A Deep Learning-Based Method for Predicting Detrimental Variants in Measured G-Quadruplex Regions in Cell Lines

Abstract

This invention discloses a deep learning-based method for predicting harmful variants located in the experimentally measured G-quadruplex region of cell lines, comprising: 1. collecting and preprocessing a dedicated dataset of harmful mutations in the G-quadruplex (G4) and its flanking sequence regions of cell lines; 2. fine-tuning the Nucleotide Transformer model using parameter-efficient fine-tuning techniques to extract deep semantic features of the sequences containing mutation sites; 3. extracting and screening epigenetic features of specific cell lines using the Enformer model; 4. constructing and training a dual-channel deep learning model containing both sequence semantic feature processing and epigenetic feature processing channels to predict harmful variants located in the G-quadruplex region of cell lines. This invention improves the accuracy of harmful variant prediction by integrating the deep semantic understanding capabilities of a large language model with cell-specific epigenetic information, thus providing a new tool for a deeper understanding of the regulatory mechanisms of G4 and the development of personalized medicine.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and deep learning technology, specifically a method for predicting harmful variants located in the experimental G-quadruplex region of cell lines based on deep learning. Background Technology

[0002] Deoxyribonucleic acid (DNA) can form not only the classic double helix structure, but also non-classical secondary structures in guanine-rich sequences, the most representative of which is the G-quadruplex. Whole-genome sequencing and high-throughput mapping studies have shown that G4-forming sequences are significantly enriched in key regulatory regions of the genome, such as gene promoters, 5'-untranslated regions, and open chromatin regions. These structures, as important cis-regulatory elements, participate in various key cellular processes, including DNA replication, transcriptional regulation, and maintaining genome stability. Research indicates that single nucleotide variants located in the G4-forming sequences and their flanking regions may disrupt normal gene regulatory networks and become potential drivers of human diseases. Therefore, accurately predicting the harmfulness of mutations in the G4 and its flanking regions is crucial for elucidating the molecular mechanisms of disease development.

