A method and system for constructing a model for identifying glycosylated RNA modifications based on nanopore sequencing

Abstract

The application provides a glycosylated RNA modification identification model construction method and system based on nanopore sequencing, and belongs to the field of RNA modification identification technology in new generation information technology. A high-confidence training set containing base feature sequences, current signal feature sequences and base modification types is established by fusing second-generation / third-generation sequencing data from a variety of public databases; a RNA modification identification model based on cross-modal feature fusion and multi-scale sequence modeling is pre-trained by using a sample subset with a larger modification type scale in the training set; and the pre-trained RNA modification identification model is fine-tuned by using a sample subset with a smaller modification type scale, so that a modification identification model suitable for glycosylated RNA is obtained. The modification identification model has good adaptability in modification feature detection of different cell lines, significantly improves the learning efficiency and prediction accuracy of rare modifications, and especially supports modification type identification of a small amount of glycosylated RNA samples.

Description

Technical Field

[0001] This invention belongs to the field of RNA modification identification technology in the new generation of information technology, and particularly relates to a method and system for constructing a glycosylated RNA modification identification model based on nanopore sequencing. Background Technology

[0002] RNA modification is a crucial post-transcriptional regulatory mechanism, widely present in both coding and non-coding RNAs. It influences key processes such as RNA splicing, nuclear export, stability, and translation efficiency, playing a vital role in cell function, developmental processes, and disease progression. Since its initial discovery in the 1950s, over 170 types of chemical modifications have been identified, including N6-methyladenosine (M... 6 A) Pseudorazine (Ψ), 5-methylcytosine (m) 5 C) and adenosine-inosine (A-to-I) RNA editing, etc. These modifications are dynamic and reversible, responding sensitively to changes in the extracellular environment and affecting physiological activities such as cell proliferation and differentiation by regulating protein synthesis. In recent years, abnormal RNA modifications have been found to be closely related to various pathological processes, such as neurodegenerative diseases, metabolic syndromes, and tumor metastasis.

[0003] In 2021, Bertozzi's team at Stanford University first revealed the existence of glycosylated RNA (glycoRNA), confirming that RNA is the third type of glycosylation carrier after proteins and lipids. GlycoRNAs are located on the cell membrane surface and may mediate novel signaling pathways, influencing intercellular communication of immune cells and neutrophil recruitment, providing new targets for tumor immunotherapy. However, compared with mature m... 6 Compared to A-modification detection systems, the identification of whole-genome modification sites for glycoRNAs is still in its early stages and lacks single-molecule dynamic detection capabilities. Nanopore sequencing, as a third-generation sequencing technology, has been successfully applied to m 6 A、m 5 While the detection of common RNA modifications such as C is possible, the current signal characteristics of glycoRNAs have not yet been systematically analyzed, and existing algorithms cannot achieve high-precision prediction.

[0004] In the prior art, patent document CN116343916A discloses a method for identifying RNA chemical modifications based on nanopore sequencing. This method establishes multiple RNA modification feature sets and constructs a deep learning model to achieve single-base resolution modification detection, exhibiting good generalization performance and applicability to the identification of RNA modifications in multiple species. However, this method primarily targets methylation modifications and does not address feature extraction and prediction of glycosylated RNA modifications. Patent document CN116525001A discloses a deep learning-based RNA modification site prediction model, employing a combination of CNN and RNN, and incorporating attention mechanisms and RNA sequence secondary structure information, significantly improving the model's classification accuracy and interpretability. However, this model does not integrate nanopore sequencing current signals with next-generation sequencing base information, failing to achieve cross-modal data integration and prediction of modification sites with limited samples.

[0005] Defects and shortcomings of existing technology:

[0006] Lack of characterization of glycosylated RNA modifications: (1) Existing technologies mainly target methylation modifications (such as m 6 A、m 5 (C) The current signal feature extraction and prediction of glycoRNA are not involved, which cannot meet the detection requirements of glycosylation modification; (2) The prediction accuracy of existing modification sites is limited, especially in the case of a small number of glycosylated RNA samples; (3) The long length of nanopore sequencing is not utilized: the existing technology does not utilize the long length of nanopore sequencing, which limits the comprehensive analysis and functional study of RNA modification sites. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a method and system for constructing a glycosylated RNA modification identification model based on nanopore sequencing.

