G-quadruplex whole-genome in silico sequencing method and system based on DNA language model

By employing a DNA language model-based whole-genome electronic sequencing method, utilizing a finely tuned DNA language model and sliding window scanning technology, combined with virtual read markers and read stacking region peak identification, the accuracy and efficiency issues of G-quadruplex whole-genome localization in existing technologies have been resolved, achieving efficient and accurate G4 enrichment.

WO2026237189A1PCT designated stage Publication Date: 2026-11-19YANGTZE DELTA REGION INST (QUZHOU) UNIV OF ELECTRONIC SCI & TECH OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2025/112837
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-16
Filing Date
2025-08-05
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Existing methods for whole-genome mapping of G-quadruplexes have limitations in their ability to identify non-classical G4s, are highly experimentally dependent, time-consuming and labor-intensive, and are prone to missing true positive sequences.

Method used

We employed a DNA language model-based G-quadruplex whole-genome electronic sequencing method. By constructing positive and negative sample datasets, we fine-tuned the DNA pre-trained language model, scanned the whole genome using a sliding window, and combined virtual read markers and read stacking region peak identification to achieve efficient G4 enrichment.

Benefits of technology

It achieves efficient detection of G4 across the entire genome, avoiding the problems of incomplete coverage and omissions in traditional technologies, and improving the accuracy and efficiency of detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2025112837_19112026_PF_FP_ABST
    Figure CN2025112837_19112026_PF_FP_ABST
Patent Text Reader

Abstract

Disclosed are a G-quadruplex whole-genome in silico sequencing method and system based on a DNA language model. The method comprises: constructing a sequence data set; fine-tuning a DNA pre-trained language model by using the sequence data set, so as to enable same to have a G4 structure recognition capability; by using the DNA language model after fine-tuning training, performing G4 prediction by performing, according to a predetermined window and a predetermined step size, sliding scanning on a whole genome to be subjected to prediction, where the predetermined window ≥ 100 nt, and the predetermined step size ≤ 10 nt; labeling sequence fragments having a prediction probability greater than a probability threshold as virtual reads, which are regarded as potential G4 signal fragments; and performing peak calling on reads-stacked regions, and outputting high-confidence G4-enriched regions. The in silico sequencing method is implemented by using the fine-tuned DNA language model, a sliding scanning mode in which the window is much larger than the step size is used, and G4 enrichment is achieved by means of virtual reads labeling and peak calling on the reads stacked regions, thus achieving efficient whole-genome detection, and overcoming the limitations of traditional technologies such as incomplete coverage, easy omission, or strong dependence on experiments.
Need to check novelty before this filing date? Find Prior Art

Description

A G-quadruplex whole-genome electronic sequencing method and system based on DNA language model Technical Field

[0001] This invention belongs to the field of bioinformatics technology, and in particular relates to a method and system for G-quadruplex whole-genome electronic sequencing based on a DNA language model. Background Technology

[0002] Existing methods for whole-genome localization of G-quadruplexes (G4) include sequence search methods, polymerase chaining (PCR) methods such as G4-seq, and antibody-mediated pull-down methods. Sequence search methods, as mentioned above, are based on classical rules and sequence scoring methods summarized from insights into resolved G4 structures, but their ability to identify non-classical G4s (including mismatches, protrusions, or incomplete quadruplexes) is limited. PCR can detect G4s throughout the genome, but due to differences between in vitro experimental conditions and real physiological environments, this method is also difficult to accurately represent G4 formation in living cells. Antibody-mediated pull-down methods are antibody-based techniques, but these are limited by the inherent cell type specificity of G4, are time-consuming and labor-intensive, and cannot provide a complete G4 map.

[0003] Machine learning methods have been proposed for G-quadruplex prediction, but few current models can achieve ideal whole-genome G4 prediction. The DeepG4 model and a previously proposed DNABERT-based G-quadruplex prediction method (publication number CN118447929B, which improves dataset construction and uses pre-trained DNABERT to enhance G-quadruplex prediction performance) can achieve whole-genome G4 prediction. However, both methods rely on large step sizes or binning the genome, making them prone to missing true positive sequences. Summary of the Invention

[0004] The purpose of this invention is to propose a G-quadruplex whole-genome electronic sequencing method and system based on a DNA language model to address the problems existing in the prior art.

