A deep learning-based method and system for analyzing the complete structure of o-linked glycans

Abstract

The application discloses an O-linked glycan complete structure analysis method and system based on deep learning, and belongs to the technical field of bioinformatics and glycoproteomics. The method comprises the following steps: acquiring tandem mass spectrometry data of O-glycopeptide samples under data-dependent mode acquisition; pre-processing the mass spectrometry data to construct a fragment ion feature matrix; inputting the feature matrix into a pre-trained multi-task deep learning model, wherein the model comprises a graph neural network coding module, a Transformer feature extraction module and a multi-task decoding module; simultaneously outputting monosaccharide composition sequence, branch topological structure, glycosidic bond connection mode and O-glycosylation site of the O-linked glycan through multi-task decoding; and outputting complete structure information after confidence evaluation of the output result. The application adopts a graph neural network to accurately model the nonlinear branch structure of the O-linked glycan, uses a Transformer to extract long-range dependence features of a spectrum, realizes end-to-end complete structure analysis through multi-task learning, does not depend on a database, can discover unknown O-linked glycan structures, and has good generalization ability and application prospect.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of bioinformatics, glycoproteomics, and deep learning, specifically to a method and system for resolving the complete structure of O-linked glycans based on deep learning. It is applicable to the automated resolution of the polypeptide amino acid sequence, O-glycosylation sites, monosaccharide composition, branching structure, and glycosidic bond linkage of complete O-glycopeptides from tandem mass spectrometry data acquired in a data-dependent mode. Background Technology

[0002] Glycosylation is one of the most common and complex post-translational modifications of proteins, occurring in more than half of all proteins in mammals. O-linked glycosylation occurs at serine or threonine residues in proteins, starting with N-acetylgalactosamine (GalNAc). O-linked glycans play crucial roles in biological processes such as immune defense, cell growth, intercellular adhesion, inflammatory responses, and signal transduction, and their aberrant modifications are closely related to tumor proliferation, invasion, and metastasis.

[0003] However, mass spectrometric analysis of O-glycopeptides is far more difficult than that of N-glycopeptides, mainly in the following aspects:

[0004] (1) O-glycosylation lacks a common sequence like N-glycosylation "NXS / T", and adjacent O-glycosylation sites are often more densely packed, making site prediction extremely difficult;

[0005] (2) O-glycopeptides have low stoichiometry and poor ionization efficiency in mass spectrometry detection, resulting in weak mass spectrometry signals;

[0006] (3) O-glycans have up to eight core structures (core 1 to core 8), complex branching structures, and a large number of structural isomers;

[0007] (4) O-linked glycans have a non-linear dendritic branching structure, and the connection modes between monosaccharides (1→2, 1→3, 1→4, 1→6, etc.) and α / β configurations are diverse, making the complete structure analysis extremely challenging.

[0008] Currently, the structural analysis of O-linked glycans mainly relies on database search methods and expert interpretation. Existing technologies have the following shortcomings:

[0009] First, database search methods are limited by the coverage of existing databases. Currently used technologies such as GlycanFinder rely on predefined rule bases and databases of known glycan structures. Due to the extremely high structural diversity of O-linked glycans, many novel glycan structures are not included in databases, making it impossible for database search methods to identify unknown structures.

[0010] Second, existing deep learning methods primarily focus on spectral prediction rather than complete structural analysis. For example, CN116825200B discloses a deep learning-based glycopeptide mass spectrum prediction device that divides glycopeptides into peptide segments and glycan chains, encodes them separately, and then predicts the mass spectrum. While this method can predict the mass spectrum of glycopeptides, it does not solve the core problem of inferring the complete structure of glycans from the mass spectrum.

[0011] Third, existing glycan structure prediction models mainly target N-linked or free glycans. For example, CandyCrunch uses an extended residual neural network to predict glycan structures from LC-MS / MS data; GlycoBERT and GlycoBART use a Transformer architecture for glycan structure prediction; and MS2Glycan uses deep learning methods to iteratively generate glycan structures. However, these models mainly target N-linked or free glycans, have limited ability to handle O-linked glycans, and fail to simultaneously resolve O-glycosylation sites.

[0012] Fourth, there is a lack of effective modeling for the branching structure of O-linked glycans. The nonlinear branching structure of O-linked glycans is difficult to describe using traditional sequence representation methods, and existing linear sequence models cannot accurately capture the branching topology information.

[0013] Fifth, existing methods cannot achieve multi-task joint analysis. Complete analysis of O-glycopeptides requires determining information at three levels: peptide sequence, glycosylation site, and glycan structure. However, existing technologies are usually only for a single task and cannot achieve complete end-to-end analysis.

[0014] In summary, there is an urgent need for a method and system that can automatically and accurately resolve the complete structure of O-linked glycans (including monosaccharide composition, branching topology, and glycosidic bond linkage) from O-glycopeptide tandem mass spectrometry data acquired in data-dependent mode, and simultaneously determine the O-glycosylation sites. Summary of the Invention

[0015] (a) Purpose of the invention

[0016] This invention addresses the problems of existing O-linked glycan structure analysis techniques, such as strong database dependence, inability to handle unknown structures, insufficient ability to model branch structures, and inability to achieve multi-task joint analysis. It provides a deep learning-based method and system for analyzing the complete structure of O-linked glycans.

