Method capable of improving methylation detection accuracy of single-molecule 5mC based on CCS

Abstract

The invention discloses a method capable of improving CCS-based single-molecule 5mC methylation detection accuracy. The method comprises the following steps: acquiring a native data set and a processed data set, and constructing a training data set and a test data set; constructing a single-molecule CpG methylation recognition module based on a graph convolution deep neural network, and performing training by using training data of different background base lengths; testing the background base length and the model performance, selecting a proper base length and training a corresponding model; and performing single-molecule 5mC methylation detection by using the trained single-molecule CpG methylation recognition module based on the graph convolution deep neural network. The whole genome methylation frequency quantitative result of the read counting mode has superior performance in all test data sets including different data sources, sequencing library lengths or sequencing depths. In addition, through high CpG methylation recognition, complex single-molecule CpG methylation modes, including methylation modes of haplotypes and repeated regions, can be found.

Description

Technical Field

[0001] The present invention belongs to the technical field of methylation detection, and in particular relates to a method for improving the accuracy of single-molecule 5mC methylation detection based on CCS. Background Art

[0002] DNA methylation is an epigenetic mechanism that changes chromatin structure and function without changing the DNA sequence. In animals, DNA methylation mainly occurs on cytosine bases, producing 5-methylcytosine (5mC) at CpG sites. This mechanism plays an important role in the normal development and functional maintenance of cells, and is involved in regulating biological processes such as genomic imprinting, X chromosome inactivation, transposable element silencing and aging. It can also interact with other epigenetic features such as nucleosome positioning, histone modification and three-dimensional genome organization. DNA methylation detection at the single-molecule level can obtain the simultaneous methylation status of bases in continuous regions on a single molecule, which helps to accurately identify the methylation patterns between alleles and in genomic repetitive regions. In the future, combining accurate single-molecule methylation detection technology with single-cell technology is expected to reveal the role of DNA methylation in cell differentiation and carcinogenesis in a deeper way.

[0003] In DNA methylation research, resolving the methylation status of single molecules is a key challenge. Next-generation sequencing (NGS) methods, such as bisulfite sequencing (BS-seq), have limitations in capturing single-molecule methylation patterns due to their short read lengths. In contrast, single-molecule (long-read) sequencing technologies, such as PacBio and Nanopore, can generate longer reads and methylation-related signals from native DNA molecules without the need for additional DNA processing. Nanopore sequencing can generate ultra-long reads of up to several megabases and has high sensitivity for detecting 5mC. However, its high sequencing error rate can lead to the loss of CpG sites and inaccurate methylation detection results.

[0004] The recently released Circular Consensus Sequencing (CCS) mode of the Pacbio Sequel II platform can generate highly accurate (>99%) reads longer than 20 kb by sequencing single molecules multiple times, which helps to overcome the low signal intensity of 5mC. Therefore, CCS is a promising technology for accurately revealing single-molecule methylation patterns. However, the accuracy of current CCS-based 5mC detection algorithms at the single-molecule level is limited (<90% accuracy for most tested data sets). These algorithms infer the methylation status of CpG motifs from the kinetic information of polymerases, including pulse width (PW) and inter-pulse duration (IPD), and this kinetic signal is affected by the background bases nearby in the sequence. Previous studies have shown that the kinetic signal of nucleotides is mainly affected by the 7 upstream bases and the 2 downstream bases. Therefore, current algorithms, such as primrose and ccsmeth, use short k-mer (≤21 bp) upstream and downstream base information for methylation identification, but many signals are not reliable and the overall accuracy is not high. Among them, the deep learning-based method ccsmeth is the most effective, achieving an accuracy of almost 0.90. In mammalian genomes, CpG sites tend to cluster and form CpG-rich regions, such as CpG islands, which tend to have synchronized methylation states between CpGs within them. Previous studies have also found that there is a significant correlation in methylation frequencies between adjacent CpG sites in mammals and plants. Based on this observation, some DNA methylation technologies have successfully improved the quantification of genome-wide methylation frequencies. The correlation of methylation frequencies between adjacent CpGs (population level) indicates that CpG methylation states may also be associated at the single-molecule level, which may provide additional information for determining the methylation state of low-confidence CpGs and improve the accuracy of methylation detection.

[0005] The second-generation bisulfite sequencing cannot capture single-molecule methylation patterns due to its short read length; the third-generation nanopore sequencing has a high sequencing error rate, which leads to the loss of some CpG sites and inaccurate detection results; the CCS of the third-generation Pacbio Sequel II platform infers the methylation state of the CpG motif based on the kinetic information of the polymerase, but this kinetic information is affected by the background bases near the sequence. The current algorithm uses the upstream and downstream base information of short k-mer (≤21 bp) for CCS reads to identify methylation, which is not accurate enough. Summary of the invention

[0006] The main purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and to provide a method for improving the accuracy of single-molecule 5mC methylation detection based on CCS.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions: One aspect of the present invention provides a method for improving the accuracy of single-molecule 5mC methylation detection based on CCS, characterized in that it comprises the following steps: Obtain native data sets and processed data sets, and construct training data sets and test data sets; Construct a single-molecule CpG methylation recognition module based on graph convolutional deep neural network and train it using training data of different background base lengths; Test the background base length and model performance, select the appropriate base length and train the corresponding model; Single-molecule 5mC methylation detection was performed using the trained single-molecule CpG methylation recognition module based on graph convolutional deep neural network.

[0008] As a preferred technical solution, the acquisition of the original data set and the processed data set, and the construction of the training data set and the test data set are specifically as follows: The native data set is obtained by downsampling a DNA sequencing vector chip containing CCS reads of the native human HG002 genome reference sample; The processed data set includes fully methylated CCS data obtained by treating DNA with CpG methyltransferase M.SssI and unmethylated data obtained from PCR amplified DNA, which are used for training the read count model; Construct training datasets using native datasets and processed datasets; The test data set includes a training data set, and also includes a native human HG002 genome reference sample data set of multiple DNA sequencing carrier chips, a native CHM13 data set, a native zebrafish data, and processed native human HG002 genome reference sample data.

[0009] As a preferred technical solution, the single-molecule CpG methylation recognition module based on graph convolutional deep neural network includes 12 blocks connected in sequence, wherein the first 11 blocks each include a batch normalization layer, a graph convolution layer and a rectified linear unit activation function, and the last block includes a batch normalization layer, a graph convolution layer and a global average pooling layer, and also includes a feedforward layer.

