Synonymous mutation harmfulness prediction method based on multi-level biological characteristics

Abstract

The invention discloses a method for predicting the harmfulness of synonymous mutation based on multilevel biological characteristics. The method comprises the following steps: 1, collecting and preprocessing a synonymous mutation sample; 2, carrying out DNA, RNA and protein level feature annotation on the obtained synonymous mutation sample by using a biological calculation tool, and carrying out pretreatment; 3, performing dimension reduction on the high-dimensional biological characteristic data by using a causal characteristic selection algorithm to obtain a plurality of causal characteristic subsets; 4, splicing an optimal feature set and performing model construction and training by using a machine learning classifier; and 5, predicting an external test set sample by using the trained optimal model to obtain a synonymous mutation harmfulness prediction score. According to the method, the features having the causal relationship with the target variable can be screened out from the multi-level high-dimensional biological features, and then the causal features of the three biological levels are fused for prediction, so that the synonymous mutation harmfulness prediction precision is improved.

Description

Technical Field

[0001] This invention belongs to the field of biotechnology, specifically a method for predicting the harmfulness of synonymous mutations based on multi-level biological characteristics. Background Technology

[0002] Point mutations refer to changes in a single nucleotide in a DNA sequence, as well as the insertion and deletion of a single nucleotide. Point mutations involving changes in a single nucleotide can be further classified as synonymous or non-synonymous mutations based on whether they alter the encoded amino acid sequence. Synonymous mutations do not change the amino acid sequence; therefore, for a long time, the vast majority of synonymous mutations were considered neutral mutations that did not alter the biological function of proteins. However, with the deepening research into synonymous mutations, many researchers have begun to focus on their functional effects. For example, Supek et al. and Kaissarian et al. demonstrated that synonymous mutations can influence the development and progression of cancer by affecting RNA splicing, expression regulation, translation rate, protein folding, and drug response. Therefore, research on the harmfulness of synonymous mutations has become increasingly important. However, distinguishing between harmful and neutral synonymous mutations through biological experiments is very time-consuming and expensive. In contrast, the development of computational methods is highly efficient and provides a foundation for further experimental verification and discovery. To meet this need, researchers have developed various excellent computational tools. These tools can be divided into two categories: broad-spectrum prediction tools applicable to a variety of mutations and specific prediction tools for synonymous mutations. While broad-spectrum tools can predict multiple mutation types, their accuracy in predicting the harmfulness of synonymous mutations is lower than that of specific tools. Currently, there are no specific tools that delve into the different impacts of DNA, RNA, and Protein-level features on synonymous mutations. This may lead to models failing to achieve optimal predictive performance due to missing information from certain feature levels. Furthermore, past studies have relied on manual feature collection for prediction, and some studies have used general feature selection methods for feature dimensionality reduction, without considering from a causal perspective which features are important for predicting the harmfulness of synonymous mutations.

[0003] Therefore, in order to solve the above-mentioned technical problems, it is urgent to propose a new technical approach. Summary of the Invention

[0004] The present invention aims to address the shortcomings of the existing technology by proposing a method for predicting the harmfulness specificity of synonymous mutations based on multi-level biological features. This method aims to screen out causal features from high-dimensional multi-level biological feature data, thereby effectively improving the predictive performance of the synonymous mutation harmfulness specificity prediction model and reducing research costs.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a method for predicting the harmfulness of synonymous mutations based on multi-level biological characteristics, characterized by the following steps: Step 1: After obtaining the synonymous mutation sample set and performing preprocessing, the preprocessed synonymous mutation sample set is obtained as follows: ,in, Indicates the first preprocessed step One synonymous mutation sample, express The tag, ;when When, it means It is a harmful mutation. When, it means It is a benign mutation; Indicates the number of synonymous mutation samples; Step 2: Use computational biology tools to analyze... Multilevel feature annotation and missing value imputation are performed to obtain DNA-level features. RNA-level characteristics Protein-level characteristics Thus obtain DNA-level feature vector matrix , RNA-level feature vector matrix , Protein-level feature vector matrix ; make , , Any biological feature vector matrix in the matrix is ​​denoted as ;make any of the first The biological feature column vector is denoted as ,in, , The number of columns representing the biological feature vector matrix; Step 3: Use the ensemble causal feature selection algorithm to select the feature from the following samples respectively. , , Feature selection is performed to obtain a subset of causal features at the DNA level. causal feature subset at the RNA level and protein-level causal feature subset ; Step 4: [Regarding...] , , After concatenation, a multi-level causal feature vector matrix is ​​obtained. Then, a five-fold cross-validation method is used to... The dataset is divided into training and validation sets, which are used to train a machine learning classifier model to obtain the optimal synonymous mutation harmfulness prediction model. This is used to obtain a harmfulness prediction score for each mutant sample.