[0003] Existing research on predicting the harmfulness of genomic mutations largely focuses on whole-genome or exon regions, exhibiting significant limitations in its applicability when dealing with the unique non-coding structural region of G4. First, most harmfulness prediction tools widely used in clinical genetics are designed for protein-coding regions, with feature engineering emphasizing changes in amino acid physicochemical properties or sequence conservation across species. However, G4 structures are primarily distributed in non-coding regions such as promoters and UTRs, and their function highly depends on the folding stability of specific secondary structures and their tight coupling with various epigenetic modifications. Broad-spectrum models often overlook the unique structural motif features of the G4 region and its specific context as a regulatory element, resulting in limited accuracy in predicting the harmfulness of mutations in this region. Second, G4 function is highly cell-specific and depends on chromatin state and transcription factor binding, while obtaining cell-specific epigenetic features is costly and subject to dimensionality noise. Traditional prediction methods, to integrate this information, typically rely on wet experimental data such as ChIP-seq or ATAC-seq from specific cell lines, which significantly limits their application. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing a deep learning-based method for predicting harmful variants in the experimentally measured G-quadruplex region of cell lines. This method aims to integrate deep sequence semantic features with key epigenetic information, thereby improving the accuracy of predicting the harmfulness of G4 region mutations. This, in turn, provides a new tool for a deeper understanding of the regulatory mechanism of G4 and the development of personalized medicine.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a method for predicting harmful variants located in the experimentally measured G-quadruplex region of cell lines based on deep learning, characterized by the following steps: Step 1: Construct the target variant sequence dataset ;in, This represents the i-th target variant sequence containing the G-quadruplex and its flanking sequences; ∈{0,1} means Corresponding harmful mutation tags; This represents the length of the target variant sequence; N is the total number of target variant sequences. Step 2, based on Using IA 3 The strategy fine-tunes the parameters of the bond projection matrix layer, value projection matrix layer, and feedforward network layer in the nucleotide converter model, resulting in a fine-tuned nucleotide converter model. Used for Process the data to obtain the i-th hidden state. Thus, the semantic feature vector of the i-th high-dimensional sequence is obtained using equation (1). Where d is the dimension of the hidden layer; (1) In equation (1), Attention mask representing a nucleotide converter model The Middle t Bit mask express The Middle t Hidden state; Step 3: Construct auxiliary training dataset D aux This is used to train the auxiliary dual-channel model, thereby obtaining the trained model iDualG4; Step 3.1: Obtain any cell line h The actual in vivo G-quadruplex sequence was used as the positive sample set. P vivo Obtained only in vitro folding but in the same cell line h Sequences that did not form G-quadruplexes in the lower body were used as negative sample sets. N vitro Thus by P vivo and N vitro composition D aux ;make D auxThe real label set is ; Step 3.2: Utilize the pre-trained epigenetic prediction model M epi For D aux Perform predictions and output a full-dimensional epigenetic feature map. And based on cell lines h The name, from Cell lines were screened h The following epigenetic information E epi ,in, C represents the number of windows in the spatial sequence dimension of the epigenetic phenotypic map. total This represents the total number of channels in the epigenetic trait map; Step 3.3: Construct an auxiliary dual-channel model, including: sequence branches for processing sequence information. Apparent branch used for processing full-dimensional appearance information and fully connected layers, and for D aux and epigenetic information E epi Processing yields D aux The predicted probability set of G-quadruplex in vivo ; Step 3.4: Construct the classification loss L of the auxiliary dual-channel model using equation (3). aux ; (3) In equation (3), D represents aux The total number of samples in the sample, express any sample The true label, express middle The predicted probability value; Step 3.5, Minimize L aux The training aided dual-channel model was used to obtain the trained model iDualG4, which can distinguish between in vivo and in vitro G-quadruplexes. Step 4, from positive sample set P vivo A subset of epigenetic information is randomly sampled from the corresponding epigenetic feature map; the epigenetic information is then used in the trained model iDualG4 with the epigenetic branch. A subset of epigenetic information is processed to obtain a subset of epigenetic channel features. The DeepSHAP algorithm is then used to calculate the contribution score of each epigenetic channel feature subset under each channel, and the index set I corresponding to the epigenetic channel features under the top K channels in terms of contribution score is selected. key ; Step 5: Utilize the pre-trained epigenetic prediction model M epi right X i Prediction is performed to obtain the i-th epigenetic feature map. and from Extract only I from key The corresponding epigenetic characteristics constitute the i-th low-dimensional epigenetic feature vector. ; Step 6: Construct the main dual-channel model, including: sequence semantic feature processing branch Ψ seq Key epigenetic features processing branch Ψ epi and classification layers, and for and Processing is performed to obtain Predicted probability of harmful mutations and with Y i Constructing a binary cross-entropy loss function This allows for the training of the primary dual-channel model, resulting in the trained harmful mutation prediction model M. main It is used to predict the G-quadruplex sequence of the input mutation and obtain the optimal harmful mutation prediction category.

[0006] The method described in this invention is also characterized in that step 1 includes: Step 1.1: Obtain the set of mutation sites V = {v1, v2, ..., v} m ,...,v M}, where v m This represents the m-th mutation site and includes genomic coordinates and a harmful mutation label; Obtain any cell line h The set of coordinate intervals of the G-quadruplex in vivo measured below. vivo ; Step 1.2: Combine the set of variant sites V with G vivo Spatial intersection is performed, and based on the human reference genome, target variant sequences corresponding to sites falling within the G-quadruplex and its flanking sequence regions are generated, thereby constructing a target variant sequence dataset. .