[0005] A G-quadruplex whole-genome electronic sequencing method based on a DNA language model includes:

[0006] Construct a sequence dataset where positive samples contain the G4 structure and negative samples do not;

[0007] The DNA pre-trained language model was fine-tuned using the constructed sequence dataset to enable it to recognize G4 structures.

[0008] The fine-tuned DNA language model is used to perform G4 prediction by sliding scan of the whole genome to be predicted according to a predetermined window and a predetermined step size, with the predetermined window ≥ 100nt and the predetermined step size ≤ 10nt.

[0009] Sequence segments with a predicted probability greater than a probability threshold are labeled as virtual reads and considered as potential G4 signal segments.

[0010] Peak identification is performed on the stacked regions of reads to output the G4 enriched region with high confidence.

[0011] In the above electronic sequencing method, during the sliding scan, each slide is a predetermined step size, and the windows overlap by a predetermined window size minus the predetermined step size of bases. The same genomic region will be covered by multiple windows. In the example with a step size of 3nt and a window size of 200nt, a 300nt genomic region will be covered by approximately 33 windows.

[0012] The sequence fragment of each window is input into the fine-tuned DNA language model, and the fine-tuned DNA language model outputs the predicted probability that the corresponding window sequence fragment contains the G4 structure.

[0013] If the predicted probability is greater than the probability threshold, the corresponding window will be marked with virtual reads. The same genomic region may be marked with virtual reads multiple times to form read stacking regions. The more times it is marked, the higher the confidence level. When the read stacking exceeds the set number, the corresponding genomic region is considered to be a G4 enriched region.

[0014] In the above electronic sequencing method, Macs3 is used to perform peak finding on the read stacking region to screen genomic regions where the number of read stackings exceeds a set number as G4 enriched regions for output.

[0015] In the above electronic sequencing method, the sequence dataset is constructed using the following method:

[0016] Endogenous G4 data from different cell lines / tissues were obtained. The endogenous G4 data consisted of multiple experimental samples, and each experimental sample contained multiple G4 sequences.

[0017] The top N peaks with the highest G4 enrichment signal values ​​in each experimental sample were selected as representative peaks.

[0018] Sequence discovery was performed on the representative peaks to determine the G4 characteristic sequence;

[0019] Based on the G4 feature motif, the most matching motif position is found for each G4 sequence obtained from all experimental samples, and a sequence of a specific length is extracted as a positive sample in a center-aligned manner according to the found motif.

[0020] Negative samples are generated by randomly extracting human genome sequences that match the sequence characteristics of positive samples, including features such as length, GC content, and repetitive content.

[0021] In the above electronic sequencing methods, 500 ≤ N ≤ 2000;

[0022] Use MEME tool to perform sequence discovery on intrinsic G4 data;

[0023] The STORM tool was used to find the best-matching motif position for each sample sequence by scoring it using a position weight matrix.

[0024] In the above electronic sequencing method, 100nt ≤ predetermined window ≤ 300nt, and 2nt ≤ predetermined step size ≤ 10nt.

[0025] In the above-mentioned G-quadruplex whole-genome electronic sequencing method based on DNA language model, 180nt ≤ predetermined window ≤ 250nt, and 2nt ≤ predetermined step size ≤ 5nt.

[0026] In the above electronic sequencing method, 180nt ≤ predetermined window ≤ 200nt, and 2nt ≤ predetermined step size ≤ 3nt.

[0027] In the above-mentioned electronic sequencing method, before the whole genome is predicted by sliding scan, the whole genome is divided into positive and negative strands. The positive and negative strands are extracted in the 5' to 3' direction to maintain the consistency of sequence representation.

[0028] The DNA pre-trained language model is a GROVER model that has been pre-trained on over 5 million genome samples.