[0006] The method for predicting the harmfulness of synonymous mutations based on multi-level biological characteristics described in this invention is also characterized in that step 3 is performed as follows: Step 3.1: Define the significance level of the conditional independence test as... ; according to Using the PCMB algorithm to Processing is performed to obtain about The set of candidate parent and child nodes as well as about Separation set ; according to Using the MBOR algorithm Processing is performed to obtain about The set of candidate parent and child nodes ; Step 3.2: Calculate the intersection of the two candidate parent node and child node sets. ; Calculate the union of two candidate parent node and child node sets as ;Calculate the difference between the sets of two candidate parent nodes and child nodes. ; Step 3.3: From Remove the incorrect parent and child nodes to obtain the corrected difference set. ; make Any number in the middle The biological feature column vector is denoted as , making it belong to But not belonging to The biological characteristic column vector is denoted as ; Using G 2 Test judgment in about Separation set In the case of a condition set and Is it conditionally dependent? If it is conditionally dependent, then it means... It is a potential parent or child node, and retains... If the conditions are independent, then use G. 2Further testing is needed to determine whether the first [condition] exists. Biological feature column vectors belong , making and Conditionally independent; if it exists, then it means It is an incorrectly selected parent or child node, and from Remove from the list; otherwise, keep. Thus, the corrected difference set is obtained. ; Step 3.4: Calculation and The union of is And serve as the final set of parent and child nodes; Step 3.5: Let The r-th biological feature column vector in the data is denoted as , making it belong to But not belonging to The The biological feature column vector is denoted as ; Using G 2 Test judgment in about Separation set In the case of a condition set and Whether the conditions are independent; if the conditions are independent, then it means It is a candidate spouse node, and will join in candidate spouse set At the same time Labeled as a column vector of biological features with multiple parent nodes; if conditionally dependent, then it represents... Nodes that are not candidate spouses are not added or marked; thus, the final set of candidate spouses is obtained. ; Step 3.6: Use the MMPC algorithm to process the biological feature column vectors marked with multiple parent nodes. Processing is performed to obtain The set of parent and child nodes, and with Merge, and obtain The set of spouses ; Step 3.7: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and The union as about Markov blanket .

[0007] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in executing the synonym mutation harmfulness prediction method, and the processor is configured to execute the program stored in the memory.

[0008] The present invention provides a computer-readable storage medium on which a computer program is stored, characterized in that the computer program, when executed by a processor, performs the steps of the synonym mutation harmfulness prediction method.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The method of this invention differs from traditional synonymous mutation harmfulness prediction methods. It collects as complete as possible synonymous mutation features at the DNA, RNA, and protein levels and uses a causal feature selection algorithm to screen out useful causal features for prediction, thus solving the problem of limited feature collection in previous synonymous mutation harmfulness prediction methods.

[0010] 2. In the process of constructing the synonymous mutation harmfulness prediction model, this invention uses a causal feature selection algorithm to reduce the dimensionality of features. This algorithm can find features with causal relationships with the harmfulness of synonymous mutations in multi-level high-dimensional biological features, making the trained prediction model superior to traditional prediction models, thereby improving the prediction performance of the synonymous mutation harmfulness prediction model. Attached Figure Description

[0011] Figure 1 This is a flowchart of the synonymous mutation harmfulness prediction method of the present invention. Detailed Implementation

[0012] In this embodiment, a method for predicting the harmfulness of synonymous mutations based on multi-level biological features is used to screen causal features from high-dimensional multi-level biological features and apply them to the task of predicting the harmfulness of synonymous mutations. Specifically, such as... Figure 1 As shown, proceed as follows: Step 1: After obtaining the synonymous mutation sample set and performing preprocessing, the preprocessed synonymous mutation sample set is obtained as follows: ,in, Indicates the first preprocessed step One synonymous mutation sample, express The tag, .when When, it means It is a harmful mutation. When, it means It is a benign mutation; This indicates the number of synonymous mutation samples. The preprocessing process includes the following steps: 1) Removal of redundant, conflicting, and common mutation samples; 2) Homologous sequence similarity control; 3) Close-by processing.

