense mutation functional classification model based on multi-omics features outputs a prediction score, and the missense mutations are divided according to the prediction score.

2. The method according to claim 1, wherein The comprehensive feature set data sources in step S1 include: one or more of genome, protein, function prediction, transcript, protein interaction, tissue specificity, amino acid or post-translational modification related data.

3. The method according to claim 2, wherein The genomic features include cell type markers from the chromHMM model and four types of conservation scores: phastCons, phyloP, GERP, and fitCons; The protein features include secondary structure information from the AlphaFold model, predictions of solvent accessibility and protein disordered regions, as well as prediction scores based on sequence and three-dimensional structure, sequence conservation scores, MoRF and SCRIBER scores; The functional prediction features include CADD, FATHMM, M-CAP, PROVEAN, SIFT, SiPhy 29-way logOdds, and LIST-S2 scores, as well as frequency information from COSMIC, HGMD, and HapMap databases; The transcriptome features include splicing scores from MaxEntScan, SpliceAI, dbscSNV, and MMSplice, the distances to transcription start and stop sites, as well as RNA motif interaction scores and transcript stability scores; The protein interaction features include 64 features obtained by processing and analyzing protein-protein interaction data in the STRING database through a graph representation learning method; The tissue-specific features include gene expression information in different tissues collected from the genotype-tissue expression database; The amino acid features include substitution scores of BLOSUM62, PAM250 and EX matrices, as well as changes in physicochemical properties such as Grantham distance, charge, volume, hydrophobicity and polarity after amino acid substitution; The post-translational modification features include phosphorylation, acetylation and methylation modification scores.

4. The method according to claim 2 or 3, wherein: The heterogeneous graph constructed by the multi-omics comprehensive feature set in step S2 includes the following entity nodes: one or more of a gene node, a protein node, a missense mutation node, a transcript node, a protein interaction node, a tissue node, an amino acid node, or a post-translational modification node; The heterogeneous graph constructed by the multi-omics comprehensive feature set includes the following edge relationships: (a) gene->missense mutation (occurrence); (b) gene->transcript (transcription); (c) protein->PPI (involvement); (d) missense mutation->amino acid (cause substitution); (e) missense mutation->protein (influence); (f) transcript->missense mutation (carrying); (g)PPI->protein (network association); (h) tissue->gene (regulation expression); (i) amino acid->protein (belong to); (j) PTM->protein (modification); defined metapathways include: (a) missense mutation-protein-PPI-protein; (b) missense mutation-protein-PTM-protein; (c) missense mutation-amino acid-protein; (d) gene-transcript-missense mutation-protein; (e) tissue-gene-missense mutation-protein.

5. The method according to claim 1, wherein The use of node-level attention in step S3 to learn the meta-path-based weight of each node in the heterogeneous graph specifically includes: Let each category of nodes in the heterogeneous graph be φ i , node φ i There are eigenvectors of different dimensions, first through the transformation matrix M i Map them to the same dimension; let the feature vector before transformation be h i , the transformed eigenvector is h ' i , and the calculation method between them is: After obtaining the feature vectors of the same dimension, the weight of each node based on the meta-path is learned through multi-head attention; for a node pair (i, j) under a given meta-path Φ, the node attention value between them is Indicates the importance of node j to node i, calculated as: where att node represents node-level attention, σ is the sigmoid activation function, || is the concatenation operation, a Φ is the node-level attention vector; the weight coefficient of each feature vector is then obtained through the softmax operation: Finally, the final embedding of node i under a specific meta-path Φ is calculated as: For a given set of meta-paths {Φ1, Φ2, ..., Φ P }, after node-level attention, we get the semantically related embeddings of P groups, expressed as 6. The method according to claim 1, wherein The calculation of the weights of different meta-paths using semantic-level attention in step S4 specifically includes: in, Represents the meta-path Φ i The attention weight, P represents the number of meta-paths, att sem Indicates semantic level attention; att sem The specific implementation is as follows: for the meta-path Φ i , first, calculate its importance The calculation formula is: Among them, q is the attention vector at the semantic level, W is the weight matrix, and b is the bias vector. The above data are all obtained through learning and are shared by all meta-paths. The attention weight of the meta-path is then calculated by the softmax function. The calculation formula is:

7. The method according to claim 1, wherein The specific formula for weighted summation of the node weights in S3 and the meta-path weights in S4 in step S5 is:

8. The method according to claim 1, wherein When the missense mutation functional classification model based on multi-omics features is obtained through training in step S6, the Adam optimizer is used to update the parameters through back propagation.

9. The method according to claim 1, wherein The step S7 of classifying missense mutations according to the prediction scores specifically includes: A threshold is set for the prediction score, and the patient is classified as gain-of-function or loss-of-function type according to the threshold interval of the prediction score.

10. A multi-omics-based missense mutation functional classification system, characterized by: The system includes the following units: Data collection unit: used to collect comprehensive multi-omics feature sets related to missense mutations; Graph construction unit: used to construct a heterogeneous graph based on the collected multi-omics comprehensive feature set. During the construction process, missense mutation nodes are initially embedded using function prediction features, while the remaining nodes are initially embedded using features of the corresponding categories. Node weight calculation unit: used to learn the meta-path-based weight of each node in the constructed heterogeneous graph using node-level attention; Meta-path weight calculation unit: used to calculate the weights of different meta-paths based on the constructed heterogeneous graph using semantic-level attention; Weighted calculation unit: used to weight the sum of node weights and meta-path weights to obtain the final embedding representation of each node; Model building unit: used to concatenate the final embedding representation and initial embedding representation of each node, input them into the multi-layer perceptron, generate prediction scores, and train a missense mutation functional classification model based on multi-omics features; Classification output unit: It is used to take the missense mutation of the multi-omics feature to be predicted as input, output the prediction score based on the missense mutation functional classification model of the multi-omics feature, and classify the missense mutation according to the prediction score.

Images