[0007] Furthermore, in step 2, the use of IA 3The strategy is to use Equation (2) to fine-tune the bond projection matrix K, value projection matrix V, and down projection matrix D in the self-attention layer of the nucleotide converter model, thereby obtaining the fine-tuned bond projection matrix. The fine-tuned projection matrix and the fine-tuned lower projection matrix : (2) In equation (2), ⊙ denotes element-wise multiplication. , , There are three scaling vectors to be learned.

[0008] Furthermore, the negative sample set N in step 3.1 vitro The construction method is as follows: Dataset A based on G4-seq high-throughput sequencing technology was obtained from the gene expression comprehensive database; From dataset A, select the set of sequence coordinates G that can form G-quadruplex structures under in vitro experimental conditions. seq ; G seq With cell lines h The set of coordinate intervals of the G-quadruplex in vivo measured below. vivo Perform a difference operation and remove the overlapping regions of the two sets to obtain the set that folds only in vitro but in cell lines. h The sequence in the lower body did not form a G-quadruplex.

[0009] Furthermore, step 3.3 includes: Step 3.3.1 The auxiliary training dataset after one-hot encoding D’ aux Processing is performed to obtain sequence channel features. F DNA ; Step 3.3.2 For E epi Processing was performed to obtain epigenetic channel characteristics. F epi ; Step 3.3.3, for F DNA and F epi After feature fusion, the input is used for prediction in a fully connected layer, and the output is D. aux The predicted probability set of G-quadruplex in vivo .

[0010] Furthermore, step 6 includes: Step 6.1, Sequence semantic feature processing channel Ψ seq Using fully connected residual networks Perform a nonlinear mapping to obtain the i-th semantic deep representation. ; Step 6.2, Processing of key epigenetic features (branch Ψ) epi Using a one-dimensional convolutional neural network and a global pooling structure in sequence to... Perform local pattern extraction to obtain the i-th deep representation of the appearance. ; Step 6.3, on the channel dimension and After splicing, the data is input into the classification layer for processing to obtain... Predicted probability of harmful mutations ; Step 6.4: Construct the binary cross-entropy loss function using equation (3). : (3) Step 6.5: Minimize the binary cross-entropy loss function The parameters of the main dual-channel model are trained until... The training of the harmful mutation prediction model M is obtained by converging until convergence is achieved. main .

[0011] Furthermore, the sequence semantic feature processing channel Ψ in step 6.1 seq It is obtained using equations (4) and (5) : (4) (5) In equations (4) and (5), Indicates random inactivation. , , , These represent the output dimensions as follows: , , , The fully connected layer, BN represents batch normalization, and Proj represents dimension-aligned projection layer. This indicates an element-wise addition operation. This represents the feature extraction operation of the feedforward network. Furthermore, the key epigenetic feature processing channel Ψ in step 6.2 epi It is obtained using equations (6)-(8) : (6) (7) (8) In equations (6)-(8), , These represent the number of filters respectively. , One-dimensional convolutional layer, This represents the feature map output by the initial convolutional operation of the first layer. and Let represent the feature maps output by the m-th and (m-1)-th residual convolutional blocks, respectively. This represents the feature map output by the last residual convolutional block in a fully connected residual network. This indicates max pooling. Indicates global average pooling. Indicates a modified linear element. Indicates the output dimension is The fully connected layer.

[0012] 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 performing the method described therein, and the processor is configured to execute the program stored in the memory.

[0013] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to perform the steps of the method described thereon.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention does not require additional cell-specific sequencing data. It uses the Enformer model to directly predict the epigenetic signals of specific cell lines from DNA sequences, overcoming the dependence of traditional methods on expensive wet experimental data such as ChIP-seq or ATAC-seq. This enables the model to predict cell lines or regions lacking experimental data, significantly reducing application costs.

[0015] 2. This invention does not directly use massive amounts of epigenetic prediction signals. Instead, it uses the DeepSHAP interpretability algorithm to accurately select the Top 10 core features most relevant to G4 harmfulness from multiple features, thereby achieving feature denoising and model optimization. This step effectively eliminates irrelevant background noise, prevents model overfitting, and makes the model more focused on key biological processes.

