Synonymous mutation harmfulness prediction method based on gene language model
By combining gene language models and deep learning frameworks, a synonymous mutation harmfulness prediction network is constructed, which solves the problem of insufficient generalization ability of existing models and achieves high-precision synonymous mutation harmfulness prediction, which is suitable for genomics research and clinical mutation screening.
Patent Information
- Application Number
- CN202511651863.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-02-10
AI Technical Summary
Existing synonym mutation harmfulness prediction models are insufficient in generalization ability and accuracy. Traditional methods rely on manually designed feature extraction, which leads to the loss of feature correlation and the neglect of global data features, making it difficult to efficiently and accurately identify the harmfulness of synonym mutations.
We use gene language models (GPN-MSA, SpliceBERT, HyenaDNA, Caduceus) to extract multiple embedding features, and construct a synonym mutation harmfulness prediction network through dynamic convolution plus bidirectional LSTM, TextCNN, fully connected layers and feature fusion modules. We then use deep learning optimization algorithms to improve prediction accuracy.
By incorporating features from multiple models, the accuracy of synonymous mutation harmfulness prediction and the generalization ability of the model are improved, providing a novel approach to synonymous mutation harmfulness prediction that is applicable to genomics research and clinical mutation screening.
Smart Images
Figure CN121506246A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of bioinformatics and artificial intelligence, specifically relating to a method for predicting the harmfulness of synonymous mutations based on a gene language model (Gene-LM). This method combines feature embeddings from multiple gene language models with deep learning structures to achieve intelligent prediction of the pathogenicity probability of synonymous mutations, falling within the scope of gene variation function prediction and disease gene research techniques. Background Technology
[0002] Synonymous mutations are a type of gene mutation that refers to changes in one or more nucleotides in the DNA sequence, but these changes do not lead to changes in the amino acid sequence of the encoded protein. This phenomenon occurs because many amino acids in the genetic code are encoded by multiple different codons, a phenomenon known as "degeneracy of the genetic code." For example, the codons "GGA" and "GGG" can both encode glycine. Therefore, if "GGA" in the original DNA sequence mutates to "GGG," although the DNA sequence changes, the encoded protein sequence remains the same; such a mutation is a synonymous mutation. Currently, the main methods for distinguishing synonymous mutations include wet experiments and computational methods. However, identifying synonymous mutations using wet experiments is time-consuming and laborious, and studying synonymous mutations across the entire genome is even more difficult. With the rapid development of advanced computing technologies such as artificial intelligence and high-throughput sequencing technologies, data-based artificial intelligence methods can predict harmful synonymous mutations more efficiently, quickly, and at a lower cost than traditional biological experimental methods. Currently, there are many computational tools available for identifying harmful synonymous mutations, which can be categorized into specific tools and broad-spectrum tools based on their applicability. Broad-spectrum tools can assess the function of point mutations and also identify the harmfulness of synonymous mutations, such as CADD v1.7 (Schubach, Max, et al. "CADD..."). v1. 7: using protein language models, regulatoryCNNs and other nucleotide-level scores to improve genome-wide variantpredictions." Nucleic acids research 52.D1 (2024): D1143-D1154.), DANN (Quang, Daniel, Yifei Chen, and Xiaohui Xie. "DANN: a deep learning approach forannotating the pathogenicity of genetic variants." Bioinformatics 31.5(2014): 761-763.), FATHMM-MKL (Shihab, Hashem A., et al. "An integrative approach to predicting the functional effects of non-coding and coding sequence variation." Bioinformatics 31.10 (2015): 1536-1543.), etc., specific tools such as EnDSM (Cheng, Na, et al. "Effect predictor of driver synonymous mutations based on multi-feature fusion and iterative feature representation learning." IEEE Journal of Biomedical and Health Informatics 28.2 (2023): 1144-1151.), PrDSM (Cheng, Na, et al. "Comparison and integration of computational methods for deleterious synonymous mutation prediction." Briefings in bioinformatics 21.3 (2020): 970-981.), usDSM (Tang, Xi, et al. "usDSM: a novel method for deleterious synonymous mutation prediction using undersampling scheme." Briefings in Bioinformatics 22.5 (2021): bbab123.) etc. To facilitate use, researchers have also developed user-friendly online prediction tools that provide rapid synonym mutation prediction services. However, the generalization ability of current models needs improvement. This challenge mainly stems from the limitations of traditional machine learning methods, which rely on manually designed feature extraction and model building. This approach not only increases the complexity and cost of data preprocessing but may also lead to the loss of correlations between features and the neglect of global features of the data, thus limiting the model's generalization ability. Summary of the Invention
[0003] This invention aims to address the shortcomings of existing technologies by proposing a gene language model-based method for predicting the harmfulness of synonymous mutations. This method leverages the powerful text understanding and reasoning capabilities of large language models to extract relevant information about gene harmfulness from gene sequences. Simultaneously, it utilizes a deep learning feature extraction and fusion framework to improve the accuracy of reasoning, thereby overcoming the problem of insufficient accuracy in predicting the harmfulness of synonymous mutations in existing methods.