[0017] (II) Technical Solution

[0018] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0019] 1. Overall Technical Approach

[0020] This invention employs a technical architecture of "graph modeling-feature extraction-multi-task decoding". First, the nonlinear branching structure of O-linked glycans is encoded as a graph structure representation, and a graph neural network (GNN) is used to capture the branch topology information. Second, the powerful feature extraction capability of the Transformer architecture is utilized to extract long-range dependent features from MS / MS fragment ion spectra. Finally, the multi-task decoding module simultaneously outputs the monosaccharide composition, branching structure, glycosidic bond connection mode, and O-glycosylation sites of O-linked glycans, achieving a complete end-to-end structural analysis.

[0021] 2. Methods and Steps

[0022] Step S1: Mass Spectrometry Data Acquisition

[0023] Tandem mass spectrometry (MS / MS) data of the O-glycopeptide sample to be analyzed were acquired in data-dependent acquisition mode. The sample was separated by liquid chromatography and then acquired by mass spectrometry in data-dependent mode to obtain a precursor ion list and corresponding MS / MS fragment spectra.

[0024] Step S2: Data Preprocessing

[0025] The MS / MS spectral data obtained in step S1 are preprocessed, specifically including:

[0026] • Peak identification: Identifying valid signal peaks in the spectrum;

[0027] • Remove isotope peaks: Remove interference from isotope peaks;

[0028] • Noise filtering: Noise signals are filtered out based on the signal-to-noise ratio threshold;

[0029] • Normalization: Normalize the fragment ion intensity to a uniform dimension.

[0030] After preprocessing, the mass-to-charge ratio (m / z) and intensity information of each fragment ion were extracted to construct a fragment ion feature matrix.

[0031] Step S3: Deep Learning Model Analysis

[0032] The preprocessed fragment ion feature matrix is ​​input into a pre-trained multi-task deep learning model. The model comprises three core modules:

[0033] (1) Graph Neural Network Encoding Module

[0034] This module encodes O-linked glycans as a directed acyclic graph structure G=(V, E), where:

[0035] • The node set V represents a monosaccharide unit, and the feature vector of each node contains monosaccharide type information (such as HexNAc, Hex, Fuc, NeuAc, NeuGc, etc.).

[0036] • The edge set E represents glycosidic bond connections, and the attributes of each edge include the connection position (such as 1→2, 1→3, 1→4, 1→6, etc.) and α / β configuration information;

[0037] • Graph neural networks capture the branch topology information of glycans by iteratively updating the feature representations of nodes and edges through a message passing mechanism.

[0038] (2) Transformer Feature Extraction Module

[0039] This module employs a multi-head self-attention mechanism to extract long-range dependent features from the fragment ion feature matrix. Specifically:

[0040] • Use the m / z and intensity of fragment ions as sequence inputs;

[0041] • Sequence position information of injected fragment ions is encoded by position;

[0042] • Multilayer Transformer encoders capture the correlations between different fragment ions through a self-attention mechanism;

[0043] Output a high-dimensional feature representation for use by the decoding module.

[0044] (3) Multi-task decoding module

[0045] This module adopts a multi-task learning architecture, sharing underlying feature representations and simultaneously outputting through multiple task-specific decoding heads:

[0046] • Task 1: Monosaccharide composition prediction – Output the types and quantities of monosaccharides contained in O-linked polysaccharides;

[0047] • Task 2: Branch Topology Prediction – Output the branch structure of the glycan, including core type and branch extension pattern;

[0048] • Task 3: Predicting Glycosidic Bond Linkage Modes – Output the linkage positions and α / β configurations of each glycosidic bond;

[0049] • Task 4: O-glycosylation site prediction - Output the specific linkage sites (Ser / Thr positions) of O-linked glycans on peptides.

[0050] During the decoding process, a beam search algorithm is used to generate candidate O-linked glycan structures, and the candidate structures are sorted and screened based on the fragment ion matching degree.

[0051] Step S4: Output Results

[0052] The multi-task decoding module outputs complete structural information of the O-linked glycan, including:

[0053] • Monosaccharide composition and sequence;

[0054] • Branching topology (represented as a graph or tree structure);

[0055] • The α / β configuration and linkage position of each glycosidic bond;

[0056] • Linkage sites of O-linked glycans on peptides.

[0057] Step S5: Confidence Assessment

[0058] The output of step S4 is evaluated for confidence, and a confidence score is predicted for each structure. When the confidence score is lower than a preset threshold, the manual verification mode is activated, and the candidate structure and its corresponding MS / MS spectrum are pushed to the user interface for manual review and confirmation.

[0059] 3. Model Training

[0060] The training dataset for the multi-task deep learning model contains no fewer than 100,000 experimentally validated MS / MS spectra of fully annotated O-linked glycan structures. Annotation information includes:

[0061] • Monosaccharide composition (types and quantities of each monosaccharide);

[0062] • Branching structure (core type, number of branches, and extension pattern);

[0063] • Glycosidic bond linkage (linkage position and α / β configuration);

[0064] • O-glycosylation site (the specific Ser / Thr position on the peptide).

[0065] Training employs a multi-task joint loss function:

[0066] L_total = α·L_composition + β·L_topology + γ·L_linkage + δ·L_site

[0067] Where L_composition, L_topology, L_linkage, and L_site are the loss functions for the four prediction tasks, and α, β, γ, and δ are the weight coefficients for each task.