[0029] A G-quadruplex whole-genome electronic sequencing system based on a DNA language model includes a scanning module for performing G4 prediction by sliding scan of the whole genome to be predicted using a fine-tuned trained DNA language model according to a predetermined window and a predetermined step size, wherein the predetermined window is ≥100nt and the predetermined step size is ≤10nt.

[0030] The electronic sequencing module is used to label virtual reads for sequence fragments with a predicted probability greater than the probability threshold, and to identify peaks in the stacked regions of reads, outputting high-confidence G4 enriched regions.

[0031] The DNA language model is obtained by fine-tuning the DNA pre-trained language model using a dataset, and the dataset consists of sequences containing G4 structures in positive samples and sequences not containing G4 structures in negative samples.

[0032] The advantages of this invention are as follows: This scheme proposes an electronic sequencing method that utilizes a fine-tuned DNA language model, employs a sliding scan method with a window much larger than the step size, and achieves G4 enrichment through virtual read markers and peak identification of read stacking regions. This enables efficient detection of the entire genome and overcomes the limitations of traditional techniques, such as incomplete coverage, easy omissions, or strong experimental dependence. Attached Figure Description

[0033] Figure 1 is the overall flowchart of this solution;

[0034] Figure 2 is a flowchart of the sequence dataset construction method in this scheme;

[0035] Figure 3 is a flowchart of the data processing workflow during the dataset construction process in this scheme;

[0036] Figure 4 is a schematic diagram of the G4-GROVER model and electronic sequencing scheme in this scheme;

[0037] Figure 5 is a comparison of the model G4-GROVER implemented in this scheme with G4detector-retrain and PENGUINN;

[0038] Figure 6A shows the attention heatmap of the G4-GROVER model in this scheme;

[0039] Figure 6B shows the eight heads in Figure 6A that show obvious interest in the region with high G content in the middle;

[0040] Figure 7 shows the high-attention tokens and predicted G4 motifs analysis of the G4-GROVER model in this scheme;

[0041] Figure 8 shows the probability distribution and genome-wide results analysis of the model implemented in this scheme at 61 locations centered on G4.

[0042] Figure 9 is a structural block diagram of the G-quadruplex whole-genome electronic sequencing system based on the DNA language model in this scheme. Detailed Implementation

[0043] This approach provides a G-quadruplex whole-genome electronic sequencing method based on a DNA language model. It utilizes large language models (LLMs) in genomics, which, through training on DNA sequence information, learn unique principles and reveal hidden biological meanings using models based on BERT or GPT architectures. GROVER is a novel DNA language model pre-trained on over 5 million genome samples and uses byte-pair tokenization to analyze the human genome. Its vocabulary is selected through a custom task—next k-mer prediction—and the model performs optimally when the final selected token size is 609.

[0044] This approach employs a pre-trained GROVER model for fine-tuning predictions on the G4 dataset, resulting in a G4-GROVER model. Based on this model, an electronic sequencing method (e-seq) for whole-genome G4 enrichment is proposed. The genome is scanned using G4-GROVER; this example uses a 200nt window size and a 3nt scan step, but other sizes can be substituted for appropriate values. Reads exceeding a threshold are then used to enrich peaks. In this example, the probability threshold for G4-GROVER is set to 0.9, and Macs3 (Model-based Analysis of ChIP-Seq3, a Python-based software) is used to identify G4 peaks, yielding G4-enriched regions in the human genome. Compared to binning methods, e-seq not only avoids missing G4 peaks but also identifies high-confidence G4 peaks supported by multiple reads, thus achieving efficient and accurate whole-genome G4 prediction.

[0045] As shown in Figure 1, the specific processing workflow given in this embodiment is as follows:

[0046] As shown in Figures 2 and 3, the dataset was constructed by downloading endogenous G4 (eG4) data from 35 different cell lines and tissues (93T449, A549, AB521M, AB551, AB555M, AB577M, AB580, AB636M, AB790, AB863M, CNCC, ESC, H1975, HACAT, HCI005, HCI009, HEK293T, HeLa.S3, HepG2, IPSC.derived_neurons, K562, NHEK, NSC, PAR1006, PAR1022, STG139, STG139M, STG143, STG195M, STG201, STG282M, STG316, STG331, VHIO098, VHIO719) from the EndoQuad database.