[0008] The technical solution adopted in this invention is as follows:

[0009] In a first aspect, this invention proposes a method for constructing a glycosylated RNA modification identification model based on nanopore sequencing, comprising:

[0010] Raw second-generation sequencing and third-generation sequencing data of high-confidence modification sites from RNA samples of multiple species were obtained from various public database sources. The modification types of RNA from multiple species include m 6 A、m 5 C, m 1 A, Am, Um, acp 3 U、m 7 G, Cm, Gm;

[0011] Align the current signals in the third-generation sequencing data with the reference genome once;

[0012] The high-confidence modification sites in the second-generation sequencing data are aligned with the current signals in the third-generation sequencing data after the first alignment. The features of each base current signal are extracted from the aligned current signals to obtain a high-confidence training set containing base feature sequences, current signal feature sequences, and each base modification type.

[0013] An RNA modification identification model based on cross-modal feature fusion and multi-scale sequence modeling was pre-trained using a subset of samples with a large number of modification types in a high-confidence training set. Then, the pre-trained RNA modification identification model was fine-tuned using a subset of samples with a smaller number of modification types to obtain a modification identification model adapted to glycosylated RNA.

[0014] Furthermore, the database from which the original next-generation sequencing data originated includes at least two of RMDisease V2.0, RMVar V2.0, RMBase V3.0, and m7GHub V2.0.

[0015] Furthermore, the original third-generation sequencing data source databases include at least two of the GEO, SRA, and SG-Nex databases.

[0016] Furthermore, high-confidence modification sites refer to sites that have been verified by at least two different experimental techniques, and the position weight matrix PWMs of the site modification type is greater than 3, with the training set containing more than one million high-confidence modification sites.

[0017] Furthermore, the characteristics of each base current signal include the mean, median, standard deviation, and dwell time of the current signal.

[0018] Furthermore, the process of aligning the current signals in the third-generation sequencing data with the reference genome includes:

[0019] The Bonito basecalling model was retrained using human cell data;

[0020] The retrained Bonito basecalling model was used to perform basecall operations on the third-generation sequencing data to complete the data calibration and achieve a one-time alignment process.

[0021] Furthermore, the secondary alignment process is implemented using the Remora current signal reanalysis tool.

[0022] Furthermore, the RNA modification identification model first uses a cross-attention mechanism to fuse base features and current signal features, where base features are used as queries and current signal features are used as key-value pairs; then, the Evo 2 StripedHyena 2 architecture is used, combined with short-range explicit convolution (SE), mid-range regularized convolution (MR), and long-range implicit convolution (LI) to perform multi-scale sequence modeling on the fusion result, and predicts the type of each base modification based on the final modeling features.

[0023] Furthermore, the sample subset used for pre-training the RNA modification identification model must contain at least m modification types. 6 A、m 5 C; The sample subset used to fine-tune the RNA modification identification model must contain at least ACP modifications. 3 U.

[0024] Secondly, this invention proposes a nanopore sequencing-based glycosylated RNA modification identification model construction system to realize the above-mentioned nanopore sequencing-based glycosylated RNA modification identification model construction method.

[0025] The beneficial effects of this invention are:

[0026] This invention obtains high-confidence modified sites from RNA samples of multiple species from various public databases, including raw second-generation sequencing data and third-generation sequencing data. It then performs a first alignment of the current signals in the third-generation sequencing data with the reference genome, and a second alignment of the high-confidence modified sites in the second-generation sequencing data with the current signals in the first-aligned third-generation sequencing data. This ensures the high confidence of the training set and lays a data foundation for model training.

[0027] This invention employs a transfer learning strategy to pre-train an RNA modification identification model based on cross-modal feature fusion and multi-scale sequence modeling using a subset of samples with a large number of modification types in a high-confidence training set. Then, it fine-tunes the pre-trained RNA modification identification model using a subset of samples with a smaller number of modification types. This significantly improves the learning efficiency and prediction accuracy of rare modifications and enhances the prediction ability of modification types in a small number of glycosylated RNA samples. It is the first to establish an algorithm for glycoRNA modification, filling the gap in the field of glycoRNA modification detection using nanopore sequencing.