[0068] 4. System Composition

[0069] This invention also provides a deep learning-based system for resolving the complete structure of O-connected glycans, comprising:

[0070] • Data acquisition module: Used to acquire tandem mass spectrometry data of the O-glycopeptide sample to be analyzed;

[0071] • Data preprocessing module: used for peak identification, isotope peak removal, noise filtering, and normalization of mass spectrometry data;

[0072] • Graph Neural Network Encoding Module: Used to encode the nonlinear branching structure of O-connected glycans into a graph structure representation;

[0073] • Transformer Feature Extraction Module: Used to extract long-range dependent features from fragment ion spectra;

[0074] • Multi-task decoding module: used to simultaneously output the monosaccharide composition sequence, branching topology, glycosidic bond linkage mode, and O-glycosylation sites of O-linked glycans;

[0075] • Confidence assessment module: Used to assess the confidence level of the output results;

[0076] • Output module: Used to output the complete structural information of O-linked glycans and their corresponding confidence scores.

[0077] The modules mentioned above adopt an end-to-end joint training approach and share the underlying feature representation.

[0078] (III) Beneficial Effects

[0079] Compared with the prior art, the present invention has the following beneficial effects:

[0080] 1. Achieve end-to-end resolution of the complete structure of O-linked glycans. This invention is the first to combine graph neural networks, Transformer, and multi-task learning to achieve end-to-end automated resolution of the complete structure of O-linked glycans (monosaccharide composition, branching topology, glycosidic bond linkage) and O-glycosylation sites from raw MS / MS spectra, overcoming the limitation of existing technologies that can only handle single tasks.

[0081] 2. Effective modeling of the nonlinear branching structure of O-connected glycans. This invention uses a graph neural network to encode O-connected glycans into a directed acyclic graph structure, which can accurately capture the branching topology information of glycans and overcome the shortcomings of traditional linear sequence models in handling branching structures.

[0082] 3. Unreliant on databases, capable of discovering unknown structures. This invention employs a generative deep learning method that does not rely on existing glycan structure databases. It can predict novel O-connected glycan structures not found in the training data, demonstrating excellent generalization ability and the ability to discover unknown structures.

[0083] 4. Fully utilize long-range dependency information in MS / MS spectra. This invention employs the Transformer architecture to extract long-range dependency features from fragment ion spectra, which can capture the correlation between different fragment ions and improve the accuracy of structure prediction.

[0084] 5. Achieve multi-task joint optimization. This invention adopts a multi-task learning architecture, where the four prediction tasks share the underlying feature representation. Through joint training, they mutually promote each other, improving the prediction accuracy of each task.

[0085] 6. Provides confidence assessment and manual verification mechanisms. This invention assesses the confidence level of the prediction results and initiates a manual verification mode for results with low confidence levels, thus balancing the degree of automation in the analysis with the reliability of the results.

[0086] References

[0087] 1. Suideng Qin, Zhixin Tian* (Inventors of the present invention). Deep structure-levelN-glycan identification using feature-induced structure diagnosis integrated with a deep learning Model. Analytical and Bioanalytical Chemistry, 2024,417, 1001–1014

[0088] 2. Xiang Menghui; Tian Zhixin (inventors of this invention). A complete O-glycopeptide relative quantitative analysis method based on mass spectrometry. Chinese Patent, Application No. CN202410864029.8, Application Date June 30, 2024.

[0089] 3. Zihan Zhang, Hua Xu* and Zhixin Tian* (Inventors of this invention). Exploration of quantitative site-specific serum O-glycoproteomics with isobaric labelling for the discovery of putative O-glycoprotein biomarkers. Proteomics-Clinical Applications, 2022, e2100095

[0090] 4. James Urban, Chunsheng Jin, Kristina A. Thomsson, Niclas G.Karlsson, Callum M. Ives, Elisa Fadda, Daniel Bojar. Predicting glycanstructure from tandem mass spectrometry via deep learning. BioRxiv, 2023,10.1101 / 2023.06.13.544793.

[0091] 5. Ejas Althaf Abtheen, Arun Singh, Shyam Sriram, Changyou Chen, Sriram Neelamegham, Rudiyanto Gunawan. Transformer-based Deep Learning for Glycan Structure Inference from Tandem Mass Spectrometry. BioRxiv, 2025,10.1101 / 2025.07.02.662857.

[0092] 6. Yang Yi; Fang Qun. A deep learning-based glycopeptide mass spectrometry prediction device. Chinese Patent, Application No. CN202310828711.7, Application Date July 7, 2023.

[0093] 7. Defeng Li, Zizheng Nie, Yanmin Liu, Xiaojun Cai, Yong Zhang,Xuefeng Cui. Iterative Glycan Structure Generation by Component Localizationand Classification from Mass Spectra. IEEE International Conference onBioinformatics and Biomedicine (BIBM), Lisbon, 2024, 10.1109 / BIBM62325.2024.10821861.

[0094] 8. James Urban, Roman Joeres, Luc Thomès, Kristina A. Thomsson,Daniel Bojar. Navigating the maze of mass spectra: a machine-learning guideto identifying diagnostic ions in O-glycan analysis. Analytical andBioanalytical Chemistry, 2025, 417, 931-943.

[0095] 9. Yu Zong, Yuxin Wang, Liang Qiao. Deep Learning Prediction ofIntact N- / O-Glycopeptide Tandem Mass Spectra Enhances Glycoproteomics.BioRxiv, 2025, 10.1101 / 2025.07.20.665808.

[0096] 10. Yu Zong, Yuxin Wang, Liang Qiao. Deep Learning Prediction of O-Glycopeptide Tandem Mass Spectra Enhances O-Glycoproteomics. AdvancedScience, 2026, 10.1002 / advs.76760

[0097] 11. Shotaro Kumozaki, Kengo Sato, and Yasubumi Sakakibara. A MachineLearning Based Approach to de novo Sequencing of Glycans from Tandem MassSpectrometry Spectrum. IEEE-ACM Transactions of Computational Biology andBioinformatics, 2015, 12, 1267-1274.