[0004] 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 a gene language model, characterized by the following steps: Step 1: Obtain the dataset of synonymous mutation samples },in, Represents the i-th synonymous mutation sample. The number of synonymous mutation samples, let The category label is ,and ;when When, it means This is a benign sample; when When, it means This is a harmful sample; Step 2: Use the gene language model GPN-MSA to analyze the samples. Feature extraction is performed to obtain samples Sequence information embedding features ,in, express The sequence length input to the GPN-MSA model, The dimension representing the embedded features of sequence information; Using the gene language model SpliceBERT to Feature extraction is performed to obtain splice embedding features ,in, express The length of the sequence input to the SpliceBERT model. Indicates the dimension of the splice embedding feature; Using the gene language model HyenaDNA to Feature extraction is performed to obtain samples Global information embedding features ,in, Indicates sample The sequence length input into the HyenaDNA model, Indicates sample The dimension of the global information embedding features; Using the gene language model Caduceus to analyze samples Feature extraction is performed to obtain Inverse complementary embedding features ,in, express The length of the sequence input into the Caduceus model. The dimension representing the inverse complementary embedding feature; Step 3: Construct a synonym mutation harmfulness prediction network, including: a dynamic convolutional layer with bidirectional LSTM module, a TextCNN module, a fully connected layer, a feature fusion module, and a prediction module, and then process the results. 、 、 Processing is performed to obtain Predicted category score ; Step 4: Construct the loss function of the synonymous mutation harmfulness classification network using equation (14). : (1) In equation (1), for Mapping function; Step 5: Optimize the synonym mutation harmfulness classification network using the Adam optimization algorithm, and calculate... The network parameters are updated until convergence is achieved, thus obtaining a synonym mutation harmfulness classification model with optimal parameters, which is used to predict the category of harmfulness of the input synonym mutation.
[0005] The method for predicting the harmfulness of synonymous mutations based on gene language models described in this invention is characterized in that step 3 is performed as follows: Step 3.1: Dynamic convolution plus bidirectional LSTM module, including: dynamic convolutional layer, bidirectional LSTM layer. Two parts, and respectively for and Processing is performed, and the corresponding results are obtained. Deep sequence features and Deep splicing features ; Step 3.2, TextCNN module Processing yields the first... i Deep information embedding features ; Step 3.3, the fully connected layer uses formula (2) for... Perform linear transformation and nonlinear activation processing to obtain the first... i Deeply inverse complementary features : (2) In equation (2), The linear weight matrix in the fully connected layer. For bias terms in fully connected layers, It is a non-linear activation function; Step 3.4, the feature fusion module will 、 、 and After feature splicing, the first... i Deep fusion vectors : Step 3.5: The prediction module uses equations (11), (12), and (13) to... Processing is performed to obtain Predicted category score : (3) (4) (5) In equations (3), (4), and (5), , , This is the weight matrix of each layer in the three fully connected layers. , , These are the bias terms for each layer in the three fully connected layers. , These are two activation features of the middle layer of a three-layer fully connected layer. It is a non-linear activation function.
[0006] Furthermore, step 3.1 is performed as follows: Step 3.1.1: The dynamic convolutional layer contains m convolutional branches and their corresponding sets of convolutional kernels. and to Processing yields the first... i A dimensionality-reduced focusing feature ,in, Indicates the first j The convolution kernel of each convolutional branch; Step 3.1.2, Use bidirectional right Processing yields the first... i A global context feature and will The Middle t The characteristics of the time step as the first i Deep sequence features ,in, ; Step 3.1.3: Follow the procedures in steps 3.1.1 and 3.1.2. Perform the same process to obtain the first... i A deep splicing feature .