[0013] Step 2: Use computational biology tools to analyze... Multilevel feature annotation and missing value imputation are performed to obtain DNA-level features. RNA-level characteristics Protein-level characteristics Thus obtain DNA-level feature vector matrix , RNA-level feature vector matrix , Protein-level feature vector matrix ;make , , Any biological feature vector matrix in the matrix is ​​denoted as ;make any of the first The biological feature column vector is denoted as ,in, , The number of columns in the biological feature vector matrix represents the total number of features. DNA-level features were extracted using nine tools: CADD, ANNOVAR, ChromHMM, Favor, siPhy, iFeatureOmgea, frtCOOL, DNAshapeR, and MathFeature. RNA-level features were extracted using eight tools: RNAfold, sincFold, RNAPKplex, SpliceAI, MMSplice, iFeatureOmgea, frtCOOL, and MathFeature. Protein-level features were extracted using three tools: iFeatureOmega, AIUPred, and NetSurfP-3.0. Missing value imputation was performed using MissForest, which fully utilizes the internal structure and relationships of the data to predict missing values, continuously improving imputation accuracy through multiple iterations, and can quickly fill missing values ​​in high-dimensional feature data.

[0014] Step 3: Use the ensemble causal feature selection algorithm to select the feature from the following samples respectively. , , Feature selection is performed to obtain a subset of causal features at the DNA level. causal feature subset at the RNA level and protein-level causal feature subset ; Step 3.1: Define the significance level of the conditional independence test as... In this embodiment Set to 0.01.

[0015] according to Using the PCMB algorithm to Processing is performed to obtain about The set of candidate parent and child nodes as well as about Separation set ; according to Using the MBOR algorithm Processing is performed to obtain about The set of candidate parent and child nodes The two algorithms learn the sets of parent and child nodes with completely opposite strictness, thus obtaining feature sets with significant differences for subsequent secondary discrimination.

[0016] Step 3.2: Calculate the intersection of the two candidate parent node and child node sets. ; Calculate the union of two candidate parent node and child node sets as ;Calculate the difference between the sets of two candidate parent nodes and child nodes. The features in this set are considered controversial features. Secondary discrimination helps to identify features that have been incorrectly selected into the parent and child node sets, which is helpful for subsequent prediction tasks.

[0017] Step 3.3: From Remove the incorrect parent and child nodes to obtain the corrected difference set. ; make Any number in the middle The biological feature column vector is denoted as , making it belong to But not belonging to The biological characteristic column vector is denoted as ; Using G 2 Test judgment in about Separation set In the case of a condition set and Is it conditionally dependent? If it is conditionally dependent, then it means... It is a potential parent or child node, and retains... If the conditions are independent, then use G.2 Further testing is needed to determine whether the first [condition] exists. Biological feature column vectors belong , making and Conditionally independent; if it exists, then it means It is an incorrectly selected parent or child node, and from Remove from the list; otherwise, keep. Thus, the corrected difference set is obtained. .

[0018] Step 3.4: Calculation and The union of is And serve as the final set of parent and child nodes; Step 3.5: Let The r-th biological feature column vector in the data is denoted as , making it belong to But not belonging to The The biological feature column vector is denoted as ; Using G 2 Test judgment in about Separation set In the case of a condition set and Whether the conditions are independent; if the conditions are independent, then it means It is a candidate spouse node, and will join in candidate spouse set At the same time Labeled as a column vector of biological features with multiple parent nodes; if conditionally dependent, then it represents... Nodes that are not candidate spouses are not added or marked; thus, the final set of candidate spouses is obtained. .