[0098] 12. Weiping Sun, Qianqiu Zhang, Xiyue Zhang, Ngoc Hieu Tran, M. ZiaurRahman, Zheng Chen, Chao Peng, Jun Ma, Ming Li, Lei Xin & Baozhen Shan. Glycopeptide database search and de novo sequencing with PEAKS GlycanFinderenable highly sensitive glycoproteomics. Nature Communications, 2023, 14, 4046. Attached Figure Description

[0099] Figure 1 This is an overall flowchart of the deep learning-based method for resolving the complete structure of O-linked glycans according to the present invention. The process begins with acquiring MS / MS mass spectrometry data obtained using a data-dependent pattern. After preprocessing, the data is input into a multi-task deep learning model. The model uses Transformer feature extraction and multi-task decoding to output the complete structural information of the O-linked glycan. Finally, the final result is output after confidence evaluation.

[0100] Figure 2 This is a schematic diagram of the network architecture of the multi-task deep learning model of the present invention. The fragment ion feature matrix, after input embedding and position encoding, is input to a multi-layer Transformer encoder to extract long-range dependent features, obtaining a shared feature representation. These shared features are then input to a graph neural network encoding module and four task-specific prediction heads to achieve multi-task parallel decoding. Finally, after beam search and structure sorting, complete structural information is output.

[0101] Figure 3 This is a schematic diagram of the graph structure encoding of O-linked glycans. The left side shows a typical core type 1 O-linked glycan structure (NeuAcα2-3Galβ1-3GalNAcα1-Ser / Thr), and the right side shows its corresponding graph structure encoding. Nodes represent monosaccharide units, edges represent glycosidic bonds, and edge attributes include the connection position (e.g., α2-3, β1-3, β1-4, etc.) and configuration information. The root node is the GalNAc linked to the Ser / Thr residues of the peptide. Nodes represent monosaccharide units (GalNAc, Gal, GlcNAc, NeuAc, Fuc, Man, etc.); edges represent glycosidic bonds (connection position + α / β configuration); and the root node is the starting monosaccharide (GalNAc) linked to the peptide.

[0102] Figure 4 This diagram illustrates a comparison of the analytical results of the method of this invention and traditional database search methods. Traditional database search methods rely on existing databases for spectral matching, cannot identify unknown structures, and have limited ability to analyze branch structures. The method of this invention employs an end-to-end generative architecture based on deep learning, does not rely on a database, and can simultaneously analyze monosaccharide composition, branch topology, glycosidic bond linkages, and O-glycosylation sites, achieving complete structural analysis of O-linked glycans. Detailed Implementation

[0103] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0104] Example 1: Implementation of the method for resolving the complete structure of O-linked glycans

[0105] This embodiment provides a method for resolving the complete structure of O-linked glycans based on deep learning, which is implemented according to the following steps.

[0106] Step S1: Mass Spectrometry Data Acquisition

[0107] Acquire tandem mass spectrometry (MS / MS) data of the O-glycopeptide sample to be analyzed in Data-Dependent Acquisition (DDA) mode. The specific implementation method is as follows:

[0108] (1) Sample preparation and liquid chromatography separation

[0109] The O-glycopeptide sample to be analyzed was digested with trypsin or proteinase K, and then separated using a C18 reversed-phase liquid chromatography column. The chromatographic conditions were as follows: mobile phase A was 0.1% formic acid aqueous solution, mobile phase B was 0.1% formic acid in acetonitrile solution, gradient elution program was 2%~35% mobile phase B, elution time was 90 minutes, and flow rate was 300 nL / min.

[0110] (2) Mass spectrometry data acquisition

[0111] Data acquisition was performed using a nano-level liquid chromatography-tandem mass spectrometry (LC-MS / MS) system (such as the Thermo Scientific Orbitrap Fusion Lumos or the Bruker timsTOF Pro). The mass spectrometry parameters were set as follows:

[0112] • Ionization mode: Electrospray ionization positive ion mode (ESI+);

[0113] • Spray voltage: 2.0~2.5 kV;

[0114] • Ion transfer tube temperature: 275~320℃;

[0115] • MS1 full scan range: m / z 350~2000, resolution 120,000 (@ m / z 200);

[0116] • Data-dependent model: Select the 20 most abundant precursor ions for MS / MS fragmentation;

[0117] • Fragmentation mode: High-energy collision-induced dissociation (HCD), with normalized collision energy set at 20%~35%;

[0118] • MS / MS resolution: 30,000 (@ m / z 200);

[0119] • Dynamic exclusion time: 20~30 seconds.

[0120] The above-mentioned data collection yielded a list of precursor ions and corresponding MS / MS fragment spectral data for each precursor ion.

[0121] Step S2: Data Preprocessing

[0122] The MS / MS spectral data obtained in step S1 is preprocessed as follows:

[0123] (1) Peak identification

[0124] A peak detection algorithm based on Continuous Wavelet Transform (CWT) was used to identify valid signal peaks in the spectrum. A signal-to-noise ratio (SNR) threshold of at least 3:1 was set, and only signal peaks with an SNR greater than this threshold were retained. The quality window for peak identification was set to ±10 ppm.

[0125] (2) Remove isotope peaks

[0126] Based on the natural abundance distribution of isotope peaks, an isotope peak cluster identification algorithm is used to remove isotope peak interference. For each isotope peak cluster, only monoisotopic peaks are retained, while isotope peaks of elements such as C, N, O, and S (e.g., M+1 and M+2 peaks) are removed. The criterion is that peaks with intensities lower than the expected intensity at the corresponding position in the theoretical isotope abundance distribution are removed.