[0007] Furthermore, step 3.1.1 is performed as follows: Step a, the first Each convolutional branch uses equation (6) to... Processing yields the first... The first convolutional branch i Weights : (6) In equation (6), Represents the activation function. represents the linear transformation matrix to be learned, and m represents the number of channels in the dynamic convolutional layer; Step b, the Each convolutional branch uses equation (7) to... Processing yields the first... The first convolutional branch i Convolutional features : (7) In equation (7), Representing the Convolution kernels of each convolution branch One-dimensional convolution operation; Step c, use formula (8) to and Processing yields the first... i Convolutional weighted fusion features : (8) Step e, using formula (9) and Processing yields the first... i Individual residual characteristics : (9) In equation (9), This is a convolution operation with a 1×1 kernel; Indicates residual connection; Step f, using formula (10) Processing yields the first... i Multi-perspective features : ) (10) In equation (10), It is a non-linear activation function; Step g, using formula (11) Perform feature reduction to obtain the first... i A dimensionality-reduced focusing feature : (11) In equation (11), For max pooling operation, It is a one-dimensional convolution operation.
[0008] Furthermore, step 3.2 is performed as follows: Step 3.2.1, use formula (12) to... Processing yields the first... j The first scale i Local information features : (12) In equation (12), Mish is the activation function; Indicates transpose; Step 3.2.2, use formula (13) to... Perform branch pooling to obtain the first... j The first scale i Local information convolutional features : (13) In equation (13), This is a one-dimensional max pooling operation; Step 3.2.3: Use equation (14) for all scales. i The convolutional features of the local information are concatenated and linearly transformed to obtain the first... i A deep global information feature : (14) In equation (14), For feature splicing, This is the weight matrix of the fully connected layer. For the bias term of the fully connected layer, This is a batch normalization operation.
[0009] 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.
[0010] The present invention provides a computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, performs the steps of the synonym mutation harmfulness prediction method.
[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Existing research on the harmfulness prediction of synonymous mutations mainly relies on manually extracted biological features, and in the field of gene language models for predicting mutation harmfulness, it primarily focuses on the harmfulness prediction of non-synonymous mutations or broad-spectrum mutations. This invention is the first to apply gene language models to the task of predicting the harmfulness of synonymous mutations, and by introducing the feature expression capabilities of gene language models, it fills a gap in this field and provides a completely new approach to the harmfulness prediction of synonymous mutations.
[0012] 2. This invention proposes a feature splicing strategy for multi-model embedding fusion. By integrating the embedding features of multiple gene language models (GPN-MSA, HyenaDNA, SpliceBERT, Caduceus), it captures multi-dimensional biological information such as sequence evolution conservation, long-range dependence, splicing correlation, and reverse complementarity, thereby achieving complementarity between information from different models. Thus, through feature splicing and deep learning fusion, the feature expression ability and prediction performance of the model are effectively improved.
[0013] 3. The GeneLM-DSM model constructed in this invention integrates the embedding features of multiple gene language models and adopts nonlinear dimensionality reduction and Sigmoid classification mechanisms, thereby achieving high-precision prediction through multi-source feature fusion, which improves the accuracy and reliability of downstream classification tasks.