[0019] Step 3.6: Use the MMPC algorithm to process the biological feature column vectors marked with multiple parent nodes. Processing is performed to obtain The set of parent and child nodes, and with Merge, and obtain The set of spouses ; Step 3.7: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and The union as about Markov blanket ;make The features selected by the integrated causal feature selection algorithm represent the features selected by the algorithm. The RNA features selected by the integrated causal feature selection algorithm are represented. This represents the protein features selected by the integrated causal feature selection algorithm. Features selected by the integrated causal feature selection algorithm are considered to have a causal relationship with the target variable, which can enhance the interpretability of subsequent prediction models.

[0020] Step 4: [Regarding...] , , After concatenation, a multi-level causal feature vector matrix is ​​obtained. Then, a five-fold cross-validation method is used to... The dataset is divided into training and validation sets, which are used to train a machine learning classifier model. LightGBM is selected as the machine learning model to obtain the optimal synonymous mutation harmfulness prediction model. This is used to obtain a harmfulness prediction score for each mutant sample.

[0021] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0022] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0023] The specific implementation method of the present invention and the effectiveness of the method of the present invention are illustrated by examples of collected synonym mutation datasets.

[0024] The dataset for this experiment consists of two parts: a training set and a test set. The positive samples in the training set are primarily from the Human Gene Mutation Database (HGMD) Professional Edition (April 2022 version), the ClinVar database (March 31, 2024), and two manually curated databases for harmful synonymous mutations: dbDSM and SynMall. The negative samples are obtained from the ClinVar database (March 31, 2024). The obtained mutation data underwent redundancy, conflict, and common sample removal, homology sequence similarity control, and close-by processing. The final training dataset consists of 1181 positive samples (pathogenic synonymous mutations) and 1181 negative samples (benign synonymous mutations), with no overlap between the positive and negative samples. This embodiment uses the above dataset to construct a prediction model according to the implementation process and names it CFDPSM.

[0025] To evaluate the model's performance, this experiment collected positive sample information from the same database. After controlling for homology sequence similarity, 119 positive samples with less than 40% sequence similarity to the training set were obtained, and then close-by processing was performed. The test set contained 119 positive samples and 119 negative samples. The composition of all datasets used in this experiment is shown in Table 1.

[0026] Table 1. Dataset Composition

[0027] This embodiment uses AUC and AUPR to measure the overall performance of the model. AUC and AUPR are not affected by thresholds and range from 0 to 1. The closer they are to 1, the better the overall performance of the model. Therefore, they are often considered important evaluation metrics.

[0028] To verify the superiority of the model, this embodiment selects comparison with the following synonym mutation broad-spectrum and specific pathogenicity prediction tools, including: CADD (Schubach, M., et al., CADD v1. 7: using protein language models, regulatory CNNs and other nucleotide-level scores to improve genome-wide variant predictions. Nucleic acids research, 2024. 52(D1): p.D1143-D1154.), DANN (Quang, D., Y. Chen, and X. Xie, DANN: a deep learning approach for annotating the pathogenicity of genetic variants. Bioinformatics, 2014. 31(5): p. 761-763.), FATHMM-MKL (Shihab, HA, et al., Anintegrative approach to predicting the functional effects of non-coding and coding sequence variation. Bioinformatics, 2015. 31(10): p. 1536-1543.), FATHMM-XF (Rogers, MF, et al., FATHMM-XF: accurate prediction of pathogenicpoint mutations via extended features. Bioinformatics, 2018. 34(3): p. 511-513.), PhD-SNPg (Capriotti, E. and P. Fariselli, PhD-SNPg: updating awebserver and lightweight tool for scoring nucleotide variants. Nucleic AcidsResearch, 2023. 51(W1): p. W451-W458.), TraP (Gelfman, S., et al., Annotating pathogenic non-coding variants in genic regions. Nature communications, 2017.8(1): p. 236.), PrDSM (Cheng, N., et al., Comparison and integration of computational methods for deleterious synonymous mutation prediction. Briefings in bioinformatics, 2020. 21(3): p. 970-981.) and frDSM (Wang, H., et al., frDSM: an ensemble predictor with effective feature representation for deleterious synonymous mutation in human genome. IEEE / ACM Transactions on Computational Biology Bioinformatics, 2022. 20(1): p. 371-377.). Among them, PrDSM and frDSM are specific tools, while the remaining six tools are broad-spectrum tools. Some methods in Table 3 failed to obtain corresponding prediction scores on the test set. Therefore, performance comparisons were performed using the largest common subset of samples from the test set with prediction scores. The largest common subset contained 209 synonymous mutation samples, of which 105 were positive and 104 were negative.