[0127] (3) Noise filtering

[0128] A dynamic threshold is set based on the baseline noise level of the spectrum to filter out high-frequency noise signals with an intensity below the threshold. The noise threshold is calculated as follows: calculate the intensity distribution of all peaks in the spectrum, take the 5th percentile of the intensity distribution as the noise baseline, and multiply it by a preset coefficient (usually 3~5) to obtain the filtering threshold.

[0129] (4) Normalization

[0130] The fragment ion intensity is normalized to a uniform dimension. Two normalization methods are specifically employed:

[0131] • Total ion current normalization: Divide the intensity of each fragment ion by the total ion current intensity of the spectrum so that the sum of the intensities of all fragment ions is 1;

[0132] • Maximum peak normalization: Divide the intensity of each fragment ion by the intensity of the highest peak in the spectrum to normalize the maximum peak intensity to 1.

[0133] After preprocessing, the mass-to-charge ratio (m / z) and normalized intensity information of each fragment ion are extracted to construct a fragment ion feature matrix. Each row of the feature matrix corresponds to a fragment ion and contains two dimensions: m / z value and normalized intensity value. For the m / z value, a fixed-width binning strategy is used, or the original m / z value is used directly as the input feature.

[0134] Step S3: Deep Learning Model Analysis

[0135] The preprocessed fragment ion feature matrix is ​​input into a pre-trained multi-task deep learning model. The implementation details of the model are as follows:

[0136] (1) Graph Neural Network Encoding Module

[0137] This module encodes O-linked glycans as a directed acyclic graph structure G=(V, E). The specific implementation is as follows:

[0138] • Node Feature Encoding: The node set V represents monosaccharide units, and the initial feature vector of each node is a one-hot encoding of the monosaccharide type. Monosaccharide types include, but are not limited to: N-acetylgalactosamine (GalNAc), galactose (Gal), N-acetylglucosamine (GlcNAc), fucose (Fuc), N-acetylneuraminic acid (NeuAc), N-hydroxyacetylneuraminic acid (NeuGc), mannose (Man), glucose (Glc), xylose (Xyl), etc. The node feature vector dimension is 10~20 dimensions.

[0139] • Edge feature encoding: The edge set E represents glycosidic bond connections. The attributes of each edge include the connection position (e.g., 1→2, 1→3, 1→4, 1→6, etc.) and α / β configuration information. The feature vector of the edge is a joint encoding of the connection position and configuration, with a dimension of 8 to 12.

[0140] • Message Passing in Graph Neural Networks: Graph Isomorphism Networks (GINs) or Graph Attention Networks (GATs) are used as the basic architecture of graph neural networks. The feature representations of nodes and edges are iteratively updated through K rounds of message passing (K is typically 3-5). The formula for each round of message passing is:

[0141] h_v^{(k+1)} = UPDATE(h_v^{(k)}, AGGREGATE({h_u^{(k)}, e_{uv}^{(k)} :u ∈ N(v)}))

[0142] Where h_v^{(k)} is the feature representation of node v at layer k, e_{uv}^{(k)} is the feature representation of edge (u,v) at layer k, N(v) is the set of neighboring nodes of node v, AGGREGATE is the aggregation function (such as summation, averaging or attention weighting), and UPDATE is the update function (such as multilayer perceptron).

[0143] • Graph Readout: After K rounds of message passing, global pooling operations (such as global summation pooling or global average pooling) are used to aggregate node features into graph-level feature vectors, which serve as a representation of the overall glycan structure.

[0144] (2) Transformer Feature Extraction Module

[0145] This module employs a multi-head self-attention mechanism to extract long-range dependent features from the fragment ion feature matrix. The specific implementation is as follows:

[0146] • Input embedding: The m / z values ​​and normalized intensity values ​​of fragment ions are mapped to d_model-dimensional embedding vectors through a linear projection layer (d_model is usually 256~512).

[0147] • Position encoding: Sine-cosine position encoding or learnable position encoding is used to inject the sequence position information of fragment ions into the embedding vector. The dimension of the position encoding is the same as the dimension of the embedding vector.

[0148] • Transformer Encoder: Employs a multi-layer Transformer encoder (typically 4-8 layers), each layer containing a multi-head self-attention mechanism (typically 8-16 attention heads) and a feedforward neural network. The calculation formula for the self-attention mechanism is:

[0149] Attention(Q, K, V) = softmax(QK^T / √d_k) V

[0150] Where Q, K, and V are the query, key, and value matrices, respectively, and d_k is the dimension of the key vector.

[0151] • Feature output: The Transformer encoder outputs a high-dimensional feature representation, which is shared as a low-level feature for the decoding module to use.

[0152] (3) Multi-task decoding module

[0153] This module adopts a multi-task learning architecture, sharing the low-level feature representation output by the Transformer encoder, and simultaneously outputting prediction results through four task-specific decoding heads. The specific implementation is as follows:

[0154] Task 1: Monosaccharide Composition Prediction – Using a Multilayer Perceptron (MLP) as the classification head, output the number (integer between 0 and 20) of each monosaccharide type contained in the O-linked glycan. The loss function used is either cross-entropy loss or mean squared error loss.

[0155] Task 2: Branch Topology Prediction – Using a graph generation decoder, progressively predict the core type (core 1 to core 8), number of branches (0 to 4), and extension pattern of each branch for the glycan. An autoregressive generation method is used, predicting the addition of one node or edge at each step.