[0016] 3. This invention combines the structural characteristics of G4 with the cell-specific epigenetic core features introduced by the Enformer model, enabling the model to perceive the chromatin environment and transcriptional regulatory background of the mutation site, thereby improving the biological rationality of the prediction. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the dual-channel deep network model of the present invention. Figure 2a This is a distribution map of the top 10 core epigenetic characteristics of the MCF-7 cell line selected based on DeepSHAP in this invention; Figure 2b This is a distribution map of the top 10 core epigenetic characteristics of the HepG2 cell line screened using DeepSHAP according to the present invention; Figure 2c This is a distribution map of the top 10 core epigenetic characteristics of the K562 cell line screened using DeepSHAP according to the present invention; Figure 3 This is a comparison chart of the performance of the dual-channel deep learning model of this invention. Detailed Implementation

[0018] In this embodiment, a method for predicting harmful variants located in the experimentally measured G-quadruplex region of a cell line based on deep learning is described, such as... Figure 1 As shown, the procedure is as follows: Step 1: Construct the target variant sequence dataset ;in, This represents the i-th target variant sequence containing the G-quadruplex and its flanking sequences; ∈{0,1} means Corresponding harmful mutation tags; This represents the length of the target variant sequence; N is the total number of target variant sequences.

[0019] Step 1.1: Obtain the set of mutation sites V = {v1, v2, ..., v} m ,...,v M}, where v m This represents the m-th mutation site and includes genomic coordinates and a harmful mutation label; Obtain any cell line h The set of coordinate intervals of the G-quadruplex in vivo measured below. vivo .

[0020] Step 1.2: Combine the set of variant sites V with G vivo Spatial intersection is performed, and based on the human reference genome, target variant sequences corresponding to sites falling within the G-quadruplex and its flanking sequence regions are generated, thereby constructing a target variant sequence dataset. .

[0021] In the specific implementation, step 1.1 involves downloading the latest VCF file from the ClinVar database, parsing the INFO field, and filtering out single nucleotide variants (SNPs) explicitly labeled as "Pathogenic" or "Benign". Simultaneously, for specific cell lines (this example uses three representative cell lines: K562, MCF-7, and HepG2), in vivo experimental data are obtained from the Gene Expression Comprehensive Database (GEO) to obtain G... vivo .

[0022] In step 1.2, spatial intersection is performed to ensure that the studied variant sites do indeed occur in key regions within the body capable of folding into G-quadruplexes. To construct the target variant sequence dataset, generated based on the human reference genome, this embodiment sets the sequence length L. target =2048 bp. For label Y i Set "Harmful" to 1 and "Harmless" to 0.

[0023] Step 2, based on Using IA 3 The strategy fine-tunes the parameters of the bond projection matrix layer, value projection matrix layer, and feedforward network layer in the nucleotide converter model, resulting in a fine-tuned nucleotide converter model. Used for Process the data to obtain the i-th hidden state. Thus, the semantic feature vector of the i-th high-dimensional sequence is obtained using equation (1). Where d is the dimension of the hidden layer; (1) In equation (1), Attention mask representing a nucleotide converter model The Middle t Bit mask express The Middle t The bit is in a hidden state.

[0024] Step 2 uses IA 3 The strategy is to use Equation (4) to fine-tune the bond projection matrix K, value projection matrix V, and down projection matrix D in the self-attention layer of the nucleotide converter model, thereby obtaining the fine-tuned bond projection matrix. The fine-tuned projection matrix and the fine-tuned lower projection matrix : (2) In equation (2), ⊙ denotes element-wise multiplication. , , There are three scaling vectors to be learned.