[0047] Motif discovery was performed using MEME (Multiple Em for Motif Elicitation). The database contained 162 experimental samples, each containing multiple G4 sequences. The top 1000 signal values ​​from each sample were selected. The number of sequences obtained varied from sample to sample, resulting in 145,440 eG4 sequences. These sequences had an average length of 372 base pairs (bp), with the shortest being 59 bp. This protocol used 100 bp DNA fragments on either side of the peak center to input into MEME for de novo DNA motif discovery. The motif with the highest confidence (i.e., the most frequent occurrences) was selected as the characteristic G4 motif. Next, STORM was used to locate the center of the G4 for each sample and align the DNA sequences.

[0048] For the positive sample dataset, the following steps are used to obtain negative (control) sequences: sequences are randomly extracted from the human genome using the gkmSVMR package to serve as negative samples, with length, GC content (%GC), and repetitive content similar to positive DNA sequences.

[0049] The new dataset was partitioned using the following steps: CD-HIT was used to remove duplicate sequences with a similarity greater than 80%. The resulting positive set contained 619,133 eG4 sequences. Negative samples were also partitioned in the same way, removing sequences with a similarity greater than 80%. The model was then trained, tested, and validated. To ensure the randomness of the positive and negative sample data distribution, chromosome 1 was chosen as the independent test set to validate the model's performance. The model input data was randomly allocated into training, testing, and validation sets in an 8:1:1 ratio.

[0050] G4 enrichment region acquisition: As shown in Figure 4A, a pre-trained GROVER is used as the base model, and the constructed dataset is used to fine-tune and train it to obtain the G4-GROVER model. The G4-GROVER model is used to extract features from the G4 sequence.

[0051] GROVER employs a Transformer model encoder structure, which links all relevant DNA bases in the context from different perspectives to extract features from the DNA sequence. The input sequence can be a single sentence or a concatenation of two sentences. The input is a raw sequence of less than 521 bytes, with [CLS] appended to the first position of the input, and two sentences separated by a special label [SEP]. During training, word embeddings are used to convert each input sequence into a latent space representation (N, 768), where N represents the number of labels following the sentence label.

[0052] The G4-GROVER model is obtained by fine-tuning a pre-trained GROVER model. Its architecture is shown in Figure 4A, containing 12 encoder layers. Each encoder layer mainly consists of a multi-head self-attention layer and a feedforward layer. When a DNA sequence is input, the tokens obtained from word segmentation are feature-mapped, and position-encoded features are added. Then, the embedded feature vectors are merged and input into the encoder. Finally, the model inputs the classification head [CLS] into the classifier for binary classification prediction. The Embeddings module takes a 200nt DNA sequence as input, and the tokens generated by the word segmenter are the feature representations. The word embedding layer encodes each token according to the genomic vocabulary of the pre-trained GROVER model, which has a vocabulary size of 609. Position embeddings and token type embeddings are added to the embeddings. Fine-tuning the GROVER model involves feeding the output of the InputEmbedding layer into the Transformer's encoder module. The encoder consists of 12 layers, each containing multiple self-attention mechanisms and layer normalization. During normalization, the input from the previous layer is normalized together with the output of the current layer. The generated output is then passed through a feedforward network and normalized to ensure stability and consistency.

[0053] The goal of this design is to maximize model capacity while ensuring that the model can be fine-tuned on a single GPU card. Specifically, the encoder structure of the Transformer model captures upstream and downstream information M of DNA bases by implementing a multi-head self-attention mechanism.

[0054] MultiHead(M)=Concat(head1,…,head h W o (1)

[0055] in:

[0056] W o and These are the learning parameters for linear projection. d k yes The number of columns in the matrix, i.e., the vector dimension. `head` calculates the next hidden state `M`. `MultiHead()` combines multiple... The independently calculated head results are then concatenated.