[0156] Task 3: Glycosidic Bond Connection Prediction – Using an MLP classification head, predict the connection position (1→2, 1→3, 1→4, 1→6, etc.) and α / β configuration for each predicted edge. Cross-entropy loss is used as the loss function.

[0157] Task 4: O-glycosylation site prediction – Using a sequence-labeled MLP classification head, predict whether each amino acid residue in the peptide sequence has undergone O-glycosylation (binary classification: yes / no). Peptide sequence information is inferred from the precursor ion mass of mass spectrometry data or directly determined from peptide fragment ions (b / y ions) in the MS / MS spectrum.

[0158] Decoding Strategy: During the decoding process, a beam search algorithm is used to generate candidate O-linked glycan structures, with the beam width set to 5~20. Candidate structures are sorted and screened based on fragment ion matching. The matching degree is calculated by comparing the fragment ions generated from the theoretical fragmentation of the candidate structure with the fragment ions in the experimental MS / MS spectra, and calculating the cosine similarity or square root cosine similarity.

[0159] Step S4: Output Results

[0160] The multi-task decoding module outputs complete structural information of the O-linked glycan, specifically including:

[0161] • Monosaccharide composition and sequence: Output the type and quantity of each monosaccharide in IUPAC abbreviation form, for example, "HexNAc(1)Hex(1)NeuAc(1)" represents a polysaccharide composed of 1 GalNAc, 1 Gal and 1 NeuAc.

[0162] • Branching Topology: Represent the branching topology of glycans using a graph or tree structure, with output in an indented text description or graphical representation. For example, the topological representation of core type 1 O-glycans is: GalNAcα1-Ser / Thr ← Galβ1-3 ← NeuAcα2-3.

[0163] • α / β configuration and linkage position of each glycosidic bond: Output each glycosidic bond information in the format of "monosaccharide Aα / βX-Y monosaccharide B", for example, "NeuAcα2-3Gal" means that NeuAc is linked to Gal in the α configuration via a 2→3 linkage.

[0164] • Linkage site of O-linked glycan on peptide: The specific Ser / Thr residue position in the output peptide sequence where O-glycosylation occurs, indicated by amino acid number, such as "Ser-12" or "Thr-8".

[0165] Step S5: Confidence Assessment

[0166] The confidence level of the output of step S4 is evaluated, and the specific implementation method is as follows:

[0167] (1) Calculation of confidence score

[0168] Calculate the confidence scores for each of the four prediction tasks in the multi-task decoding module:

[0169] • Monosaccharide composition confidence: Based on the probability distribution output by softmax, the highest probability value is taken as the confidence level;

[0170] • Branch topology confidence: Based on the cumulative probability of candidate structures in the bundle search, the normalized probability value is taken;

[0171] • Glycosidic bond confidence: the geometric mean of the predicted probabilities of each glycosidic bond;

[0172] • Confidence of glycosylation sites: the geometric mean of the predicted probabilities of each glycosylation site.

[0173] The overall confidence score is the weighted average of the four scores mentioned above.

[0174] (2) Manual verification mechanism

[0175] When the overall confidence score falls below a preset threshold (typically set to 0.7 or 0.85), manual verification mode is activated. The top 35 candidate structures and their corresponding MS / MS spectra are pushed to the user interface for professional review and confirmation via a visual interface. Once the manual review is passed, the review results are fed back into the model training dataset for continuous model optimization.

[0176]

[0177] Example 2: Training Implementation of a Multi-Task Deep Learning Model

[0178] This embodiment provides a training method for a multi-task deep learning model, and the specific implementation is as follows:

[0179] (a) Construction of the training dataset

[0180] (1) Data source

[0181] The training dataset contains no fewer than 100,000 experimentally validated MS / MS spectra of fully annotated O-linked glycan structures. Data sources include:

[0182] • Publicly available glycoproteomics databases (such as GlycoPOST, UniCarb-DB, etc.);

[0183] • O-glycopeptide mass spectrometry datasets from published literature;

[0184] • O-glycopeptide mass spectrometry data accumulated by the inventors in previous research [References 1-3];

[0185] • Manually labeled data verified through experiments.

[0186] (2) Data labeling

[0187] Each MS / MS spectrum should be labeled with the following information:

[0188] • Monosaccharide composition: the types and quantities of each monosaccharide (GalNAc, Gal, GlcNAc, Fuc, NeuAc, NeuGc, etc.);

[0189] • Branch structure: Core type (core 1 to core 8), number of branches, and extension patterns of each branch;

[0190] • Glycosidic bond linkage: the linkage positions of each glycosidic bond (1→2, 1→3, 1→4, 1→6, etc.) and α / β configuration;

[0191] • O-glycosylation site: The specific Ser / Thr location on the peptide where O-glycosylation occurs.

[0192] The annotation work was carried out with the assistance of glycopeptide search engines such as pGlyco3, and was cross-validated by at least two professionals.

[0193] (3) Data partitioning

[0194] The labeled dataset was divided into training, validation, and test sets in an 8:1:1 ratio. The training set did not contain any glycopeptide sequences from the validation and test sets to ensure the objectivity of the model evaluation.

[0195] (II) Model Training Strategy

[0196] (1) Training environment and hyperparameters

[0197] Model training was performed on a workstation equipped with an NVIDIA RTX 3090 or A100 GPU. The training hyperparameters were set as follows:

[0198] • Number of training epochs: 100-200 epochs;

[0199] • Batch size: 128~256;

[0200] Initial learning rate: 1×10⁻ 4 ;

[0201] • Learning rate scheduling strategy: cosine annealing or linear decay;

[0202] • Optimizer: AdamW (weight decay factor set to 1×10⁻) 5 );

[0203] • Gradient clipping: The maximum gradient norm is set to 1.0.