[0025] Step 3: Construct auxiliary training dataset D aux This is used to train the auxiliary dual-channel model, thereby obtaining the trained model iDualG4; Step 3.1: Obtain any cell line h The actual in vivo G-quadruplex sequence was used as the positive sample set. P vivo Obtained only in vitro folding but in the same cell line h Sequences that did not form G-quadruplexes in the lower body were used as negative sample sets. N vitro Thus by P vivo and N vitro composition D aux ;make D aux The real label set is .

[0026] Step 3.2: Utilize the pre-trained epigenetic prediction model M epi For D aux Perform predictions and output a full-dimensional epigenetic feature map. And based on cell lines h The name, from Cell lines were screened h The following epigenetic information E epi ,in, C represents the number of windows in the spatial sequence dimension of the epigenetic phenotypic map. total This represents the total number of channels in the epigenetic trait map.

[0027] Step 3.3: Construct an auxiliary dual-channel model, including: sequence branches for processing sequence information. Apparent branch used for processing full-dimensional appearance information And fully connected layers.

[0028] Step 3.3.1 The auxiliary training dataset after one-hot encoding D’ aux Processing is performed to obtain sequence channel features. F DNA ; Step 3.3.2 For Eepi Processing was performed to obtain epigenetic channel characteristics. F epi ; Step 3.3.3, for F DNA and F epi After feature fusion, the input is used for prediction in a fully connected layer, and the output is D. aux The predicted probability set of G-quadruplex in vivo .

[0029] Step 3.4: Construct the classification loss L of the auxiliary dual-channel model using equation (3). aux ; (3) In equation (3), D represents aux The total number of samples in the sample, express any sample The true label, express middle The predicted probability value; Step 3.5, Minimize L aux The training of the auxiliary dual-channel model resulted in the iDualG4 model, which can distinguish between in vivo and in vitro G-quadruplexes.

[0030] Step 3.6, from positive sample set P vivo A subset of epigenetic information is randomly sampled from the corresponding epigenetic feature map; the epigenetic information is then used in the trained model iDualG4 with the epigenetic branch. A subset of epigenetic information is processed to obtain a subset of epigenetic channel features. The DeepSHAP algorithm is then used to calculate the contribution score of each epigenetic channel feature subset under each channel, and the index set I corresponding to the epigenetic channel features under the top K channels in terms of contribution score is selected. key .

[0031] Step 3 aims to utilize an auxiliary model to identify key regulatory factors that can distinguish between the in vivo specific living cell environment and the in vitro environment, thereby filtering noise from massive epigenetic signals and identifying the truly effective biological characteristics involved in cell regulation. The pre-trained epigenetic prediction model Enformer used in this invention outputs C... total=5313 full-dimensional epigenetic feature maps. Directly using thousands of dimensions of epigenetic features would introduce a lot of noise. This invention constructs an auxiliary model iDualG4 and uses the DeepSHAP algorithm to find 10 key epigenetic features, such as Figure 2a Epigenetic characteristics distribution map corresponding to MCF-7 cell line, set I key The corresponding biological characteristics include histone modifications (H3K4me3, H3K9ac, H3K4me2, H2AFZ), DNase hypersensitive sites (DNASE), RNA polymerase II subunit (POLR2A), and transcriptional regulatory factors (SIN3A, MYC). Figure 2b Epigenetic characteristics distribution map corresponding to the HepG2 cell line, set I key The corresponding biological characteristics include histone modifications (H3K4me3, H3K4me2, H3K9ac), RNA-binding proteins (HNRNPLL, RBFOX2, HNRNPK, PCBP1), RNA polymerase II subunits (POLR2G, POLR2AphosphoS5), and DNA repair proteins (XRCC5). Figure 2c The corresponding epigenetic feature distribution map of the K562 cell line shows that the biological features corresponding to the Ikey set include transcription factors (PHF8, ZBTB7A, MAX, EGR1, GABPA, SP1), RNA-binding protein (RBFOX2), DNase hypersensitive site (DNASE), and histone modification (H3K4me3).