[0010] As a preferred technical solution, the input and output of the CpG methylation recognition module based on graph convolutional deep neural network are specifically: Given a read segment R = r0r1...rn-1 This sequence is composed of the alphabet Σ = (A, C, G, T). Given a CpG site p, where 0 ≤ p < n, a background base Kp of length k is extracted, i.e., Kp = r p-h r p-h+1 ...r p+h-1 , where h = k / 2; When p < h or n - p < h, the first base r0 or the last base r of the sequence is extracted n-1 ; Construct a graph Gp for Kp as the input; In the graph Gp, each base r in Kp i represents a node Ni, and two adjacent bases ri and ri +1 are connected by a directed edge ei = r i ← r i+1 So the graph Gp contains k nodes and k - 1 directed edges. The edges ei in Gp do not contain any features; each node Ni has 15 features fi ∈ ℝ¹ 5 , which include base information, sequencing quality, pulse interval time and pulse width of the positive strand, and pulse interval time and pulse width of the negative strand, and are summarized in a two-dimensional tensor of shape k × 15; The output contains two probability values p0 and p1, and p0 + p1 = 1, representing the probabilities of the methylated and non-methylated states of the CpG site respectively.

[0011] As a preferred technical solution, the training process of the single-molecule CpG methylation recognition module based on the graph convolutional deep neural network is specifically as follows: It is trained through the cross-entropy loss function and the SGD optimizer. The learning rate of the SGD optimizer is 0.1, and the weight decay is 10 -5 , and the momentum and Nesterov accelerated gradient are 0.9; After each training epoch, the StepLR learning rate scheduler is used to decay the learning rate by 0.1 times; each single-molecule CpG methylation recognition module based on the graph convolutional deep neural network is trained for two training epochs, or a pre-trained model can be selected as the starting point of training.

[0012] As a preferred technical solution, the background base length and model performance test are specifically as follows: The data in the CCS reads of bisulfite sequencing performed on the native human HG002 genome reference sample in the native dataset were used to identify fully methylated and completely unmethylated CpG sites, and positive and negative references were constructed, respectively. Several data were randomly sampled from the native dataset to evaluate the model multiple times, and the randomly sampled data did not include the data used for training.

[0013] As a preferred technical solution, the selection of a suitable base length and training of a corresponding model are specifically as follows: The K-20 model, i.e., the model with 20-mer background base length, was selected, and positive and negative CpG features were extracted from the dataset for training and testing respectively; The K-20 model was used as a pre-training model for training the K-400 model, where the K-400 model refers to a model with a background base length of 400-mer, and the native dataset was supplemented with the processed dataset for training and testing, where the training datasets included: using the K-20 model to select CpG sites in the dataset with a methylation probability ≥ 0.9 or ≤ 0.1, as positive and negative samples respectively, and for the remaining CpG sites, CpG sites that were aligned to the reference CpG sites and had a methylation frequency ≥ 95% or ≤ 5% were selected as positive and negative samples respectively, where the methylation frequency was determined by BS-seq.

[0014] As a preferred technical solution, single-molecule CpG methylation correlation analysis is performed, specifically: A number of data were extracted from a DNA sequencing vector chip containing native human data and a DNA sequencing vector chip containing native zebrafish data, and the K-20 model, i.e., a model with a 20-mer background base length, was used to determine the methylation status of high-confidence CpGs, i.e., CpGs with a methylation probability ≤ 0.1 were considered negative, and CpGs with a methylation probability ≥ 0.9 were considered positive; Treating the methylated and unmethylated states as two different alleles, the correlation of methylation between high-confidence CpGs at different distances is calculated, specifically: ; Among them, F MM is the frequency of methylation of two CpGs at a given distance, F M is the frequency of methylation of a single CpG in a CpG at a given distance; Chi-square tests were performed to assess statistical significance and corrected for multiple testing using the Benjamini-Hochberg test.

[0015] As a preferred technical solution, the methylation frequency quantification module of hifimeth includes a read segment counting mode; The read counting mode only considers read data with alignment quality MAPQ ≥ 10, specifically: the number of reads exceeding the positive probability threshold at a CpG site is recorded as the methylated read number Nm, and the number of reads below the negative probability threshold is recorded as the unmethylated read number Nu, and the methylation frequency of CpG is calculated using Nm / (Nm + Nu).

[0016] Another aspect of the present invention further provides a system capable of improving the accuracy of single-molecule 5mC methylation detection based on CCS, which is applied to the above-mentioned method capable of improving the accuracy of single-molecule 5mC methylation detection based on CCS, comprising a data acquisition module, a model building module, a background base length selection module and a methylation detection module; The data acquisition module is used to acquire the original data set and the processed data set, and construct the training data set and the test data set; The model building module is used to build a single-molecule CpG methylation recognition module based on a graph convolutional deep neural network, and is trained using training data of different background base lengths; The background base length selection module is used to test the background base length and model performance, select a suitable base length and train the corresponding model; The methylation detection module is used to perform single-molecule 5mC methylation detection using a trained single-molecule CpG methylation recognition module based on a graph convolutional deep neural network.

[0017] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This application uses human and zebrafish CCS data to detect 5mC at the single-molecule level with an average accuracy of 94.7% and an average F1 score of 94.2%. The genome-wide methylation frequency quantification results of the read count mode of the present invention are superior in all test data sets including different data sources, sequencing library lengths or sequencing depths. In addition, the high-accuracy CpG methylation recognition of the present application can discover complex single-molecule CpG methylation patterns, including methylation patterns of haplotypes and repeat regions. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of a method for improving the accuracy of single-molecule 5mC methylation detection based on CCS in an embodiment of the present invention.

[0019] Figure 2 It is a curve diagram of the correlation of CpG methylation status at the single molecule level within the distance of 300 bp and 10000 bp in the embodiment of the present invention.

[0020] Figure 3Schematic diagram of the effect of background length and CpG density on the performance of a deep neural network for DNA methylation inference in an embodiment of the present invention.

[0021] Figure 4 It is a schematic diagram comparing the performances of hifimeth, ccsmeth and primrose at the single molecule level in the embodiments of the present invention.

[0022] Figure 5 is a schematic diagram of using hifimeth to quantify the CpG methylation frequency of the whole genome in an embodiment of the present invention, wherein count and model represent quantification using a read count mode and a site frequency model mode, respectively.

[0023] Figure 6 Schematic diagram of differential methylation analysis of hifimeth in an embodiment of the present invention.

[0024] Figure 7 It is a schematic diagram of the methylation recognition results based on CCS using different probability thresholds in an embodiment of the present invention.

[0025] Figure 8 It is a schematic diagram of the evaluation results of typing methylation frequency in an embodiment of the present invention.

[0026] Fig. 9 It is a schematic diagram of the typing methylation pattern of the CYREN gene in HG002 in the embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application.

[0028] Example: like Figure 1 As shown, this embodiment provides a method (hifimeth) for improving the accuracy of single-molecule 5mC methylation detection based on CCS, comprising the following steps: S1. Obtain the original data set and the processed data set, and build the training data set and the test data set.