[0204] (2) Joint loss function for multiple tasks

[0205] End-to-end training is performed using a multi-task joint loss function:

[0206] L_total = α·L_composition + β·L_topology + γ·L_linkage + δ·L_site

[0207] in:

[0208] • L_composition is the cross-entropy loss for predicting monosaccharide composition;

[0209] • L_topology is the cross-entropy loss of branch topology prediction (or the log-likelihood loss of graph generation).

[0210] • L_linkage is the cross-entropy loss predicted by the glycosidic bond linkage mode;

[0211] • L_site is the binary cross-entropy loss predicted by O-glycosylation sites;

[0212] • α, β, γ, and δ are the weight coefficients for each task, with initial values ​​set to 0.2, 0.3, 0.3, and 0.2 respectively. These values ​​can be dynamically adjusted during training based on the performance of the validation set.

[0213] (3) Training strategy

[0214] • Pre-training strategy: Given the limited training data for O-glycan proteomics, a pre-training strategy is adopted to alleviate the data shortage problem. First, pre-training is performed on a large-scale N-connected glycan mass spectrometry dataset, and then fine-tuning is performed on the O-connected glycan dataset.

[0215] • Data augmentation: Perform data augmentation operations such as random noise injection and intensity random scaling on the MS / MS spectra in the training set to improve the model's generalization ability.

[0216] • Early stopping mechanism: When the validation set loss no longer decreases within 10 consecutive epochs, training is terminated early to prevent overfitting.

[0217] • Model saving: Saves the best-performing model checkpoints on the validation set for subsequent inference parsing.

[0218] (4) Evaluation indicators

[0219] The model performance was evaluated on the validation and test sets, and the evaluation metrics included:

[0220] • Monosaccharide composition accuracy: The proportion of spectra whose predicted monosaccharide composition is completely consistent with the actual labeled composition;

[0221] • Branch topology accuracy: The proportion of spectra whose predicted branch topology is completely consistent with the actual annotation;

[0222] • Glycosidic bond accuracy: The proportion of spectra in which the predicted glycosidic bond linkage is completely consistent with the actual annotation;

[0223] • Glycosylation site accuracy: The proportion of spectra in which predicted O-glycosylation sites are completely consistent with the actual annotations;

[0224] • Complete Structure Accuracy: The percentage of spectra that are correct for all four prediction tasks.

[0225]

[0226] Example 3: Implementation of the O-linked polysaccharide complete structure analysis system

[0227] This embodiment provides a deep learning-based system for resolving the complete structure of O-connected glycans. The various modules of the system can be implemented using software, hardware, or a combination of both.

[0228] System Composition

[0229] (1) Data acquisition module

[0230] The data acquisition module is used to acquire tandem mass spectrometry data of the O-glycopeptide sample to be analyzed. This module includes a mass spectrometry data import interface, supporting the import of mass spectrometry data in the following formats:

[0231] • Raw mass spectrometry data formats: .raw (Thermo), .d (Bruker), .wiff (SCIEX), etc.;

[0232] • Peak list formats: .mgf (Mascot Generic Format), .mzML, .mzXML, etc.

[0233] This module also supports direct data exchange with liquid chromatography-tandem mass spectrometry (LC-MS / MS) to enable real-time data acquisition.

[0234] (2) Data preprocessing module

[0235] The data preprocessing module is used to perform peak identification, isotope peak removal, noise filtering, and normalization on the mass spectrometry data. The specific implementation of this module is described in step S2. The preprocessed data is stored in the form of a fragment ion feature matrix for subsequent use by the deep learning model.

[0236] (3) Graph Neural Network Encoding Module

[0237] The graph neural network encoding module is used to encode the nonlinear branching structure of O-linked glycans into a graph structure representation. This module adopts the graph neural network architecture described in step S3 of Example 1, encoding the glycan into a directed acyclic graph G=(V, E), where nodes represent monosaccharide units, edges represent glycosidic bond connections, and the attributes of the edges include connection positions and α / β configuration information.

[0238] (4) Transformer Feature Extraction Module

[0239] The Transformer feature extraction module is used to extract long-range dependent features from fragment ion spectra. This module adopts the Transformer encoder architecture described in step S3 of Example 1, and captures the correlation between different fragment ions through a multi-head self-attention mechanism, outputting high-dimensional feature representations.

[0240] (5) Multi-task decoding module

[0241] The multi-task decoding module is used to simultaneously output the monosaccharide composition sequence, branching topology, glycosidic bond linkage, and O-glycosylation sites of O-linked glycans. This module adopts the multi-task decoding architecture described in step S3 of Example 1, shares the low-level feature representation output by the Transformer encoder, and outputs prediction results in parallel through four task-specific decoding heads.

[0242] (6) Confidence assessment module

[0243] The confidence assessment module is used to evaluate the confidence level of the output results. This module calculates the confidence score for each prediction task and the overall confidence score, and triggers a manual verification process when the confidence level is lower than a preset threshold.

[0244] (7) Output module

[0245] The output module outputs the complete structural information of the O-linked glycan and its corresponding confidence score. Supported output formats:

[0246] • Text format: Text output is described using IUPAC abbreviations and structures;

[0247] • Graphical format: Visualize the glycan structure in the form of a tree diagram or network diagram;

[0248] • Structured data format: Output in JSON or XML format for easy integration with other bioinformatics tools.