[0057] The inputs and outputs of the multi-head self-attention layer are summed and residual connections are made to prevent network degradation. Then, layer normalization is performed to normalize the activation values ​​of each layer.

[0058] The feedforward network layer is a two-layer fully connected layer. The first layer uses ReLU as the activation function, while the second layer does not use an activation function. The corresponding formula is as follows:

[0059] max(0,XW1+b1)W2+b2(3)

[0060] Here, X represents the input to the feedforward network layer, which is also the output of the normalization layer. The final output matrix of the feedforward network layer has the same dimension as X. The result of adding the input and output of the feedforward network is then normalized to prevent network degradation.

[0061] The entire encoder module is executed 12 times, with 12 layers. Ot represents the final hidden state of label t. Finally, the output of the encoder module's hidden states serves as the features for the final DNA sequence extraction.

[0062] Prediction is performed using fully connected layers. After feature fusion, the fully connected layers map the fused features to a low-dimensional representation space to learn the relationship between samples and labels. The features are fed into one or more fully connected layers, and activation functions are applied to achieve binary classification of G4.

[0063] Finally, the Adam optimizer and cross-entropy loss were used as the loss function. All training was performed on a single 80G NVIDIA A800 GPU card with a learning rate of 1e-5 and a batch size of 64. Thus, the finely tuned G4-GROVER model was obtained through the above methods.

[0064] As shown in Figure 4B, this embodiment uses a finely tuned G4-GROVER to slide scan the genome with windows of 200 nt size and a step size of 3 nt. During the sliding scan, each slide is 3 nt, and the windows overlap by 197 base pairs, so the same genomic region will be covered by multiple windows. The sequence fragment of each window is input into the finely tuned G4-GROVER, which outputs the predicted probability that the corresponding window sequence fragment contains a G4 structure. If the predicted probability is greater than a probability threshold, the corresponding window is labeled with virtual reads (real reads are short DNA fragment sequences generated through physical experiments; this scheme proposes the concept of virtual reads, which are generated based on the predicted probability output by the G4-GROVER model and the designed probability threshold, and are used to represent high-probability candidate G4 regions in the genome scan). The same genomic region may be labeled with virtual reads multiple times, forming read stacking regions. The more times it is labeled, the higher the confidence level. When the read stacking exceeds a set number, the corresponding genomic region is considered a G4-rich region.

[0065] In this embodiment, a probability threshold of 0.9 is selected. Potential G4 sequences are chosen based on this threshold, with each sequence considered a "read." Reads are stacked on top of each nucleotide. Peak finding is then performed on these reads, and finally, G4 enrichment is determined from the found peaks. Specifically, Macs3 is used to perform peak finding on the read stacking regions to filter genomic regions where reads stack more than a set number of times as the output G4 enriched regions.

[0066] To verify the feasibility and effectiveness of this solution, a performance evaluation was conducted in this embodiment. This embodiment fine-tuned and trained a large language model G4-GROVER on millions of sequence samples. As can be seen from A in Figure 5, the model accuracy is as high as AUROC = 0.994.

[0067] For comparison, our model was benchmarked against state-of-the-art sequence-based models PENGUINN and G4detector. Both models are based on convolutional neural networks (CNNs) and were initially trained on the G4-seq dataset, but their convolutional architectures and input sequence lengths differ significantly. Specifically, G4detector employs one-dimensional convolutional layers with different kernel configurations, while PENGUINN contains multiple convolutional layers with varying kernel sizes. Furthermore, PENGUINN uses a fixed starting sequence length of 200bp, while G4detector initially uses 124bp.

[0068] To evaluate performance, G4-GROVER was first compared with PENGUINN using an independent test dataset. As shown in Figures 5A and 5B, G4-GROVER significantly outperforms PENGUINN. For G4detector, it was retrained and evaluated on the dataset used in this example, with its input sequence length adjusted to 200bp to maintain data consistency. The resulting model is called G4detector-retrain. As shown in Figures 5A and 5B, G4-GROVER also exhibits a significant advantage over both G4detector-retrain and PENGUINN.