[0014] 4. The method of this invention can not only be applied to the prediction of the harmfulness of synonymous mutations, but also extended to other types of genetic variation effect prediction tasks. It has strong generalization ability and scalability, and provides a universal technical framework for genomics research and clinical mutation screening. Attached Figure Description
[0015] Figure 1 This is a flowchart illustrating the method of the present invention. Detailed Implementation
[0016] In this embodiment, a method for predicting the harmfulness of synonymous mutations based on gene language models uses synonymous mutation sequences as a dataset. Four gene language models (GPN-MSA, HyenaDNA, SpliceBERT, and Caduceus) are employed to extract corresponding embedding features. Three feature extraction modules are used to extract features from the four embedding features, obtaining deep-level extracted features to acquire deeper semantic information related to mutation harmfulness. By concatenating the four deep-level extracted features, performing dimensionality reduction, and calculating prediction scores, the harmfulness of synonymous mutations can be predicted. The framework diagram is shown below. Figure 1 As shown, the specific steps are as follows: Step 1: Obtain the dataset of synonymous mutation samples },in, Represents the i-th synonymous mutation sample. The number of synonymous mutation samples, let The category label is ,and ;when When, it means This is a benign sample; when When, it means This is a harmful sample; Step 2: Use the gene language model GPN-MSA to analyze the samples. Feature extraction is performed to obtain samples Sequence information embedding features ,in, express The sequence length input to the GPN-MSA model, The dimension representing the embedded features of sequence information; Using the gene language model SpliceBERT to Feature extraction is performed to obtain splice embedding features ,in, express The length of the sequence input to the SpliceBERT model. Indicates the dimension of the splice embedding feature; Using the gene language model HyenaDNA to Feature extraction is performed to obtain samples Global information embedding features ,in, Indicates sample The sequence length input into the HyenaDNA model, Indicates sample The dimension of the global information embedding features; Using the gene language model Caduceus to analyze samples Feature extraction is performed to obtain Inverse complementary embedding features ,in, express The length of the sequence input into the Caduceus model. The dimension representing the inverse complementary embedding feature; The acquisition and function of embedding features are as follows: First, synonymous mutation sequence data meeting the requirements of the gene language model must be provided. After submitting the sequence data, the gene language model learns and understands the input data information based on its own parameters. Finally, by obtaining the information from the last hidden layer of the model, the corresponding embedding features can be obtained. The embedding features contain the gene language model's understanding of the input sequence data.
[0017] Step 3: Construct a synonym mutation harmfulness prediction network, including: a dynamic convolutional layer with bidirectional LSTM module, a TextCNN module, a fully connected layer, a feature fusion module, and a prediction module, and then process the results. 、 、 Processing is performed to obtain Predicted category score ; Step 4: Construct the loss function of the synonymous mutation harmfulness classification network using equation (1). : (1) In equation (1), for Mapping function; Step 5: Train the network using five-fold cross-validation, and optimize the synonym mutation harmfulness classification network using the Adam optimization algorithm, and calculate... The network parameters are updated until convergence is achieved, thus obtaining a synonym mutation harmfulness classification model with optimal parameters, which is used to predict the category of harmfulness of the input synonym mutation.
[0018] 2. The method for predicting the harmfulness of synonymous mutations based on a gene language model according to claim 1, characterized in that step 3 is performed according to the following steps: Step 3.1: Dynamic convolution plus bidirectional LSTM module, including: dynamic convolutional layer, bidirectional LSTM layer. Two parts, and respectively for and Processing is performed, and the corresponding results are obtained. Deep sequence features and Deep splicing features ; Step 3.2, TextCNN module Processing yields the first... i Deep information embedding features ; Step 3.3, the fully connected layer uses formula (2) for... Perform linear transformation and nonlinear activation processing to obtain the first... i Deeply inverse complementary features : (2) In equation (2), The linear weight matrix in the fully connected layer. For bias terms in fully connected layers, It is a non-linear activation function; Step 3.4, the feature fusion module will 、 、 and After feature splicing, the first... i Deep fusion vectors : Step 3.5: The prediction module uses equations (3), (4), and (5) to... Processing is performed to obtain Predicted category score : (3) (4) (5) In equations (3), (4), and (5), , , This is the weight matrix of each layer in the three fully connected layers. , , These are the bias terms for each layer in the three fully connected layers. , These are two activation features of the middle layer of a three-layer fully connected layer. It is a non-linear activation function.
[0019] 3. The method for predicting the harmfulness of synonymous mutations based on a gene language model according to claim 2, characterized in that step 3.1 is performed according to the following steps: Step 3.1.1: The dynamic convolutional layer contains m convolutional branches and their corresponding sets of convolutional kernels. and to Processing yields the first... i A dimensionality-reduced focusing feature ,in, Indicates the first j The convolution kernel of each convolutional branch; Step 3.1.2, Use bidirectional right Processing yields the first... i A global context feature and will The Middle t The characteristics of the time step as the first i Deep sequence features ,in, ; Step 3.1.3: Follow the procedures in steps 3.1.1 and 3.1.2. Perform the same process to obtain the first... i A deep splicing feature .