[0029] In this embodiment, the training dataset includes a native dataset and a treated dataset. The native training data is obtained by downsampling a DNA sequencing flowcell (hereinafter referred to as flowcell) containing native human HG002 (HG002 refers to the reference genome of a human male sample) CCS reads from the study by Baid et al. The insert size of these reads is 15 kb (native human data, hereinafter referred to as ND, with a coverage of approximately 13.5×). The treated NA12898 dataset is from the study by Ni et al. and is used for read model training. This dataset contains fully methylated CCS data (hereinafter referred to as FMD) obtained by treating DNA with the CpG methyltransferase M.SssI and unmethylated data (hereinafter referred to as UMD) obtained by PCR amplifying DNA.

[0030] The test CCS dataset involves the above training datasets and also includes the native HG002 dataset from five flowcells from the studies by Baid et al. and the Human Pangenome Reference Consortium (HPRC), the native CHM13 dataset from three flowcells (library insert size of 20 kb), the native zebrafish data from one flowcell, the treated HG002 data (for the training of primrose), and the M2&W2 data.

[0031] S2. Construct a single-molecule CpG methylation recognition module based on a graph convolutional deep neural network and train it using training data with different background base lengths.

[0032] Furthermore, the single-molecule CpG methylation recognition module based on the graph convolutional deep neural network includes 12 blocks (blocks) connected in sequence and a total of several parameters (a total of 368,928 parameters). Among them, the first 11 blocks each include a batch normalization layer, a graph convolutional layer, and a rectified linear unit activation function. The last block includes a batch normalization layer, a graph convolutional layer, and a global average pooling layer, and also includes a feed-forward layer. The batch normalization layer and the graph convolutional layer are implemented through the pyg library.

[0033] Furthermore, the input and output of the CpG methylation recognition module based on the graph convolutional deep neural network are specifically as follows: Given a read R = r0r1...r n-1 , which is composed of the alphabet Σ = (A, C, G, T), and given a CpG site p (0 ≤ p < n), in this embodiment, a k-mer Kp = r p-h r p-h+1 ...r p+h-1, where h = k / 2. When p < h or n - p < h, the first base r0 or the last base r of the sequence is extracted n-1 . Construct a graph Gp for Kp, which is used as the input of hifimeth. In the graph Gp, each base r in Kp i represents a node Ni. Two adjacent bases ri and ri +1 are connected by a directed edge ei = r i ← r i+1 . Therefore, the graph Gp contains k nodes and k - 1 directed edges. The edges ei in Gp do not contain any features. Each node Ni has 15 features fi ∈ ℝ¹ 5 , and these features include base information, sequencing quality, interpulse duration (IPD) and pulse width (PW) of the positive strand, IPD and PW of the negative strand. These features are summarized in a two-dimensional tensor of shape k × 15.

[0034] The output contains two probability values p0 and p1, representing the probabilities of the methylated and non-methylated states of the CpG site respectively. Obviously, p0 + p1 = 1. If a reference genome (optional) is given, the CCS bam file containing polymerase kinetics information will be aligned to the reference genome by the pbmm2 software. Then the positions of each identified CpG in the reference genome and the read will be output, as well as p0 and p1. In addition, a BAM file will be generated, which contains the 5mC base modification values in the MM and ML tags, following the tag format described in the SAM specification file.

[0035] Further, this embodiment is trained by the cross-entropy loss function and the SGD (stochastic gradient descent) optimizer, and the training process is as shown in Figure 3 A in it. The SGD optimizer has a learning rate of 0.1, a weight decay of 10 -5 , a momentum of 0.9 and Nesterov accelerated gradient. After each epoch (after a training round), the StepLR learning rate scheduler is used to decay the learning rate by 0.1 times to accelerate the convergence speed during training. Each single-molecule CpG methylation recognition module based on the graph convolutional deep neural network is trained for two epochs. It is also possible to choose to use a pre-trained model as the starting point of training.

[0036] S3. Test the background base length and the model performance, where the model performance test includes test accuracy, AUC, precision, recall and F1 score.

[0037] A training dataset of 100,000 reads from ND was used, and HG002 BS-seq (referring to bisulfite sequencing performed on the HG002 genome reference sample) data in CCS reads was used to identify fully methylated and fully unmethylated CpG sites to construct positive and negative references, respectively. The model was evaluated by randomly sampling 10,000 reads (about 1.5Gb) from the ND dataset, which did not include the 100,000 reads used for training. This evaluation was repeated three times to ensure the robustness of the results.

[0038] S4. Select the appropriate base length and train the corresponding model.

[0039] The K-20 model, which is a model with a 20-mer background base length, was selected and trained: positive and negative CpG features were extracted from 400,000 FMD and 400,000 UMD reads, respectively. The test dataset includes 5,000 FMD and 5,000 UMD reads. The K-20 model was used as a pre-trained model for training the K-400 model (trained with a 400-mer background base length).

[0040] Training the K-400 model (i.e., a model with a 400-mer background base length): The training dataset mainly uses the native DNA dataset, supplemented by a small amount of processed datasets. For the native data, 500,000 reads were extracted from the ND dataset, and the K-20 model was used to identify 5mC. First, CpG sites with a methylation probability ≥ 0.9 or ≤ 0.1 were selected from the K-20 model as positive and negative samples, respectively. Secondly, for the remaining CpG sites in these 500,000 reads, those CpG sites that were aligned to the reference CpG sites and had a methylation frequency ≥ 95% or ≤ 5% (determined by BS-seq) were selected as positive and negative samples, respectively. For the processed data, 60,000 reads were extracted from the FMD and UMD datasets as positive and negative samples, respectively, which constituted the third part of the training dataset. In addition, CpG features of 10,000, 5,000, and 5,000 reads were extracted from the ND, FMD, and UMD datasets, respectively, and used as test datasets.

[0041] S5. Perform correlation analysis of single-molecule CpG methylation.

[0042] 100,000 reads were randomly extracted from a flowcell of native human data (ND) and a flowcell of native zebrafish data, respectively. The K-20 model was applied to these extracted reads to determine only the methylation status of high-confidence CpGs in the reads, that is, CpGs with a methylation probability ≤ 0.1 were considered negative, and CpGs with a methylation probability ≥ 0.9 were considered positive. The correlation of methylation between high-confidence CpGs at different distances was calculated, similar to the r in linkage disequilibrium analysis. 2 Here, the methylated and unmethylated states are treated as two different alleles. Then the r between CpGs at a given distance is calculated using the following formula: 2 value: ; Among them, F MM is the frequency of methylation of two CpGs at a given distance, F M is the frequency of methylation of a single CpG within a given distance of CpG. Chi-square tests were performed to assess statistical significance and corrected for multiple testing using the Benjamini-Hochberg method.

[0043] S6. Methylation frequency quantification module.