[0249] System integration and deployment

[0250] The above modules employ an end-to-end joint training approach, sharing the underlying feature representations. System deployment methods include:

[0251] • Local deployment: The complete system is deployed on the user's local computing server, which is suitable for scenarios with high data confidentiality requirements;

[0252] • Cloud deployment: The system is deployed on a cloud computing platform, and users can submit mass spectrometry data and obtain analysis results through a web interface or API.

[0253] • Containerized deployment: The system and its dependent environments are packaged using Docker containers to ensure cross-platform portability and repeatability.

[0254] The system provides a graphical user interface (GUI), through which users can complete the entire process of data import, parameter setting, parsing task submission, result viewing, and manual verification.

[0255]

[0256] Example 4: Application Scenario

[0257] The following example illustrates the effectiveness of this invention through a specific application scenario.

[0258] Sample: O-glycopeptide sample extracted from human serum.

[0259] Mass spectrometry data acquisition: MS / MS spectral data were acquired using a nanoliter liquid chromatography-tandem mass spectrometer in data-dependent mode, following the method in step S1 of Example 1.

[0260] Data preprocessing: Following the method in step S2 of Example 1, the collected mass spectrometry data were processed by peak identification, isotope peak removal, noise filtering, and normalization to construct a fragment ion feature matrix.

[0261] Model Analysis: The preprocessed fragment ion feature matrix is ​​input into a pre-trained multi-task deep learning model. After graph neural network encoding, Transformer feature extraction, and multi-task decoding, the model outputs candidate O-connected glycan structures.

[0262] Results Output: The complete structural information output by the model includes:

[0263] • Monosaccharide composition: GalNAc(1)Gal(1)NeuAc(1)

[0264] • Branch topology: Core 1 type, no branch extension

[0265] • Glycosidic bond linkage: GalNAcα1-Ser / Thr, Galβ1-3GalNAc, NeuAcα2-3Gal

[0266] • O-glycosylation site: Ser-4 in the peptide sequence “TVPSSSTAA”

[0267] Confidence assessment: The overall confidence score is 0.92, which is higher than the preset threshold of 0.85, so the final analysis result is output directly.

[0268] Verification of analysis results: Through manual review and database comparison, the analysis results are consistent with the actual structure, proving the accuracy and reliability of the method of the present invention.

[0269]

[0270] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for resolving the complete structure of O-connected glycans based on deep learning, characterized in that, Includes the following steps: S1. Acquire tandem mass spectrometry data of the O-glycopeptide sample to be analyzed, wherein the mass spectrometry data is MS / MS spectrum data obtained in data-dependent mode acquisition mode; S2. Preprocess the MS / MS spectral data, including: peak identification, isotope peak removal, noise filtering and normalization, extract the mass-to-charge ratio and intensity information of each fragment ion, and construct a fragment ion feature matrix; S3. Input the preprocessed fragment ion feature matrix into a pre-trained multi-task deep learning model, wherein the multi-task deep learning model comprises: The graph neural network encoding module is used to encode the nonlinear branching structure of O-connected glycans into a graph structure representation; The Transformer feature extraction module is used to extract long-range dependent features from fragment ion spectra. The multi-task decoding module is used to simultaneously output the monosaccharide composition sequence, branching topology, glycosidic bond linkage mode, and O-glycosylation site of O-linked glycans. S4. The multi-task decoding module outputs complete structural information of the O-linked glycan, including: monosaccharide composition and sequence, branch topology, α / β configuration and linkage position of each glycosidic bond, and linkage sites of the O-linked glycan on peptide segments; S5. Evaluate the confidence level of the output results of the multi-task decoding module and output the confidence score of each structure prediction.

2. The method according to claim 1, characterized in that, In step S3, the graph neural network encoding module encodes O-linked glycans into a directed acyclic graph structure, where nodes represent monosaccharide units, edges represent glycosidic bond connections, and the attributes of the edges include the connection positions of the glycosidic bonds and α / β configuration information.

3. The method according to claim 1, characterized in that, The training dataset of the multi-task deep learning model contains no less than 100,000 experimentally validated and fully annotated MS / MS spectra of O-linked glycan structures. The annotation information includes monosaccharide composition, branching structure, glycosidic bond linkage mode, and O-glycosylation sites.

4. The method according to claim 1, characterized in that, In step S3, the multi-task decoding module uses a cluster search algorithm to generate candidate O-linked glycan structures and sorts and filters the candidate structures based on fragment ion matching degree.

5. The method according to claim 1, characterized in that, The method further includes: when the confidence score output by the multi-task deep learning model is lower than a preset threshold, a manual verification mode is activated, and the candidate structure and the corresponding MS / MS spectrogram are pushed to the user interface for manual review and confirmation.

6. A deep learning-based system for resolving the complete structure of O-connected glycans, characterized in that, include: The data acquisition module is used to acquire tandem mass spectrometry data of the O-glycopeptide sample to be analyzed; The data preprocessing module is used to perform peak identification, isotope peak removal, noise filtering, and normalization on the mass spectrometry data. The graph neural network encoding module is used to encode the nonlinear branching structure of O-connected glycans into a graph structure representation; The Transformer feature extraction module is used to extract long-range dependent features from fragment ion spectra. The multi-task decoding module is used to simultaneously output the monosaccharide composition sequence, branching topology, glycosidic bond linkage mode, and O-glycosylation site of O-linked glycans. The confidence assessment module is used to assess the confidence level of the output results. The output module is used to output the complete structural information of O-linked glycans and their corresponding confidence scores.

7. The system according to claim 6, characterized in that, The graph neural network encoding module, Transformer feature extraction module, and multi-task decoding module adopt an end-to-end joint training approach, sharing the underlying feature representation.

Images