[0028] This invention employs a cross-attention mechanism to achieve cross-modal feature fusion, supporting multi-scale sequence modeling from single nucleotides to millions of bases. Compared with existing RNA modification identification models (such as TandemMod, Nanom6A, and RNA-FM), the modification identification model of this invention exhibits good adaptability and superior performance in detecting modification features in different cell lines, providing reliable support for functional studies and clinical translational applications of RNA modifications. Attached Figure Description

[0029] Figure 1 This is a schematic diagram illustrating the process of integrating second-generation sequencing data and third-generation sequencing data in this invention;

[0030] Figure 2 This is a schematic diagram of the framework of the RNA modification identification model used in this invention;

[0031] Figure 3 This is a flowchart of the method for constructing a glycosylated RNA modification identification model based on nanopore sequencing according to the present invention. Detailed Implementation

[0032] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0033] The accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0034] The flowchart shown in the attached diagram is merely an illustrative example and does not necessarily include all steps. For example, some steps may be broken down, while others may be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0035] This invention addresses the shortcomings of existing technologies, such as the lack of feature analysis for glycosylated RNA modifications and low prediction accuracy for traditional RNA modification types. It proposes a method and system for constructing a glycosylated RNA modification identification model based on nanopore sequencing. This invention integrates second-generation and third-generation sequencing data to construct a high-confidence training set and utilizes cross-modal feature fusion and transfer learning strategies to significantly improve the prediction ability for RNA modification types with limited samples. This invention employs an RNA modification identification model based on cross-modal feature fusion and multi-scale sequence modeling, fully leveraging the long read length characteristics of nanopore sequencing to achieve high-precision prediction of multiple modification types.

[0036] like Figure 3 As shown, a method for constructing a glycosylated RNA modification identification model based on nanopore sequencing includes:

[0037] S1. Raw second-generation sequencing and third-generation sequencing data of high-confidence modification sites of RNA samples from multiple species were obtained from various public database sources. The modification types of RNA from multiple species include m 6 A、m 5 C, m 1 A, Am, Um, acp 3 U、m 7 G, Cm, Gm;

[0038] S2 aligns the current signal in the third-generation sequencing data with the reference genome;

[0039] S3, the high-confidence modification sites in the second-generation sequencing data are aligned with the current signals in the third-generation sequencing data after the first alignment. The features of each base current signal are extracted from the aligned current signals to obtain a high-confidence training set containing base feature sequences, current signal feature sequences, and each base modification type.

[0040] S4. A sample subset with a large number of modification types in the high-confidence training set is used to pre-train an RNA modification identification model based on cross-modal feature fusion and multi-scale sequence modeling. Then, the pre-trained RNA modification identification model is fine-tuned using a sample subset with a smaller number of modification types to obtain a modification identification model adapted to glycosylated RNA.

[0041] In one specific embodiment of the present invention, step S1 is the basis for integrating second-generation sequencing data and third-generation sequencing data. First, raw data on high-confidence modification sites of RNA samples from multiple species are obtained from various public database sources. For example... Figure 1 As shown, the databases from which the original next-generation sequencing data originated include RMDisease V2.0, RMVar V2.0, RMBase V3.0, and m7GHub V2.0, covering 62 RNA modification types (including at least m...). 6 A、m 5 C, m 1 A, Am, Um, acp 3 U、m 7 The dataset (G, Cm, Gm) involves 73 species. The original third-generation sequencing data comes from databases including GEO, SRA, and SG-Nex. Combining this with public databases ensures the quality and diversity of the training data. To screen for high-confidence RNA modification sites of various types, validation using multiple experimental techniques is necessary, such as m... 5 C-RIP-seq sequencing technology, m 6Sites can only be preserved if validated by at least two different experimental techniques, such as ACE-seq, MePMe-seq, MeRIP-seq, and miCLIP. Furthermore, the position weight matrix (PWMs) for each site modification type must be greater than 3. The PWM is also known as the position-specific weight matrix (PSWM) or position-specific scoring matrix (PSSM). The RMBase v3.0 database uses PWMs to assign scores to each RNA modification site, enriching annotations with genomic information, cell lines, and experimental datasets, and revealing their evolutionary conservation across different mammals.

[0042] To ensure data quality, step S2 aligns the current signals in the third-generation sequencing data with the reference genome, completing the basecall operation. The basecall operation is a crucial step in the third-generation sequencing data processing workflow. Its core task is to convert the current signals detected by the sequencer into corresponding base sequences (A, C, G, U / T) to align high-confidence modification sites with the current signals. Generally, the basecall operation uses complex computational algorithms (usually deep neural networks) to analyze the current signals. For each feature in the current signal, it identifies which base or base combination it most likely represents, and then connects these identified bases in the correct order to form a nucleotide sequence, including an estimate of the probability of correctly identifying each base in the sequence.