[0044] This embodiment includes two methylation frequency modes: read counting and site-frequency model modes.

[0045] By default, only reads with alignment quality (MAPQ) ≥ 10 are considered. In read counting mode, the number of reads above a positive probability threshold (default is 0.5) at a CpG site is recorded as the number of methylated reads (Nm), and the number of reads below a negative probability threshold (default is 0.5) is recorded as the number of unmethylated reads (Nu). Then, the methylation frequency of CpG is calculated as Nm / (Nm + Nu).

[0046] The site frequency model used the ND dataset with approximately 13.5× coverage and the K-400 count pattern methylation frequency data on chromosomes 15 to 20 as the training set. The methylation frequencies on chromosomes 21 and 22 were used as the test set.

[0047] S7. Use the trained single-molecule CpG methylation recognition module based on graph convolutional deep neural network to perform single-molecule 5mC methylation detection.

[0048] In another aspect of this embodiment, the following tests are performed on the method in the above embodiment (hifimeth) and the methods in the prior art (two methods based on 21-mer: primrose and ccsmeth) to enable those skilled in the art to more intuitively understand the advantages and beneficial effects of the present application.

[0049] (1) Correlation of single-molecule CpG methylation status.

[0050] The method of this example is used below to analyze the correlation between the methylation status of single molecule (CCS reads) CpGs at different distances in humans and zebrafish.

[0051] First, a 20-mer model was trained using CCS reads (F1 = 0.866) from processed NA12898 DNAs as training and test data. By applying strict thresholds for methylated (≥0.9) and unmethylated (≤0.1) CpG identification, the 20-mer model identified the methylation status of approximately 52.1% of CpG sites with a positive identification accuracy of 97.6% and a negative identification accuracy of 97.3% (e.g. Figure 7 Then, Pacbio Sequel II CCS reads from two flowcells, one from a human cell line (HG002) and the other from zebrafish (Danio rerio), were used to identify the methylation status of CpG sites using the above thresholds.

[0052] Correlation analysis showed that in humans and zebrafish, the methylation status of adjacent CpG sites on a single molecule was significantly positively correlated (e.g. Figure 2 Within the distance of 10 bp and 200 bp, the methylation status of CpG sites had a significant (FDR < 0.05) correlation (r 2 ), which are not less than 0.869 (human), 0.771 (zebrafish) and 0.431 (human), 0.318 (zebrafish), respectively. Like the correlation between methylation frequencies, the correlation between single-molecule methylation states also decays with increasing distance, and the rate of decline slows down significantly with increasing distance. CpG methylation correlation (r 2 ) decays faster than in the HG002 data. At distances of ~30 bp and ~15 bp, r 2It drops rapidly from 1 to 0.50 (chi-square test FDR < 0.05), and drops to 0.10 (FDR < 0.05) at a distance of ~1500 bp (human) and ~660 bp (zebrafish). The method of this application can demonstrate that the methylation status of single-molecule adjacent CpG sites in humans and zebrafish is significantly correlated, which provides a theoretical basis for using longer k-mer features to improve CpG methylation status recognition at the single-molecule level.

[0053] (2) Determine the optimal background length for 5mC detection.

[0054] In this example, different k-mers were used to train the same native training data set, and the k-mers ranged from 20 bp to 440 bp, with an increase of 20 bp per step.

[0055] Comparison of real and simulated random (Poisson distribution) CpG motif distribution in the human genome Figure 3 The cumulative ratio of the distance between a CpG and its first upstream or downstream CpG in the human genome and the probability of observing at least one additional CpG (calculated based on the observed cumulative ratio) at different background lengths [(distance + 1) × 2] are shown in Figure 2. Figure 3 As shown in C.

[0056] The results show that the AUC (area under the ROC curve) score of the model increases with the increase of k-mer length (e.g. Figure 3 As shown in D in Figure , all models were trained and tested using the same dataset, with three replicates), confirming that considering the background length has a significant impact on CpG methylation detection. The rate of increase in the AUC score slows down with increasing k-mer length, approaching a maximum value at approximately 400 bp. Considering the limited benefit of continuing to increase the background length and the increased computational burden, 400 bp was selected as the optimal background length, and the K-400 model was trained. To confirm that the improvement in performance is due to considering more neighboring CpGs, the AUC of the k-400 model for CpG sites with different numbers (0 to 49) of background CpGs was evaluated here. Figure 3 As shown in the figure, when tested on 6 flowcells of the native HG002 CCS dataset, the AUC of the K-400 model increased from a minimum of 0.904 ± 0.012 (0 background CpG) to a maximum of 0.996 ± 0.001 (>= 21 background CpG). The effect of background CpG approaches saturation at about 21. The rate of increase in AUC caused by background CpG decreases rapidly from the first (AUC increase of 0.015 ± 0.002) to the 21st (AUC increase < 0.001) (as shown in the figure). Figure 3The left panel shows the AUC of the 400 bp hifimeth model under different numbers of background CpGs, and the right panel shows the increase in AUC for each additional background CpG number. In humans, the probability of containing at least one additional background CpG site in a 400-mer environment is 98.4% (e.g. Figure 7 As shown, A represents the performance of the hifimeth K-20 model under different probability thresholds, and B represents the accuracy and ratio of high-confidence and low-confidence CpGs. Different methods use the same threshold: ≤0.1 is high-confidence negative, ≥0.9 is high-confidence positive; >0.1 and <0.9 are low-confidence), so the K-400 model can theoretically assist in the methylation identification of most CpG sites.

[0057] (3) Evaluate the performance of hifimeth in detecting 5mC at the single-molecule level.

[0058] In this example, hifimeth was compared with two 21-mer-based methods, primrose and ccsmeth, at the read level using native CCS datasets. The test datasets included human CCS reads from 9 flowcells and zebrafish CCS reads from 1 flowcell. The test used genome-wide fully methylated and unmethylated CpG sites in BS-seq (human HG002 cell line and zebrafish) or nanopore sequencing (human cell line CHM13) data as reference positive and negative CpG sites, respectively. For the processed datasets (NA12898, HG002, M2&W2), the CpGs in the sampled 10,000 FMD reads and 10,000 UMD reads were marked as true positives and true negatives, respectively. SAMtools v1.17 software was used for each processed dataset, and the "shuf" command in bash was used to sample the test dataset three times. A methylation probability threshold (default 0.5) was used to distinguish positive and negative calls in three softwares: hifimeth, ccsmeth v0.3.2, and primrose v1.3.0. Methylation status was classified into true positive (TP), true negative (TN), false positive (FP), and false negative (FN) according to the CpG sites marked in the CCS reads. Then, the AUC score, precision, accuracy, recall, and F1 score in hifimeth, ccsmeth, and primrose were calculated for each test dataset (native or processed) using Python scripts.