[0069] G4-GROVER's superior performance is primarily attributed to its use of an attention mechanism, which effectively integrates local and global contextual information from the sequence. Furthermore, its unsupervised pre-training on a large amount of human genome data enables the model to learn highly generalizable sequence representations, significantly improving prediction accuracy compared to existing methods.

[0070] Figures 6A, 6B, and 7 show the attention heatmap and the high-attention token and predicted G4 motif analysis diagram of the model implemented in this scheme, respectively.

[0071] Figure 6A shows an attention head view of an endogenous G4 in the test set. It can be seen that in the 4th layer of the G4-GROVER model, heads 10, 13, 14, 26, 27, 58, 96 and 122 show obvious attention to the regions with high G content in the middle (i.e. AGGG and TGGG).

[0072] Figure 6B extracts the attention matrices for the eight heads in Figure 6A, including the G-run attention scores. The first AGGG receives the most attention from the model's 10th, 14th, 26th, and 96th heads. The second AGGG motif is primarily handled by head 27, indicating that the model captures the positional features of the AGGG. The third TGGG receives the most attention from the model's 10th and 26th heads. The fourth AGGG receives the most attention from the model's 58th head.

[0073] Figure 7 analyzes the markers whose analytical attention exceeds two standard deviations outside the mean, revealing that the AGGG score of the model is abnormally significant. Simultaneously, motif analysis of G4 in the test set also shows a preference for high-purine patterns such as AGGG.

[0074] Finally, this embodiment performed a genome-wide identification and comparison of G4 candidates.

[0075] Besides DeepG4, few machine learning models can achieve genome-wide G4 prediction. DeepG4 divides the genome into 200bp bins and labels the bins with active G4 regions. However, bin-based scanning can miss true G4s. The small-incremental continuous scanning method implemented in this approach effectively avoids missing true positives. In Figure 8A, when using a step size of 3nt, G4s are rarely missed.

[0076] Using G4-GROVER with a 200nt sliding window and a 3nt step size to scan the entire genome, 52,083,920 candidate reads were obtained. These reads successfully recalled 603,428 eG4 structures, corresponding to a recall rate of 97.46%, demonstrating the effectiveness and robustness of this method in comprehensively capturing G4 structures. Peak finding was then performed using Macs3, identifying a total of 1,615,830 G4 peaks.

[0077] The aforementioned enriched peak (G4-GROVER Peak) was compared with candidate G4s obtained from other genome-wide G4 prediction methods (including G4-Hunter, G4-seq, and Pqsfinder). As shown in Figure 8B, there is a high degree of overlap between the predictions from G4Hunter1.5 and Pqsfinder, indicating that these two sequence search-based methods predict G4s with similar characteristics, but also miss more experimentally detected G4s. The G4-GROVER prediction shows an even higher overlap with the G4s identified experimentally by G4-seg.

[0078] Furthermore, as shown in Figure 9, this embodiment also provides a G-quadruplex whole genome electronic sequencing system based on a DNA language model, including a scanning module 1, used to perform G4 prediction by sliding scan of the whole genome to be predicted according to a predetermined window and a predetermined step size using a fine-tuned trained DNA language model, wherein the predetermined window is ≥100nt and the predetermined step size is ≤10nt.

[0079] Electronic sequencing module 2 is used to label virtual reads for sequence fragments with a predicted probability greater than the probability threshold, and to identify peaks in the stacked regions of reads, outputting a high-confidence G4 enriched region.

[0080] The DNA language model was obtained by fine-tuning the DNA pre-trained language model using a dataset, which consisted of sequences containing G4 structures in positive samples and sequences not containing G4 structures in negative samples.