[0029] Table 2 Comparison of CFDPSM based on missing test sets and other methods

[0030] Table 3 Comparison of CFDPSM and other methods based on missing test sets

[0031] Tables 2 and 3 show that the CFDPSM method of this invention outperforms other methods in both AUC and AUPR values. These results demonstrate that collecting multi-level biological features and screening for causal features, and applying them to the task of predicting the harmfulness of synonymous mutations, is effective.

Claims

1. A method for predicting the deleteriousness of synonymous mutations based on multiple levels of biological features, characterized by, is performed by the following steps: Step 1: After obtaining a synonymous mutation sample set and preprocessing, a preprocessed synonymous mutation sample set is obtained , wherein represents the preprocessed i-th synonymous mutation sample, represents the label of ; when , it represents is a harmful mutation, when , it represents is a benign mutation; represents the number of synonymous mutation samples;​ Step 2: Multi-level feature annotation and missing value filling using computational biology tools to obtain DNA-level features , RNA-level features , and protein-level features , thereby obtaining a DNA-level feature vector matrix , an RNA-level feature vector matrix , and a protein-level feature vector matrix . ; Let , , be any biological feature vector matrix in ; let be any th biological feature column vector in , where , represents the number of columns of the biological feature vector matrix. Step 3: Feature selection is performed on the DNA level causal feature subset , , the RNA level causal feature subset , and the protein level causal feature subset using the integrated causal feature selection algorithm, respectively. ; Step 4: Concatenate the output of step 3 and the input sequence to get the multi-level causal feature vector matrix , , After concatenation, the multi-level causal feature vector matrix is obtained , and the five-fold cross-validation method is used to divide into a training set and a validation set, and used to train the machine learning classifier model, thereby obtaining the optimal synonymous mutation harmfulness prediction model , which is used to obtain the harmfulness prediction score of each mutation sample.

2. The method of predicting the deleteriousness of synonymous mutations based on multi-level biological features according to claim 1, wherein, Step 3 is performed by the following steps: Step 3.1: Define condition-independent test significance level as ; According to , the PCMB algorithm is used to process , and the candidate parent node and child node set about and the separation set about are obtained. According to , the MBOR algorithm is used to process , to obtain the candidate parent node and child node set about ; Step 3.2: Calculate the intersection of the two candidate parent and child node sets ; Calculate the union of the two candidate parent and child node sets as ; Calculate the difference of the two candidate parent and child node sets as ; Step 3.3: Remove erroneous parent and child nodes from ;​ Let anyth biological feature column vector be denoted by , let biological feature column vector belong to but not belong to be denoted by With G 2 Test determines in As for Separate set of As a condition set, And Whether conditional dependence; If condition dependent, then indicates is a potential parent or child node and is retained ; if condition independent, then utilizes G 2 to test further to determine if there is a second biological feature column vector that belongs to the set and is conditionally independent; if so, then indicates is an erroneous parent or child node and is removed from ; otherwise, is retained ; thereby resulting in a revised difference set ; Step 3.4: Calculation and The union of is And serve as the final set of parent and child nodes; Step 3.5: Let be the rth biological feature column vector in , let be the rth biological feature column vector in , and let be the rth biological feature column vector in ; G 2 inspection determines whether is a candidate spouse node and adds to the candidate spouse set of as a condition set, and marks as a biological feature column vector with multiple parent nodes. is a candidate spouse node and adds to the candidate spouse set of as a condition set, and marks as a biological feature column vector with multiple parent nodes. If condition dependent, then represent Not a candidate partner node, do not join and mark; thus resulting in final candidate partner set ; Step 3.6: Using MMPC algorithm to process the biological feature column vector marked as having multiple parent nodes to obtain the parent and child node set of , and merging with to obtain the spouse set of ;​ Step 3.7: The union of and is taken as A Markov blanket with respect to .​ 3. An electronic device comprising a memory and a processor, characterized in that The memory is configured to store a program supporting the processor to perform the synonymous mutation harmfulness prediction method of claim 1 or 2.

4. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program, when executed by the processor, performs the steps of the synonymous mutation harmfulness prediction method of claim 1 or 2.

Images