[0059] Native datasets: On all test datasets, hifimeth outperforms ccsmeth and primrose in terms of precision, accuracy, recall, and F1 score (e.g. Figure 4 As shown in A, the evaluation results of zebrafish data with 1 flowcell and human data with 9 flowcells, including 6 flowcells of HG002 cell line (2 libraries of 15 kb, 20 kb and 24 kb each) and 3 flowcells of CHM13 cell line, the AUC value (area under the ROCcurve) of hifimeth (0.987 ± 0.004) is higher and more stable (lower variance) than primrose (0.913 ± 0.009) and ccsmeth (0.961 ± 0.010) (as shown in Figure 2A). Figure 4 B in (for CHM13, the figure shows the average of three flowcells) and Table 1). On human data, the average accuracy of hifimeth is 94.6% ± 0.8%, and the average F1 score is 94.2% ± 0.6%, which is higher than primrose (accuracy 83.3% ± 1.0%, F1 score 81.3% ± 1.1%) and ccsmeth (accuracy 89.5% ± 1.1%, F1 score 87.9% ± 0.9%); although the K-400 model was only trained on human data, on the zebrafish dataset, the accuracy of hifimeth is 95.8%, and the F1 score is 97.5%, which is higher than primrose (accuracy 82.7%) and ccsmeth (accuracy 86.5%), and the improvement in accuracy is more significant than that of human data. Figure 4 D in Figure 3 shows a comparison of three treated data sets, including CCS reads of fully methylated DNA treated with methyltransferase (as a positive control) and CCS reads of fully unmethylated DNA amplified by PCR (as a negative control).

[0060] ; Processed Datasets: The strength of hifimeth was also evaluated at the read level using three CCS datasets of processed DNA from different human cell lines, including HG002 (partial training data for ccsmeth and hifimeth), NA12898 (training data for primrose), and M2&W2 data from the study by Tse et al. These processed DNAs consist of fully methylated reads and fully unmethylated reads, which were treated with M.SssI methyltransferase and PCR amplification, respectively. Despite being trained primarily on native CCS reads, the K-400 model for hifimeth outperformed primrose and ccsmeth, which used a larger proportion of processed data in training, on all tested datasets (e.g. Figure 4 D in and Table 1).

[0061] In summary, compared with primrose and ccsmeth, the performance of the K-400 model of hifimeth at the single-molecule level has been significantly improved, which is better than the short k-mer model and has advantages in different types of CCS data.

[0062] (4) High accuracy of hifimeth at the single-molecule level.

[0063] An important goal of methylation detection is to quantify the methylation frequency of the entire genome. Sub-accuracy and recall can theoretically produce better CpG site methylation frequency quantification results. Here, subsets of CCS reads with sequencing depths of 10×, 15×, 20×, and 25× were randomly sampled from 4 types of libraries, including HG00215 kb, 20kb (only for 10× and 15×), 24 kb, and CHM13 20kb libraries. The sampling method was: evaluate their sequencing depth by using the 'samtools depth' command on the aligned BAM file of each flowcell, then calculate the proportion of reads that should be sampled for each expected sequencing depth, and use the 'samtools view -s' command to sample the same number of bases for the flowcell of each type of library. Using these downsampled datasets, for hifimeth and ccsmeth, the whole genome methylation frequency profiles were calculated using their frequency quantification modules; for primrose, pb-CpG-toolsv2.2.0 was used for calculation. On each test dataset, the read count mode and frequency model mode were applied to each software. Then, this example calculated the Pearson correlation coefficient between the methylation frequency spectra obtained by hifimeth, ccsmeth and primrose and the corresponding BS-seq and nanopore sequencing results.

[0064] Performance of read count mode: The performance of primrose, ccsmeth, and hifimeth in quantifying the methylation frequency of CpG sites in the whole genome by read count mode was compared. The quality of the results was evaluated by correlation analysis with the BS-seq results of HG002 at about 100× or the nanopore sequencing data results of CHM13 at about 33×. The results showed that hifimeth had the highest correlation coefficient on all tested datasets, whether compared with BS-seq or nanopore sequencing data, followed by ccsmeth and primrose (e.g. Figure 5 In addition, hifimeth has greater advantages than ccsmeth and primrose in datasets with lower sequencing depth (e.g. Figure 5Figure A shows the correlation coefficients of methylation frequency quantification by hifimeth, ccsmeth, and primrose (y-axis), with nanopore sequencing data from the human CCS dataset at different sequencing depths (x-axis) as the reference quantification results. The CCS data were randomly downsampled from three HG002 libraries with different insert sizes (15 kb, 20 kb, and 24 kb) and one CHM13 library (20 kb). For data with sequencing depth <25×, the average of three replicates is shown, and the results of single tests are shown in Table 2). At 15× CCS reads, hifimeth had a correlation of ≥0.90 with the results of BS-seq (HG002) or ONT (CHM13) in most (3 / 4) of the libraries tested. In the 24 kb library of HG002 and the 20 kb library of CHM13, the performance of hifimeth with 10× or 15× reads exceeded the performance of ccsmeth and primrose with 25× reads. In addition, their genome-wide site frequency quantification results on zebrafish CCS data of about 23.3× were compared. The correlation coefficients of the genome-wide methylation frequencies obtained by the six methods on zebrafish CCS data flowcell and BS-seq data were improved by 0.075 and 0.110 respectively compared with ccsmeth and primrose (e.g. Figure 5 As shown in B in ).

[0065] Performance of site frequency model mode: Both ccsmeth and primrose contain site frequency models that take into account the methylation frequencies of upstream and downstream CpG sites in order to improve the correlation with BS-seq data. The performance of the read count mode of hifimeth on the HG00224kb library and zebrafish data is similar to that of the site frequency model of ccsmeth. This example uses the method used by ccsmeth to train a site frequency model for hifimeth, which slightly improves the correlation coefficient of hifimeth with BS-seq or nanopore sequencing data on all test data sets (such as Figure 5 The frequency model patterns of hifimeth are better than those of primrose and equal to those of ccsmeth on all tested datasets. Specifically, the frequency model patterns of hifimeth are better than those of ccsmeth on two datasets (HG002 24kb library and zebrafish data), similar to those of ccsmeth on HG002 15kb library, and worse than those of ccsmeth on HG002 20kb library and CHM13 library (see Figure 2A). Figure 5 ). Figure 5Figure C shows a heat map and correlation analysis comparing the methylation frequencies of the genotypes (paternal and maternal haplotypes) of HG002CCS-hifimeth and Nanopore data.