[0043] In this embodiment, the Bonito basecalling model (v0.8.1-equivalent; bonito train–epochs 8–batch 64–lr 5e-4) was retrained using human cell RNA, followed by regular basecall (bonito basecaller–rna–reference GRCh38.p14.genome.fa–batchsize 256) for data calibration.

[0044] Step S3 performs a secondary alignment between high-confidence modification sites in the second-generation sequencing data and current signals in the third-generation sequencing data after the first alignment. The current signals are then re-analyzed using the Remora tool (v3.2.0, k-mer model: RNA R9.4 180mV 70bps; parameters: do rough rescale = True, scale iters = 0 for global scaling). Dynamic time warping and refitting are performed on the current signals and bases, extracting the mean, median, standard deviation, and dwell time of the bases and current signals as the corresponding current signal features. The Remora tool primarily aims to detect DNA or RNA base modifications and improve the accuracy of base identification in certain complex regions. Essentially, it utilizes deep neural networks to focus on specified key sites based on the aforementioned basecall localization, extracting raw signal fragments containing rich contextual information, and applying a specially trained model to identify whether a base modification exists at that site or to more accurately infer its true base identity. Its output provides fine-grained, site-specific modification information in the form of modification probabilities, further improving the accuracy of modified sites.

[0045] In addition, the impact of different input base sequence lengths on model performance was tested through a system test, and 100bp was determined as the optimal input length, which balances computational efficiency and resource consumption.

[0046] In one specific embodiment of the present invention, step S4 establishes an RNA modification identification model based on cross-modal feature fusion and multi-scale sequence modeling. For example... Figure 2 As shown, the RNA modification identification model mainly consists of two parts: a cross-modal feature fusion part and a multi-scale sequence modeling part.

[0047] The cross-modal feature fusion part uses a cross-attention mechanism to fuse base features and current signal features. Specifically, the base sequence features are used as queries and the current signal features are used as key-value pairs to achieve adaptive weighting and fusion of features.

[0048] The multi-scale sequence modeling part adopts the StripedHyena 2 architecture of Evo 2, combining short-range explicit convolution (SE), mid-range regularized convolution (MR), and long-range implicit convolution (LI) to perform multi-scale sequence modeling on the fused results. This architecture supports multi-scale sequence modeling from single nucleotides to millions of bases. Here, the StripedHyena 2 architecture of Evo 2 is suitable for modeling ultra-long biological sequences, while fusing sequence features at different scales (short / medium / long). Short-range explicit convolution captures local patterns from single nucleotides to short fragments, using lightweight convolution kernels (such as 1D convolution); mid-range regularized convolution models the complex structure of medium-length sequences, introducing a regularization mechanism to improve generalization ability; long-range implicit convolution covers global dependencies over ultra-long ranges. The StripedHyena 2 architecture of Evo 2 is a standard technique in this field and will not be elaborated further here.

[0049] To improve the learning efficiency and prediction accuracy of rare modifications, especially the prediction ability of modification types in a small number of glycosylated RNA samples, this invention first utilizes a subset of samples with a large number of modification types (m) from a high-confidence training set. 6 A、m 5 C) Pre-trained RNA modification identification model. During pre-training, the modification type is m 6 A、m 5 The sample C (base features and current signal features) is used as input. The multi-scale sequence modeling results are modified by the classification layer, and the pre-training process is achieved by using the cross-entropy loss function.

[0050] The model employs a transfer learning strategy, based on the shared patterns of current signal features, to transfer known high-confidence RNA sample modification types (m... 6 A、m 5 C) pre-trained weights are transferred to a small number of modified types (n<1000:Am,m) 1 A,Um,Cm,Gm,m 7 G;n<30:acp 3 On U), utilize a smaller subset of samples (Am, m) with different modification types. 1 A, Um, Cm, Gm, m 7 G, acp 3 U) Fine-tuning the pre-trained RNA modification identification model, using regularization and early stopping strategies, ensures the stability and prediction accuracy of transfer learning.

[0051] The modification identification model trained in this invention supports accurate prediction of multiple types of modifications, and is particularly well-suited for the identification of glycosylated RNA modifications.