[0020] 4. The method for predicting the harmfulness of synonymous mutations based on a gene language model according to claim 3, characterized in that step 3.1.1 is performed according to the following steps: Step a, the first Each convolutional branch uses equation (6) to... Processing yields the first... The first convolutional branch i Weights : (6) In equation (6), Represents the activation function. represents the linear transformation matrix to be learned, and m represents the number of channels in the dynamic convolutional layer; Step b, the Each convolutional branch uses equation (7) to... Processing yields the first... The first convolutional branch i Convolutional features : (7) In equation (7), Representing the Convolution kernels of each convolution branch One-dimensional convolution operation; Step c, use formula (8) to and Processing yields the first... i Convolutional weighted fusion features : (8) Step e, using formula (9) and Processing yields the first... i Individual residual characteristics : (9) In equation (4), This is a convolution operation with a 1×1 kernel; Indicates residual connection; Step f, using formula (10) Processing yields the first... i Multi-perspective features : ) (10) In equation (10), It is a non-linear activation function; Step g, using formula (11) Perform feature reduction to obtain the first... i A dimensionality-reduced focusing feature : (11) In equation (11), For max pooling operation, It is a one-dimensional convolution operation.
[0021] 5. The method for predicting the harmfulness of synonymous mutations based on a gene language model according to claim 2, characterized in that step 3.2 is performed according to the following steps: Step 3.2.1, use formula (12) to... Processing yields the first... j The first scale i Local information features : (12) In equation (12), Mish is the activation function; Indicates transpose; Step 3.2.2, use formula (13) to... Perform branch pooling to obtain the first... j The first scalei Local information convolutional features : (13) In equation (13), This is a one-dimensional max pooling operation; Step 3.2.3: Use equation (14) for all scales. i The convolutional features of the local information are concatenated and linearly transformed to obtain the first... i A deep global information feature : (14) In equation (14), For feature splicing, This is the weight matrix of the fully connected layer. For the bias term of the fully connected layer, This is a batch normalization operation.
[0022] 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.
[0023] 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.
[0024] The dataset for this experiment consists of two parts: a training set and a test set. The training set contains 2362 samples, and the test set contains 238 samples. The dataset used in this experiment is shown in Table 1.
[0025] Table 1. Dataset Composition
[0026] The validation criteria used include recall, specificity (SPE), precision (PR), accuracy (ACC), Matthews correlation coefficient (MCC), and F1 score (F1), which are calculated as follows: In this model, TP (True positive) represents the number of true positives, i.e., the number of true pathogenic synonymous mutations correctly predicted as pathogenic synonymous mutations; TN (True negative) represents the number of true negatives, i.e., the number of true benign synonymous mutations correctly predicted as benign synonymous mutations; FP (False positive) represents the number of false positives, i.e., the number of mutations that were originally pathogenic synonymous mutations but were predicted as benign synonymous mutations; and FN (False negative) represents the number of false negatives, i.e., the number of mutations that were originally benign synonymous mutations but were predicted as pathogenic synonymous mutations. In addition, AUC and AUPR were used in this experiment to measure the overall performance of the model. Generally, the six metrics given in the above formulas are affected by thresholds; samples greater than or equal to the threshold are predicted as positive, while samples less than the threshold are considered as negative. The default threshold is 0.5, but it can be manually adjusted. AUC and AUPR, however, are not affected by thresholds and range from 0 to 1. The closer to 1, the better the overall model performance, and therefore they are often considered more important evaluation metrics.