[0081] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A DNA language model-based G-quadruplex whole-genome electronic sequencing method, characterized in that, include: Construct a sequence dataset where positive samples contain the G4 structure and negative samples do not; The DNA pre-trained language model was fine-tuned using the constructed sequence dataset to enable it to recognize G4 structures. The fine-tuned DNA language model is used to perform G4 prediction by sliding scan of the whole genome to be predicted according to a predetermined window and a predetermined step size, with the predetermined window ≥ 100nt and the predetermined step size ≤ 10nt. Virtual reads are labeled for sequence segments whose predicted probability is greater than a probability threshold; Peak identification is performed on the stacked regions of reads to output high-confidence G4 enriched regions; During the sliding scan, each slide is a predetermined step size, and the windows overlap by a predetermined window size minus the predetermined step size of bases. The same genomic region will be covered by multiple windows. The sequence fragment of each window is input into the fine-tuned DNA language model, and the fine-tuned DNA language model outputs the predicted probability that the corresponding window sequence fragment contains the G4 structure. If the predicted probability is greater than the probability threshold, the corresponding window will be marked with virtual reads. The same genomic region may be marked with virtual reads multiple times to form read stacking regions. The more times it is marked, the higher the confidence level. When the read stacking exceeds the set number, the corresponding genomic region is considered to be a G4 enriched region.

2. The electronic sequencing method of claim 1, wherein, Use Macs3 to perform peak finding on read stacking regions to filter genomic regions where read stacking exceeds a set number of times as G4 enriched regions output.

3. The method of electronic sequencing of claim 1, wherein, Construct a sequence dataset using the following method: Endogenous G4 data from different cell lines / tissues were obtained. The endogenous G4 data consisted of multiple experimental samples, and each experimental sample contained multiple G4 sequences. The top N peaks with the highest G4 enrichment signal values ​​from each experimental sample were selected as representative peaks. Sequence discovery was performed on the representative peaks to determine the G4 characteristic sequence; Based on the G4 feature motif, the most matching motif position is found for each G4 sequence, and a sequence of a specific length is extracted as a positive sample according to the motif center alignment method. Negative samples are generated by randomly extracting human genome sequences and matching the sequence characteristics of positive samples.

4. The method of electronic sequencing according to claim 3, wherein, 500≤N≤2000; Use MEME tool to perform sequence discovery on intrinsic G4 data; The STORM tool was used to find the best-matching motif position for each sample sequence by scoring it using a position weight matrix.

5. The method of electronic sequencing of claim 1, wherein, 100nt≤ReservedWindow≤300nt, 2nt≤ReservedStep≤10nt.

6. The method of electronic sequencing according to claim 5, wherein, 180nt≤Reserved Window≤250nt, 2nt≤Reserved Step Size≤5nt.

7. The method of electronic sequencing according to claim 6, wherein, 180nt≤Reserved Window≤200nt, 2nt≤Reserved Step Size≤3nt.

8. The method of electronic sequencing of claim 1, wherein, Before performing a sliding scan to predict the entire genome, the entire genome is divided into positive and negative strands; The DNA pre-trained language model is a GROVER model that has been pre-trained on over 5 million genome samples.

9. A DNA language model-based G-quadruplex whole-genome electronic sequencing system, characterized by, include, The scanning module (1) is used to perform G4 prediction by sliding the DNA language model trained by fine-tuning according to a predetermined window and a predetermined step size, and the predetermined window is ≥100nt and the predetermined step size is ≤10nt. The electronic sequencing module (2) is used to label virtual reads for sequence fragments with a predicted probability greater than the probability threshold, and to identify peaks in the stacked regions of reads, outputting a high-confidence G4 enriched region. The DNA language model is obtained by fine-tuning the DNA pre-trained language model using a dataset, and the dataset consists of sequences containing G4 structures in positive samples and sequences not containing G4 structures in negative samples. During the sliding scan, each slide is a predetermined step size, and the windows overlap by a predetermined window size minus the predetermined step size of bases. The same genomic region will be covered by multiple windows. The sequence fragment of each window is input into the fine-tuned DNA language model, and the fine-tuned DNA language model outputs the predicted probability that the corresponding window sequence fragment contains the G4 structure. If the predicted probability is greater than the probability threshold, the corresponding window will be marked with virtual reads. The same genomic region may be marked with virtual reads multiple times to form read stacking regions. The more times it is marked, the higher the confidence level. When the read stacking exceeds the set number, the corresponding genomic region is considered to be a G4 enriched region.