[0052] This invention utilizes a transfer learning strategy to transfer m-based 6 A and m 5Pre-training weights trained on C data were transferred to a small number of glycosylated RNA sample modification types (e.g., Am, m). 1 A, Um, Cm, Gm, m 7 G and acp 3 The model was fine-tuned on a limited dataset. By optimizing model performance through regularization and early stopping strategies, the learning efficiency and prediction accuracy for rare modifications were significantly improved, enhancing the predictive ability for modification types in a small number of glycosylated RNA samples. This is the first algorithm specifically designed for glycoRNA modifications, filling a gap in the field of glycoRNA modification detection using nanopore sequencing.

[0053] In this embodiment, the modification identification model was validated across cell lines to evaluate its predictive ability and generalization performance. Cross-cell line validation was performed on untrained cell lines, and multiple performance metrics were used to evaluate the model's predictive ability and generalization performance. Specifically, the model performance was tested using H9 cell line data and compared with training data (HeLa cell line). Multiple evaluation metrics (such as ROCAUC, PRAUC, F1, and MCC) were used to comprehensively evaluate the model performance. Test results show that the model can adapt to the modification characteristics of different cell lines and has good generalization performance. Compared with other advanced models, the glycosylated RNA modification identification model of this invention demonstrates better generalization performance in m... 6 A and m 5 In the C modification detection task, this model outperforms other cutting-edge RNA modification identification models (such as TandemMod, Nanom6A, and RNA-FM). By employing a 70:10:20 split strategy for training, validation, and testing datasets, and using the Adam optimizer (learning rate 1.5e-5, weight decay 0.01) for 50 training rounds, along with an early stopping strategy based on the validation set, performance metrics ROC_AUC, PR_AUC, F1, and MCC are all significantly improved. Detailed comparison results are shown in Table 1.

[0054] Table 1. Performance comparison of the modification identification model of this invention with other cutting-edge modification identification models.

[0055]

[0056] The RNA modification identification model of this invention and an RNA modification identification model constructed without the transfer learning strategy of this invention were compared in different modification detection tasks. The former showed significant improvements in performance metrics ROC_AUC, PR_AUC, F1, and MCC. The transfer learning method of this invention ensures the stability and prediction accuracy of transfer learning. Detailed comparison results are shown in Table 2.

[0057] Table 2 compares the performance of models built with and without transfer learning strategies.

[0058]

[0059] The experimental data above demonstrate that, compared to existing RNA modification identification models (such as TandemMod, Nanom6A, and RNA-FM), the modification identification model of this invention exhibits good adaptability in detecting modification features across different cell lines. The transfer learning strategy significantly improves the learning efficiency and prediction accuracy of rare modifications, and enhances the predictive ability for modification types in a small number of glycosylated RNA samples.

[0060] Based on the same inventive concept, this embodiment also provides a system for constructing a glycosylated RNA modification identification model based on nanopore sequencing, including:

[0061] The nanopore sequencing data collection module is used to acquire raw second-generation and third-generation sequencing data of high-confidence modification sites from RNA samples of multiple species from various public database sources. The modification types of the RNA from these multiple species include m... 6 A、m 5 C, m 1 A, Am, Um, acp 3 U、m 7 G, Cm, Gm;

[0062] The data preprocessing module is used to align the current signal in the third-generation sequencing data with the reference genome in one step.

[0063] The high-confidence training set module is used to perform a second alignment between the high-confidence modification sites in the second-generation sequencing data and the current signals in the third-generation sequencing data after the first alignment. It extracts the features of each base current signal from the aligned current signals to obtain a high-confidence training set containing base feature sequences, current signal feature sequences, and each base modification type.

[0064] The RNA modification identification model training module is used to pre-train an RNA modification identification model based on cross-modal feature fusion and multi-scale sequence modeling using a sample subset with a large number of modification types in the high-confidence training set. Then, the pre-trained RNA modification identification model is fine-tuned using a sample subset with a smaller number of modification types to obtain a modification identification model adapted to glycosylated RNA.

[0065] For the system embodiments, since they basically correspond to the method embodiments, relevant details can be found in the descriptions of the method embodiments; the implementation methods of the remaining modules will not be repeated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0066] The system embodiments of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar device. The system embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution.