[0027] To verify the superiority of the model, several excellent tools were selected for comparison, including five general methods: CADD v1.7 (Schubach, Max, 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 52.D1 (2024): D1143-D1154.), DANN (Quang, Daniel, Yifei Chen, and Xiaohui Xie. "DANN: a deep learning approach for annotating the pathogenicity of genetic variants." Bioinformatics 31.5 (2014): 761-763.), and FATHMM-MKL (Shihab, Hashem A., et al. "An integrative approach to predicting the functional effects of non-coding and coding sequence variation." Bioinformatics 31.10 (2015). 1536-1543.), PhD-SNPg (Capriotti, Emidio, and Piero Fariselli. "PhD-SNPg: updating a webserver and lightweight tool for scoring nucleotide variants." Nucleic Acids Research51.W1 (2023): W451-W458.), FATHMM-XF (Rogers, Mark F., et al. "FATHMM-XF:accurate prediction of pathogenic point mutations via extended features."Bioinformatics 34.3 (2018): 511-513.); and 6 specific methods: TraP (Gelfman, Sahar, et al."Annotating pathogenic non-coding variants in genic regions." Naturecommunications 8.1 (2017): 236.), EnDSM (Cheng, Na, et al. "Effect predictor ofdriver synonymous mutations based on multi-feature fusion and iterativefeature representation learning." IEEE Journal of Biomedical and HealthInformatics 28.2 (2023): 1144-1151.), frDSM (Wang, Huadong, et al. "frDSM: anensemble predictor with effective feature representation for deleterioussynonymous mutation in human genome." IEEE / ACM Transactions on ComputationalBiology and Bioinformatics 20.1 (2022): 371-377.), PrDSM (Cheng, Na, et al. "Comparison and integration of computational methods for deleterioussynonymous mutation prediction." Briefings in bioinformatics 21.3 (2020):970-981.), usDSM (Tang, Xi, et al. "usDSM: a novel method for deleterioussynonymous mutation prediction using undersampling scheme." Briefings inBioinformatics 22.5 (2021): bbab123.) and seDSM (Wang, Lihua, et al."Deleterious synonymous mutation identification based on selective ensemble strategy." Briefings in Bioinformatics 24.1 (2023). Broad-spectrum methods represent approaches that focus not only on the harmfulness prediction of synonymous mutations, while specific methods study only the harmfulness prediction of synonymous mutations. Tables 2 and 3 present a complete performance comparison of GeneLM-DSM with other tools across eight evaluation metrics (RECALL, SPE, PRE, F1, MCC, ACC, AUC, and AUPR). Specifically, the following results are presented to facilitate comparison and analysis of differences between the various tools. Due to missing values in the training set for some methods, comparisons were made with all methods on the largest common subset of the test set, as shown in Table 2. Table 2 Comparison of GeneLM-DSM based on the largest common subset of the test set with other methods
[0028] In Table 2, GeneLM-DSM achieved an AUC of 0.938, a 0.02 improvement over usDSM (AUC = 0.918), which was the best-performing specific tool previously. It was also compared with methods that did not have missing values on the test set, as shown in Table 3. Table 3 Comparison of GeneLM-DSM based on test set and methods without missing values
[0029] In Table 3, GeneLM-DSM (AUC=0.928) has an AUC 0.008 higher than EnDSM (AUC=0.920), the best-performing prediction tool. The results and analysis above show that GeneLM-DSM has strong robustness and generalization ability.
Claims
1. A method for predicting the harmfulness of synonymous mutations based on a gene language model, characterized in that, The procedure is as follows: Step 1: Obtain the dataset of synonymous mutation samples },in, Represents the i-th synonymous mutation sample. The number of synonymous mutation samples, let The category label is ,and ;when When, it means This is a benign sample; when When, it means This is a harmful sample; Step 2: Use the gene language model GPN-MSA to analyze the samples. Feature extraction is performed to obtain samples Sequence information embedding features ,in, express The sequence length input to the GPN-MSA model. The dimension representing the embedded features of sequence information; Using the gene language model SpliceBERT to Feature extraction is performed to obtain splice embedding features ,in, express The length of the sequence input to the SpliceBERT model. Indicates the dimension of the splice embedding feature; Using the gene language model HyenaDNA to Feature extraction is performed to obtain samples Global information embedding features ,in, Indicates sample The sequence length input into the HyenaDNA model, Indicates sample The dimension of the global information embedding features; Using the gene language model Caduceus to analyze samples Feature extraction is performed to obtain Inverse complementary embedding features ,in, express The length of the sequence input into the Caduceus model. The dimension representing the inverse complementary embedding feature; Step 3: Construct a synonym mutation harmfulness prediction network, including: a dynamic convolutional layer with bidirectional LSTM module, a TextCNN module, a fully connected layer, a feature fusion module, and a prediction module, and then process the results. 、 、 Processing is performed to obtain Predicted category score ; Step 4: Construct the loss function of the synonymous mutation harmfulness classification network using equation (14). : (1) In equation (1), for Mapping function; Step 5: Optimize the synonym mutation harmfulness classification network using the Adam optimization algorithm, and calculate... The network parameters are updated until convergence, thus obtaining a synonym mutation harmfulness classification model with optimal parameters, which is used to predict the category of harmfulness of the input synonym mutation.