[0066] Methylation frequency quantification at the haplotype level: Given a bam file with typing variation or haplotype markers, hifimeth can achieve quantification of site methylation frequency at the haplotype level. The performance of hifimeth at the haplotype level was evaluated here using the CCS reads of HG002, and the HG002 CCS data of six flowcells were merged and downsampled to obtain five different sequencing depths (10×-60×), and the sampling method was as described above. This example also obtained the whole genome typing variation (VCF file) from Jarvis et al. studying the data. For comparison, this example obtains a comparison BAM file of about 40× nanopore HG002 data from the OxfordNanopore Open Data project. This example uses whathap v1.7 (39) to assign haplotype information (HP tags) to reads in the aligned BAM files of the hifimeth and nanopore data using the genotyped VCF files, genotyped methylation frequency profiles for the paternal (HP:i:1) and maternal (HP:i:2) haplotypes using the read count mode and site frequency model mode of hifimeth, and genotyped methylation frequency profiles on the haplotype-tagged nanopore BAM files using modbam2bed. Finally, haplotype methylation correlation analysis was performed between the CCS-hifimeth and nanopore results using CpGs with at least 5 reads per haplotype in both datasets. The results showed that from 10× to 74× sequencing depth, the correlation coefficient between the methylation frequency using the hifimeth counting mode and the methylation frequency of the approximately 40× nanopore data increased from 0.864 (10×) to 0.924 (74×) for paternal or maternal haplotypes, and from 0.893 to 0.938 using the frequency model mode (e.g. Figure 5 C and Figure 8 The number of CpG sites that can be used for quantification (≥5 reads per haplotype) also increased from 7.05 million to 24.8 million (as shown in Table 2). At a sequencing depth of about 30×, the correlation coefficient reached 0.903 (counting mode), and the number of CpG sites that can be used for quantification exceeded 95% (23.6 million) at a sequencing depth of about 74× (as shown in Table 2). Figure 8and Table 2). In addition, differentially methylated regions (DMRs) between the two haplotypes were detected. The frequency quantification results of CCS-hifimeth (counting mode) and nanopore were analyzed using DSS v2.48.0, with thresholds set at p < 0.001 and allele methylation difference > 0.2, and then the DMRs identified by the two methods were intersected using bedtools v2.30.0. According to the haplotype with higher methylation frequency, DMRs were further classified as maternal hypermethylated regions (-M) or paternal hypermethylated regions (-P). In this example, the reference genome was divided into 1 MB continuous non-overlapping windows to analyze the distribution of these regions in the genome. Figure 6 A in the figure shows the distribution of allelic differential methylation regions (DMRs) and read methylation clustering regions in the whole genome. The proportion of maternal or paternal hypermethylated regions shown is calculated by dividing the total length of maternal (or paternal) hypermethylated DMRs in each 1 MB window by 1,000,000. The outermost ring shows the 24 chromosomes of GRCh38 with cytogenetic bands from the UCSC genome browser. The next four histograms show the proportion of maternal (-M) or paternal (-P) hypermethylated regions detected by nanopore sequencing (ONT) and CCS-hifimeth in 1 Mb windows (vertical axis ranges from 0 to 0.01). The innermost two circles show the methylation clustering regions (≥400 bp) of the reads of the HG002 (~74×) and CHM13 (~25×) CCS datasets. The results show that based on the methylation identification results of hifimeth, this example detected 15,908 differentially methylated regions (DMRs) with allelic differences ≥20%, of which 65.6% overlapped with the allelic DMRs of the nanopore data. The distribution of genomic hypermethylated paternal and maternal regions was significantly correlated between CCS-hifimeth and nanopore data (p<0.001, r =0.855 and 0.874). These allelic DMRs also overlapped with 79.7% (114) of the 143 imprinted DMRs reported by Akbari et al.

[0067] Specifically, Figure 8 A in Figure 2 shows the correlation of methylation frequencies between the CCS dataset and the Nanopore data at different sequencing depths identified using HiFiMeth. Three replicates were sampled at depths ranging from 10 to 60×, and the average correlation is shown. Figure 8B in the figure indicates the proportion of CpGs that can be quantified for both haplotypes at different CCS sequencing depths. A CpG is considered quantifiable only when the number of reads for any haplotype is ≥5. The calculated ratio is relative to the number of CpGs that can be quantified at a depth of ~74×. The correlation coefficients and the number of CpGs that can be quantified for each sampling data are shown in Table 2.

[0068] ; Like ccsmeth, hifimeth identifies the methylation status of all CpGs based on kinetic information on CCS reads and has advantages over NGS-based methods in terms of genome coverage, especially in repetitive regions. Overall, due to its improved performance at the single-molecule level, hifimeth is able to achieve high-quality genome-wide methylation frequency quantification at relatively low CCS sequencing depth, outperforming the short k-mer model.

[0069] (5) Revealing complex single-molecule CpG methylation patterns.

[0070] Highly accurate single-molecule methylation detection can explore diverse methylation patterns in homologous sequences. Methylation-based clustering can separate parental allele reads in imprinted genomic regions without using typing genetic variation. This example attempts to distinguish paternal and maternal reads in genomic regions with ≥40% allele methylation based only on the methylation recognition results of hifimeth.

[0071] A similar approach to that of Gershman et al. was used here with some modifications. The steps were: First, the reference genome was divided into 200 bp continuous non-overlapping bins, and those bins with high CpG density (≥5 CpGs per bin) were selected for testing. The average methylation frequency (Mbin) of each bin was calculated by dividing the total number of methylated CpGs in all bins that overlapped with CCS reads (MAPQ ≥ 10) by the total number of CpGs. If a read covered at least 5 CpGs, the methylation level of the read (Rbin) was calculated by dividing the number of methylated CpGs in the read by the total number of CpGs. If the number of reads with Rbin ≥ Mbin + 0.2 or Rbin ≤ Mbin - 0.2 accounts for ≥80% of all reads covering a bin (≥5 CpGs), the reads with Rbin ≥ Mbin + 0.2 are assigned to the high methylation group (HMG), and the reads with Rbin ≤ Mbin -0.2 are assigned to the low methylation group (LMG). Fisher's exact test was further used to compare the number of methylated and unmethylated CpGs in HMG and LMG reads, and bins with p<0.001 were regarded as methylation clustering bins. When assigning HMG or LMG to reads, weighting was performed according to the number of CpGs contained in the clustering bins, and reads containing more HMG clustering bins were assigned to HMG, and vice versa. Then this example further used bedtools v2.30.0 to merge adjacent bins with a distance of no more than 200bp to screen clustered regions (≥400bp).