[0067] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. Those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for constructing a glycosylated RNA modification identification model based on nanopore sequencing, characterized in that, include: Raw second-generation sequencing and third-generation sequencing data of high-confidence modification sites from RNA samples of multiple species were obtained from various public database sources. The modification types of RNA from multiple species include m 6 A、m 5 C, m 1 A, Am, Um, acp 3 U、m 7 G, Cm, Gm; Align the current signals in the third-generation sequencing data with the reference genome once; The high-confidence modification sites in the second-generation sequencing data are aligned with the current signals in the third-generation sequencing data after the first alignment. The features of each base current signal are extracted from the aligned current signals to obtain a high-confidence training set containing base feature sequences, current signal feature sequences, and each base modification type. An RNA modification identification model based on cross-modal feature fusion and multi-scale sequence modeling was pre-trained using a sample subset with a large number of modification types in the high-confidence training set. Then, the pre-trained RNA modification identification model was fine-tuned using a sample subset with a smaller number of modification types to obtain a modification identification model adapted to glycosylated RNA. The sample subset used for pre-training the RNA modification identification model must contain at least m modification types. 6 A、m 5 C; The sample subset used to fine-tune the RNA modification identification model must contain at least ACP modifications. 3 U; The RNA modification identification model first uses a cross-attention mechanism to fuse base features and current signal features, where base features are used as queries and current signal features are used as key-value pairs. Then, the Evo 2 StripedHyena 2 architecture is used to perform multi-scale sequence modeling on the fusion result by combining short-range explicit convolution (SE), mid-range regularized convolution (MR), and long-range implicit convolution (LI). Based on the final modeling features, the model predicts the type of each base modification.

2. The method for constructing a glycosylated RNA modification identification model based on nanopore sequencing according to claim 1, characterized in that, The database from which the original next-generation sequencing data is sourced includes at least two of the following: RMDisease V2.0, RMVar V2.0, RMBase V3.0, and m7GHub V2.

0.

3. The method for constructing a glycosylated RNA modification identification model based on nanopore sequencing according to claim 1, characterized in that, The original third-generation sequencing data source databases include at least two of the GEO, SRA, and SG-Nex databases.

4. The method for constructing a glycosylated RNA modification identification model based on nanopore sequencing according to claim 1, characterized in that, High-confidence modification sites refer to sites that have been verified by at least two different experimental techniques, and the site modification type position weight matrix PWMs > 3, and the training set contains more than one million high-confidence modification sites.

5. The method for constructing a glycosylated RNA modification identification model based on nanopore sequencing according to claim 1, characterized in that, The characteristics of each base current signal include the mean, median, standard deviation, and dwell time of the current signal.

6. The method for constructing a glycosylated RNA modification identification model based on nanopore sequencing according to claim 1, characterized in that, The process of aligning the current signals in third-generation sequencing data with the reference genome includes: The Bonito basecalling model was retrained using human cell data; The retrained Bonito basecalling model was used to perform basecall operations on the third-generation sequencing data to complete the data calibration and achieve a one-time alignment process.

7. The method for constructing a glycosylated RNA modification identification model based on nanopore sequencing according to claim 1, characterized in that, The secondary alignment process was implemented using the Remora current signal reanalysis tool.

8. A system for constructing a glycosylated RNA modification identification model based on nanopore sequencing, used to implement the method for constructing a glycosylated RNA modification identification model based on nanopore sequencing as described in claim 1, characterized in that, include: The nanopore sequencing data collection module is used to acquire raw second-generation and third-generation sequencing data of high-confidence modification sites from RNA samples of multiple species from various public database sources. The modification types of the RNA from these multiple species include m... 6 A、m 5 C, m 1 A, Am, Um, acp 3 U、m 7 G, Cm, Gm; The data preprocessing module is used to align the current signal in the third-generation sequencing data with the reference genome in one step. The high-confidence training set module is used to perform a second alignment between the high-confidence modification sites in the second-generation sequencing data and the current signals in the third-generation sequencing data after the first alignment. It extracts the features of each base current signal from the aligned current signals to obtain a high-confidence training set containing base feature sequences, current signal feature sequences, and each base modification type. The RNA modification identification model training module is used to pre-train an RNA modification identification model based on cross-modal feature fusion and multi-scale sequence modeling using a sample subset with a large number of modification types in the high-confidence training set. Then, the pre-trained RNA modification identification model is fine-tuned using a sample subset with a smaller number of modification types to obtain a modification identification model adapted to glycosylated RNA.

Images