[0032] Step 4: Utilize the pre-trained epigenetic prediction model M epi right X i Prediction is performed to obtain the i-th epigenetic feature map. and from Extract only I from key The corresponding epigenetic characteristics constitute the i-th low-dimensional epigenetic feature vector. .

[0033] Step 5: Construct the main dual-channel model, including: sequence semantic feature processing branch Ψ seq Key epigenetic features processing branch Ψ epi and classification layers, and for and Processing is performed to obtain Predicted probability of harmful mutations and with Y i Constructing a binary cross-entropy loss function This allows for the training of the primary dual-channel model, resulting in the trained harmful mutation prediction model M. main .

[0034] Step 5.1, Sequence semantic feature processing channel Ψ seq Using fully connected residual networks Perform a nonlinear mapping to obtain the i-th semantic deep representation. ; Step 5.1 Sequence semantic feature processing channel Ψ seq It is obtained using equations (4) and (5) : (4) (5) In equations (4) and (5), Indicates random inactivation. , , , These represent the output dimensions as follows: , , , The fully connected layer, BN represents batch normalization, and Proj represents dimension-aligned projection layer. This indicates an element-wise addition operation. This represents the feature extraction module of the feedforward network. In Ψ seq In the channel, the parameters are configured as follows: =512, =256, =128, =64. Dropout rate set to 0.3. After the residual module Add operation, it is finally compressed into a 64-dimensional vector. .

[0035] Step 5.2, Processing of key epigenetic features (branch Ψ) epi Using a one-dimensional convolutional neural network and a global pooling structure in sequence to... Local pattern extraction is performed, and the i-th deep representation of the appearance is obtained using equations (6)-(8). ; (6) (7) (8) In equations (6)-(8), , These represent the number of filters respectively. , One-dimensional convolutional layer, This represents the feature map output by the initial convolutional operation of the first layer. and Let represent the feature maps output by the m-th and (m-1)-th residual convolutional blocks, respectively. This represents the feature map output by the last residual convolutional block in a fully connected residual network. This indicates max pooling. Indicates global average pooling. Indicates a modified linear element. Indicates the output dimension is A fully connected layer. Ψ epi The channel processing dimension is 16×10 for the apparent feature map, and the convolution kernel size is set to 3×3. The number of convolution kernels in equation (7) is... =64; Equation (8) represents stacked residual convolutional layers (e.g., when m=2,3, Set to 64 and 128), finally pass through the global average pooling layer of equation (9), and combine with the fully connected layer to output a 64-dimensional ( The vector with =64) .

[0036] Step 5.3, on the channel dimension and After concatenation, the data is input into the classification layer for processing to obtain... Predicted probability of harmful mutations ; Step 5.4: Construct the binary cross-entropy loss function using equation (9). : (9) Step 5.5: Minimize the binary cross-entropy loss function The parameters of the main dual-channel model are trained until... The training of the harmful mutation prediction model M is obtained by converging until convergence is achieved. main It is used to predict the G-quadruplex sequence of the input mutation and obtain the optimal harmful mutation prediction category.

[0037] In terms of channel dimension and The concatenated 128-dimensional fusion vector is successively passed through fully connected layers with 128 and 64 neurons, and finally through an output layer with a sigmoid activation function, outputting the predicted probability of harmful mutations (between 0 and 1). The network training employs the Adam optimizer, combined with an EarlyStopping mechanism to automatically terminate training and prevent overfitting.

[0038] Figure 3 This invention demonstrates the harmful mutation prediction model M proposed in this paper.main Histograms comparing the performance of the proposed model with that of the control model in three cell lines: HepG2, MCF-7, and K562. Evaluation metrics included precision, recall, F1 score, accuracy, area under the ROC curve (AUC-ROC), and area under the PR curve (AUPR). The results demonstrate that among all tested cell lines, the proposed model achieved the best predictive performance on the aforementioned key evaluation metrics.

[0039] 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.

[0040] 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.