[0072] In the HG002 CCS dataset (approximately 74×), a total of 9738 cluster bins and 1259 cluster regions were identified. Among them, 6822 (70.0%) cluster bins and 1179 (93.6%) cluster regions overlapped with allelic DMRs (e.g. Figure 6 ), indicating that methylation-based clustering effectively captures allele differences in methylation patterns. In addition, at least one clustering bin was found in the promoter regions of 727 protein-coding genes, identifying potential imprinted genes, including 24 previously reported genes. For example, CYREN, which has not been recorded as an imprinted gene before, contains two clustering bins in its promoter and gene regions in HG002, with paternal reads mainly assigned to the high methylation group (HMG) and maternal reads assigned to the low methylation group (LMG) (as shown in Figure 2). Fig. 9 Similar hyper- and hypomethylated groups of reads were also observed for CYREN in CHM13.

[0073] Methylation-based read clustering has the theoretical potential to distinguish reads in low allelic variation or homozygous genomic regions. In HG002, 42 of the 80 clustered regions that did not overlap with any allelic differentially methylated regions (DMRs) had reads with fewer than 5 haplotype markers in at least one haplotype due to low local allelic variation. CHM13, a cell line with an almost homozygous genome and a female karyotype, still showed the presence of genomic imprinting effects. Using hifimeth on approximately 25× of CHM13 CCS reads, a total of 3,006 clustered regions were identified, of which 1,376 were located on the heavily imprinted X chromosome with approximately 15.6-fold enrichment compared to autosomes (Fisher's exact test, p<0.001). In the DXZ4 tandem repeat region of CHM13 chrX (e.g., Figure 6 In B), the low-methylated group (LMG) and high-methylated group (HMG) reads successfully captured the differential methylation patterns of active and inactive X chromosomes, similar to the observations using nanopore data at approximately 33×. These results highlight the ability of hifimeth to effectively distinguish allelic methylation patterns in regions of low allelic variation levels.

[0074] By using CCS-hifimeth, methylation-based read clustering was also able to reveal differential methylation patterns that were not associated with allelic differences. Of the remaining 38 clustered regions in HG002, 14 were located on chromosome X (e.g. Figure 6 As shown in A in Figure 3 ). Since HG002 is a male cell line, these 14 clustered regions cannot be attributed to differences between alleles. In addition, 5 of these regions are likely to be repetitive sequences with a sequencing depth of more than 1.5× in the genome. Of the remaining 24 regions on the autosomes, 7 had an average sequencing depth of more than 1.5×. After considering and excluding other factors, the methylation-based read clustering in the remaining 17 regions may represent differences at the cellular level.

[0075] In addition, the CpG methylation recognition of hifimeth can effectively capture the differential methylation patterns in tandem repeat regions. For example, in the DXZ4 region of CHM13, this example observed different methylation patterns in approximately 3kb repeat units in the reads of the high methylation group (HMG) and the low methylation group (LMG). Figure 6As shown in B, each arrow at the top represents a repeat unit with a length of ~3kb. The top panel shows the methylation frequency (points) of the low-methylation group and the high-methylation group at each CpG, as well as the average methylation frequency (curve) smoothed by modbamtools (Razaghi et al., 2022). The single-molecule graph at the bottom shows the CpG methylation status of each read segment in the low-methylation group and the high-methylation group, which are inferred to be the inactive X chromosome (Xi) and the active X chromosome (Xa), respectively.

[0076] High-precision hifimeth methylation calls provide a method to reveal complex single-molecule methylation patterns at high resolution, including changes in methylation status within reads. Although reads overlapping clustered regions are assigned to high methylation groups (HMG) or low methylation groups (LMG) based on the main cluster bins (weighted by the number of CpGs), it should be noted that a read can show different clustering patterns in different cluster bins. In HG002 and CHM13, this example observed at least two reads with such methylation state switches in 1,010 and 4,623 different genomic regions, respectively. In the DXZ4 region of CHM13, this example identified multiple clustered regions where HMG reads showed low methylation levels and LMG reads showed high methylation levels. Figure 6 The three example regions marked in B in Figure 2 demonstrate this phenomenon. In addition, in the CYREN gene of HG002, this example observed an allelic differentially methylated CpG (chr7:135,168,118-135,168,119, p = 1.1E-4), whose methylation level was opposite to the distribution of HMG and LMG (as shown in Figure 2). Fig. 9 In addition, taking the HG002 CYREN gene as an example, a highly significant allelic differential methylation CpG was observed at one of its splice sites (chr7:135,169,124-135,169,125) (by Fisher's exact test, p = 2.1E-11). Interestingly, the CpGs upstream and downstream did not show significant allelic methylation level differences (as shown in Table 2). Fig. 9 ). In the CHM13 data, this isolated CpG also exhibits different methylation patterns for HMG and LMG reads. Collectively, these findings suggest that CpG methylation regulation in natural DNA can be highly complex and that hifimeth has the power to reveal methylation patterns at the resolution of a single CpG motif.

[0077] Specifically, Fig. 9Represents the methylation patterns of the CYREN gene in HG002. From top to bottom: GenCode gene structure annotation, CYREN is located on the reverse strand of the GRCh38 reference genome; Methylation frequency panel: Each point represents the methylation frequency of the paternal or maternal parent on the corresponding CpG, and the two curves represent the smoothed average methylation frequency calculated using modbamtools (Razaghi et al., 2022); The histogram shows the -log10 (p value) of the Fisher exact test for the distribution of methylated and unmethylated calls on a single CpG between the two haplotypes; The bottom panel shows the single-molecule methylation patterns of the reads of the paternal and maternal haplotypes, as well as the unmethylated and methylated CpGs. In the bottom panel, the arrows indicate two CpGs with significantly different methylation states compared to the surrounding CpGs. The box indicates two clustering bins in the region detected by methylation-based read clustering.

[0078] In another embodiment of the present application, a system is provided that can improve the accuracy of single-molecule 5mC methylation detection based on CCS, the system comprising a data acquisition module, a model building module, a background base length selection module, and a methylation detection module; The data acquisition module is used to acquire the original data set and the processed data set, and construct the training data set and the test data set; The model building module is used to build a single-molecule CpG methylation recognition module based on a graph convolutional deep neural network, and is trained using training data of different background base lengths; The background base length selection module is used to test the background base length and model performance, select a suitable base length and train the corresponding model; The methylation detection module is used to perform single-molecule 5mC methylation detection using a trained single-molecule CpG methylation recognition module based on a graph convolutional deep neural network.

[0079] It should be noted here that the system provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure is divided into different functional modules to complete all or part of the functions described above. This system is a method applied to the above embodiment to improve the accuracy of single-molecule 5mC methylation detection based on CCS.

[0080] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0081] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A method for improving the accuracy of single-molecule 5mC methylation detection based on CCS, characterized in that: The steps include: Obtain native data sets and processed data sets, and construct training data sets and test data sets; Construct a single-molecule CpG methylation recognition module based on graph convolutional deep neural network and train it using training data of different background base lengths; Test the background base length and model performance, select the appropriate base length and train the corresponding model; Single-molecule 5mC methylation detection was performed using the trained single-molecule CpG methylation recognition module based on graph convolutional deep neural network.