2. The method for predicting the harmfulness of synonymous mutations based on a gene language model according to claim 1, characterized in that, Step 3 is performed as follows: Step 3.1: Dynamic convolution plus bidirectional LSTM module, including: dynamic convolutional layer, bidirectional LSTM layer. Two parts, and respectively for and Processing is performed, and the corresponding results are obtained. Deep sequence features and Deep splicing features ; Step 3.2, TextCNN module... Processing yields the first... i Deep information embedding features ; Step 3.3, the fully connected layer uses formula (2) for... Perform linear transformation and nonlinear activation processing to obtain the first... i Deeply inverse complementary features : (2) In equation (2), The linear weight matrix in the fully connected layer. For bias terms in fully connected layers, It is a non-linear activation function; Step 3.4, the feature fusion module will 、 、 and After feature splicing, the first... i Deep fusion vectors : Step 3.5: The prediction module uses equations (11), (12), and (13) to... Processing is performed to obtain Predicted category score : (3) (4) (5) In equations (3), (4), and (5), , , This is the weight matrix of each layer in the three fully connected layers. , , These are the bias terms for each layer in the three fully connected layers. , These are two activation features of the middle layer of a three-layer fully connected layer. It is a non-linear activation function.
3. The method for predicting the harmfulness of synonymous mutations based on a gene language model according to claim 2, characterized in that, Step 3.1 is performed as follows: Step 3.1.1: The dynamic convolutional layer contains m convolutional branches and their corresponding sets of convolutional kernels. and to Processing yields the first... i A dimensionality-reduced focusing feature ,in, Indicates the first j The convolution kernel of each convolutional branch; Step 3.1.2, Use bidirectional right Processing yields the first... i A global context feature and will Middle t The characteristics of the time step as the first i Deep sequence features ,in, ; Step 3.1.3: Follow the procedures in steps 3.1.1 and 3.1.
2. Perform the same process to obtain the first... i A deep splicing feature .
4. The method for predicting the harmfulness of synonymous mutations based on a gene language model according to claim 3, characterized in that, Step 3.1.1 is performed as follows: Step a, the first Each convolutional branch uses equation (6) to... Processing yields the first... The first convolutional branch i Weights : (6) In equation (6), Represents the activation function. represents the linear transformation matrix to be learned, and m represents the number of channels in the dynamic convolutional layer; Step b, the Each convolutional branch uses equation (7) to... Processing yields the first... The first convolutional branch i Convolutional features : (7) In equation (7), Representing the Convolution kernels of each convolution branch One-dimensional convolution operation; Step c, use formula (8) to and Processing yields the first... i Convolutional weighted fusion features : (8) Step e, using formula (9) and Processing yields the first... i Individual residual characteristics : (9) In equation (9), This is a convolution operation with a 1×1 kernel; Indicates residual connection; Step f, using formula (10) Processing yields the first... i Multi-perspective features : ) (10) In equation (10), It is a non-linear activation function; Step g, using formula (11) Perform feature reduction to obtain the first... i A dimensionality-reduced focusing feature : (11) In equation (11), For max pooling operation, It is a one-dimensional convolution operation.
5. The method for predicting the harmfulness of synonymous mutations based on a gene language model according to claim 2, characterized in that, Step 3.2 is performed as follows: Step 3.2.1, use formula (12) to... Processing yields the first... j The first scale i Local information features : (12) In equation (12), Mish is the activation function; Indicates transpose; Step 3.2.2, use formula (13) to... Perform branch pooling to obtain the first... j The first scale i Local information convolutional features : ) (13) In equation (13), This is a one-dimensional max pooling operation; Step 3.2.3: Use equation (14) for all scales. i The convolutional features of the local information are concatenated and linearly transformed to obtain the first... i A deep global information feature : (14) In equation (14), For feature splicing, This is the weight matrix of the fully connected layer. For the bias term of the fully connected layer, This is a batch normalization operation.
6. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the synonymous mutation harmfulness prediction method according to any one of claims 1-5, the processor being configured to execute the program stored in the memory.
7. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, performs the steps of the synonym mutation harmfulness prediction method according to any one of claims 1-5.