Claims

1. A method for predicting harmful variants located in the experimentally measured G-quadruplex region of cell lines based on deep learning, characterized in that, The procedure is as follows: Step 1: Construct the target variant sequence dataset ;in, This represents the i-th target variant sequence containing the G-quadruplex and its flanking sequences; ∈{0,1} means Corresponding harmful mutation tags; This represents the length of the target variant sequence; N is the total number of target variant sequences. Step 2, based on Using IA 3 The strategy fine-tunes the parameters of the bond projection matrix layer, value projection matrix layer, and feedforward network layer in the nucleotide converter model, resulting in a fine-tuned nucleotide converter model. Used for Process the data to obtain the i-th hidden state. Thus, the semantic feature vector of the i-th high-dimensional sequence is obtained using equation (1). Where d is the dimension of the hidden layer; (1) In equation (1), Attention mask representing a nucleotide converter model The Middle t Bit mask express The Middle t Hidden state; Step 3: Construct auxiliary training dataset D aux This is used to train the auxiliary dual-channel model, thereby obtaining the trained model iDualG4; Step 3.1: Obtain any cell line h The actual in vivo G-quadruplex sequence was used as the positive sample set. P vivo Obtained only in vitro folding but in the same cell line h Sequences that did not form G-quadruplexes in the lower body were used as negative sample sets. N vitro Thus by P vivo and N vitro composition D aux ;make D aux The real label set is ; Step 3.2: Utilize the pre-trained epigenetic prediction model M epi For D aux Perform predictions and output a full-dimensional epigenetic feature map. And based on cell lines h The name, from Cell lines were screened h The following epigenetic information E epi ,in, C represents the number of windows in the spatial sequence dimension of the epigenetic phenotypic map. total This represents the total number of channels in the epigenetic trait map; Step 3.3: Construct an auxiliary dual-channel model, including: sequence branches for processing sequence information. Apparent branch used for processing full-dimensional appearance information and fully connected layers, and for D aux and epigenetic information E epi Processing yields D aux The predicted probability set of G-quadruplex in vivo ; Step 3.4: Construct the classification loss L of the auxiliary dual-channel model using equation (3). aux ; (3) In equation (3), D represents aux The total number of samples in the sample, express any sample The true label, express middle The predicted probability value; Step 3.5, Minimize L aux The training aided dual-channel model was used to obtain the trained model iDualG4, which can distinguish between in vivo and in vitro G-quadruplexes. Step 4, from positive sample set P vivo A subset of epigenetic information is randomly sampled from the corresponding epigenetic feature map; the epigenetic information is then used in the trained model iDualG4 with the epigenetic branch. A subset of epigenetic information is processed to obtain a subset of epigenetic channel features. The DeepSHAP algorithm is then used to calculate the contribution score of each epigenetic channel feature subset under each channel, and the index set I corresponding to the epigenetic channel features under the top K channels in terms of contribution score is selected. key ; Step 5: Utilize the pre-trained epigenetic prediction model M epi right X i Prediction is performed to obtain the i-th epigenetic feature map. and from Extract only I from key The corresponding epigenetic characteristics constitute the i-th low-dimensional epigenetic feature vector. ; Step 6: Construct the main dual-channel model, including: sequence semantic feature processing branch Ψ seq Key epigenetic features processing branch Ψ epi and classification layers, and for and Processing is performed to obtain Predicted probability of harmful mutations and with Y i Constructing a binary cross-entropy loss function This allows for the training of the primary dual-channel model, resulting in the trained harmful mutation prediction model M. main It is used to predict the G-quadruplex sequence of the input mutation and obtain the optimal harmful mutation prediction category.

2. The method according to claim 1, characterized in that, Step 1 includes: Step 1.1: Obtain the set of mutation sites V = {v1, v2, ..., v} m ,...,v M }, where v m This represents the m-th mutation site and includes genomic coordinates and a harmful mutation label; Obtain any cell line h The set of coordinate intervals of the G-quadruplex in vivo measured below. vivo ; Step 1.2: Combine the set of variant sites V with G vivo Spatial intersection is performed, and based on the human reference genome, target variant sequences corresponding to sites falling within the G-quadruplex and its flanking sequence regions are generated, thereby constructing a target variant sequence dataset. .