2. A method for improving the accuracy of single-molecule 5mC methylation detection based on CCS according to claim 1, characterized in that: The acquisition of the original data set and the processed data set, and the construction of the training data set and the test data set are specifically as follows: The native data set is obtained by downsampling a DNA sequencing vector chip containing CCS reads of the native human HG002 genome reference sample; The processed data set includes fully methylated CCS data obtained by treating DNA with CpG methyltransferase M.SssI and unmethylated data obtained from PCR amplified DNA, which are used for training the read count model; Construct training datasets using native datasets and processed datasets; The test data set includes a training data set, and also includes a native human HG002 genome reference sample data set of multiple DNA sequencing carrier chips, a native CHM13 data set, a native zebrafish data, and processed native human HG002 genome reference sample data.

3. A method for improving the accuracy of single-molecule 5mC methylation detection based on CCS according to claim 1, characterized in that: The single-molecule CpG methylation recognition module based on graph convolutional deep neural network includes 12 blocks connected in sequence, wherein the first 11 blocks each include a batch normalization layer, a graph convolution layer and a rectified linear unit activation function, and the last block includes a batch normalization layer, a graph convolution layer and a global average pooling layer, and also includes a feedforward layer.

4. A method for improving the accuracy of single-molecule 5mC methylation detection based on CCS according to claim 3, characterized in that: The input and output of the CpG methylation recognition module based on graph convolutional deep neural network are specifically: Given a read segment R = r0r1...r n-1 , the sequence consists of the alphabet Σ = (A, C, G, T), and given a CpG site p, with 0 ≤ p < n, extract a background base Kp = r of length k p-h r p-h+1 ...r p+h-1 , where h = k / 2; When p < h or n - p < h, extract to the first base r0 or the last base r0 of the sequence n-1 ; Construct a graph Gp as input for Kp; In the graph Gp, each base r in Kp i Both represent a node Nᵢ, and the two adjacent bases rᵢ and rᵢ +1 Through a directed edge eᵢ = r i ← r i+1 The graph Gp contains k nodes and k - 1 directed edges. The edge eᵢ in Gp does not contain any features. Each node Nᵢ has 15 features fᵢ ∈ ℝ¹ 5 ,These features include base information, sequencing quality, pulse interval time and pulse width of the positive chain, pulse interval time and pulse width of the negative chain, which are summarized in a two-dimensional tensor with shape k × 15; The output contains two probability values ​​p0 and p1, and p0 + p1 = 1, which respectively represent the probabilities of the methylated and unmethylated states of the CpG site.

5. A method for improving the accuracy of single-molecule 5mC methylation detection based on CCS according to claim 1, characterized in that: The training process of the single-molecule CpG methylation recognition module based on graph convolutional deep neural network is as follows: The training was performed using the cross entropy loss function and the SGD optimizer with a learning rate of 0.1 and a weight decay of 10. -5 , momentum and Nesterov acceleration gradient are 0.9; The StepLR learning rate scheduler was used to decay the learning rate by 0.1 after each training round; each single-molecule CpG methylation recognition module based on graph convolutional deep neural network was trained for two training rounds, or a pre-trained model was selected as the starting point for training.

6. A method for improving the accuracy of single-molecule 5mC methylation detection based on CCS according to claim 1, characterized in that: The background base length and model performance test are specifically: The data in the CCS reads of bisulfite sequencing performed on the native human HG002 genome reference sample in the native dataset were used to identify fully methylated and completely unmethylated CpG sites, and positive and negative references were constructed, respectively. Several data were randomly sampled from the native dataset to evaluate the model multiple times, and the randomly sampled data did not include the data used for training.

7. A method for improving the accuracy of single-molecule 5mC methylation detection based on CCS according to claim 1, characterized in that: The specific steps of selecting a suitable base length and training a corresponding model are as follows: The K-20 model, i.e., the model with 20-mer background base length, was selected, and positive and negative CpG features were extracted from the dataset for training and testing respectively; The K-20 model was used as a pre-training model for training the K-400 model, where the K-400 model refers to a model with a background base length of 400-mer, and the native dataset was supplemented with the processed dataset for training and testing, where the training datasets included: using the K-20 model to select CpG sites in the dataset with a methylation probability ≥ 0.9 or ≤ 0.1, as positive and negative samples respectively, and for the remaining CpG sites, CpG sites that were aligned to the reference CpG sites and had a methylation frequency ≥ 95% or ≤ 5% were selected as positive and negative samples respectively, where the methylation frequency was determined by BS-seq.

8. A method for improving the accuracy of single-molecule 5mC methylation detection based on CCS according to claim 1, characterized in that: Single-molecule CpG methylation correlation analysis was performed, specifically: A number of data were extracted from a DNA sequencing vector chip containing native human data and a DNA sequencing vector chip containing native zebrafish data, and the K-20 model, i.e., a model with a 20-mer background base length, was used to determine the methylation status of high-confidence CpGs, i.e., CpGs with a methylation probability ≤ 0.1 were considered negative, and CpGs with a methylation probability ≥ 0.9 were considered positive; Treating the methylated and unmethylated states as two different alleles, the correlation of methylation between high-confidence CpGs at different distances is calculated, specifically: ； Among them, F MM is the frequency of methylation of two CpGs at a given distance, F M is the frequency of methylation of a single CpG in a CpG at a given distance; Chi-square tests were performed to assess statistical significance and corrected for multiple testing using the Benjamini-Hochberg test.

9. A method for improving the accuracy of single-molecule 5mC methylation detection based on CCS according to claim 1, characterized in that: The methylation frequency quantification module of hifimeth includes a read segment counting mode; The read counting mode only considers read data with alignment quality MAPQ ≥ 10, specifically: the number of reads exceeding the positive probability threshold at a CpG site is recorded as the methylated read number Nm, and the number of reads below the negative probability threshold is recorded as the unmethylated read number Nu, and the methylation frequency of CpG is calculated using Nm / (Nm + Nu).

10. A system capable of improving the accuracy of single-molecule 5mC methylation detection based on CCS, characterized in that: A method for improving the accuracy of single-molecule 5mC methylation detection based on CCS, applied to any one of claims 1-9, comprising a data acquisition module, a model building module, a background base length selection module and a methylation detection module; The data acquisition module is used to acquire the original data set and the processed data set, and construct the training data set and the test data set; The model building module is used to build a single-molecule CpG methylation recognition module based on a graph convolutional deep neural network, and is trained using training data of different background base lengths; The background base length selection module is used to test the background base length and model performance, select a suitable base length and train the corresponding model; The methylation detection module is used to perform single-molecule 5mC methylation detection using a trained single-molecule CpG methylation recognition module based on a graph convolutional deep neural network.

Images