3. The method according to claim 1, characterized in that, Step 2 uses IA 3 The strategy is to use Equation (2) to fine-tune the bond projection matrix K, value projection matrix V, and down projection matrix D in the self-attention layer of the nucleotide converter model, thereby obtaining the fine-tuned bond projection matrix. The fine-tuned projection matrix and the fine-tuned lower projection matrix : (2) In equation (2), ⊙ denotes element-wise multiplication. , , There are three scaling vectors to be learned.

4. The method according to claim 1, characterized in that, The negative sample set N in step 3.1 vitro The construction method is as follows: Dataset A based on G4-seq high-throughput sequencing technology was obtained from the gene expression comprehensive database; From dataset A, select the set of sequence coordinates G that can form G-quadruplex structures under in vitro experimental conditions. seq ; G seq With cell lines h The set of coordinate intervals of the G-quadruplex in vivo measured below. vivo Perform a difference operation and remove the overlapping regions of the two sets to obtain the set that folds only in vitro but in cell lines. h The sequence in the lower body did not form a G-quadruplex.

5. The method according to claim 1, characterized in that, Step 3.3 includes: Step 3.3.1 The auxiliary training dataset after one-hot encoding D’ aux Processing is performed to obtain sequence channel features. F DNA ; Step 3.3.2 For E epi Processing was performed to obtain epigenetic channel characteristics. F epi ; Step 3.3.3, for F DNA and F epi After feature fusion, the input is used for prediction in a fully connected layer, and the output is D. aux The predicted probability set of G-quadruplex in vivo .

6. The method according to claim 1, characterized in that, Step 6 includes: Step 6.1, Sequence semantic feature processing channel Ψ seq Using fully connected residual networks Perform a nonlinear mapping to obtain the i-th semantic deep representation. ; Step 6.2, Processing of key epigenetic features (branch Ψ) epi Using a one-dimensional convolutional neural network and a global pooling structure in sequence to... Perform local pattern extraction to obtain the i-th deep representation of the appearance. ; Step 6.3, on the channel dimension and After splicing, the data is input into the classification layer for processing to obtain... Predicted probability of harmful mutations ; Step 6.4: Construct the binary cross-entropy loss function using equation (3). : (3) Step 6.5: Minimize the binary cross-entropy loss function The parameters of the main dual-channel model are trained until... The training of the harmful mutation prediction model M is obtained by converging until convergence is achieved. main .

7. The method according to claim 6, characterized in that, The sequence semantic feature processing channel Ψ in step 6.1 seq It is obtained using equations (4) and (5) : (4) (5) In equations (4) and (5), Indicates random inactivation. , , , These represent the output dimensions as follows: , , , The fully connected layer, BN represents batch normalization, and Proj represents dimension-aligned projection layer. This indicates an element-wise addition operation. This represents the feature extraction operation of the feedforward network.

8. The method according to claim 6, characterized in that, The key epigenetic feature processing channel Ψ in step 6.2 epi It is obtained using equations (6)-(8) : (6) (7) (8) In equations (6)-(8), , These represent the number of filters respectively. , One-dimensional convolutional layer, This represents the feature map output by the initial convolutional operation of the first layer. and Let represent the feature maps output by the m-th and (m-1)-th residual convolutional blocks, respectively. This represents the feature map output by the last residual convolutional block in a fully connected residual network. This indicates max pooling. Indicates global average pooling. Indicates a modified linear element. Indicates the output dimension is The fully connected layer.

9. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports a processor in executing the method of any one of claims 1-8, the processor being configured to execute the program stored in the memory.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the method according to any one of claims 1-8.

Images