1 Introduction
According to International Union of Pure and Applied Chemistry (IUPAC) nomenclature, polysaccharides (glycans) are macromolecules consisting of a large number of monosaccharide residues linked by glycosidic bonds, with the term “polysaccharide” commonly applied to chains containing more than ten monosaccharide residues. However, “glycan” is also used for the saccharide component of glycoproteins regardless of chain length. Therefore, both terms are retained in this review: “polysaccharide” primarily refers to independent carbohydrate macromolecules, whereas “glycan” is used for glycoconjugate-associated carbohydrate chains and glycan-specific datasets or studies. These structurally diverse carbohydrates participate in a wide range of biological processes, including cell adhesion, immune regulation, molecular recognition, host–microbe interactions, and intracellular signaling [
1,
2]. Beyond their essential biological functions, glycans possess remarkable physicochemical characteristics, such as biocompatibility, biodegradability, gel-forming capacity, rheological tunability, and molecular self-assembly behavior, which have led to their widespread applications in food systems, biomedical engineering, pharmaceutical delivery, tissue scaffolding, intelligent packaging, and advanced biomaterials [
3,
4]. In food science, polysaccharides are extensively utilized as stabilizers, emulsifiers, thickeners, edible coatings, and fat replacers, while numerous natural glycans exhibit antioxidant, immunomodulatory, hypoglycemic, hypolipidemic and prebiotic activities, highlighting their substantial potential for functional foods and precision nutrition [
4,
5].
Despite their significances, glycoscience has progressed more slowly than genomics and proteomics because glycans are synthesized through non-template driven pathways, resulting in extensive branching, linkage variability, stereochemical diversity, and structural heterogeneity [
1,
6]. These characteristics make glycan structural elucidation and structure-function analysis extremely challenging. Conventional analytical techniques, including nuclear magnetic resonance (NMR), mass spectrometry (MS), chromatography, and crystallographic approaches, provide valuable structural information but remain limited in resolving highly complex glycan structure and dynamic conformations [
1,
2,
7]. Furthermore, the rapidly increasing volume and complexity of glycomics data have exceeded the capabilities of traditional computational and statistical methods.
Recent advances in artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL), are transforming glycoscience by enabling efficient analysis of high-dimensional and heterogeneous datasets (Fig. 1). AI particularly ML and DL are rapidly transforming in food science and biomacromolecular research. AI-driven approaches have demonstrated remarkable capability in handling high-dimensional, nonlinear and heterogeneous datasets generated from spectroscopy, imaging, omics and molecular characterization technologies. In recent years, ML algorithms have been increasingly applied in food processing, quality evaluation, biomarker discovery, structural elucidation and intelligent material design. For example, AI-driven approaches have demonstrated strong potential in glycan extraction optimization, sequence elucidation, structural prediction, glycan-protein interaction analysis, biomarker discovery and intelligent design of glycan-based materials [
8–
11]. In particular, deep learning models combined with MS, NMR nanopore sensing, and glycomics technologies have significantly improved the accuracy and throughput of glycan characterization and functional prediction [
10,
11]. However, compared with protein and other biomacromolecules, AI-assisted glycan research remains relatively underdeveloped and insufficiently summarized.
As shown in Fig. 1A, current AI-related research in food science is mainly concentrated in areas such as classification, food quality, computer vision, fermentation, and chemometrics, demonstrating the broad integration of AI technologies into food and biomolecular science. Fig. 1B further illustrates the exponential increase in AI and ML related publications during the past decade, reflecting the rapidly growing scientific and industrial interest in intelligent data-driven technologies. Nevertheless, Fig. 1C indicates that existing review studies remain largely focused on proteins, food quality detection and general food engineering applications whereas comprehensive reviews dedicated to AI-assisted glycan research are still limited. This research gap highlights the urgent need for a systematic overview of AI-driven strategies in glycoscience, particularly in the areas of glycan extraction, structural elucidation, molecular interaction prediction and functional material design. Therefore, this review provides a comprehensive overview of recent advances in AI-assisted glycan research. Particular emphasis is placed on AI-driven strategies for glycan extraction, structural elucidation, three-dimensional (3D) structure prediction, glycan-protein interaction analysis, biomarker discovery and functional material design. In addition, current challenges and future perspectives for integrating AI with glycoscience are discussed to promote the development of predictive, precision and mechanism-guided glycan research in food science, biomedicine, and advanced materials.
2 Challenges in the Polysaccharide Field
2.1 Limitations of molecular-level characterization methods
As one of the three major classes of biological macromolecules, glycans remain much less understood than proteins and nucleic acids in terms of chemical structure, sequence composition, 3D conformation, and intermolecular interactions. This gap mainly arises from their extraordinary structural complexity and heterogeneity, together with the limitations of current experimental and computational methods (Table 1 and Fig. 2). As a result, the slow progress in glycan structural elucidation has constrained the rational de novo design of glycans and hindered deeper understanding of their biological functions and applications.
The sequence architecture of glycans is highly complex. Monosaccharides can form multiple linkage isomers through glycosidic bonds, while the stereochemistry of the anomeric carbon generates α- and β-anomers. In addition, derivatization further expands structural diversity [
1]. Although spectroscopic and chromatographic methods provide information on functional groups and molecular weight, they are insufficient for resolving polysaccharide sequences [
1,
5]. Therefore, fine-structure elucidation usually relies on the combined use of methylation analysis, selective cleavage methods, NMR, and MS. Methylation analysis and selective cleavage can reveal monosaccharide composition, linkage positions, branching features, and polymerization degree [
1]. NMR provides key information on monosaccharide type, substituents, and anomeric configuration, while two-dimensional NMR supports signal assignment, linkage identification, sequence analysis, and isomer differentiation. However, severe signal overlap in highly branched glycans often limits its resolving power [
2]. MS platforms, including MALDI–MS and LC–MS/MS, are powerful tools for determining molecular weight, monosaccharide composition, partial sequences, and certain linkage features, yet conventional MS remains limited in distinguishing structural isomers with identical
m/z values [
7,
9,
10]. Although IM–MS improves isomer separation, its application is constrained by finite peak capacity and limited accessibility [
9]. Nanopore sensing has emerged as a promising strategy for single-molecule glycan sequencing, but its practical use is still restricted by low sensitivity, insufficient isomeric resolution, and challenges related to translocation kinetics and sample purity [
11].
Glycans also exhibit highly flexible 3D structures, with low energy barriers between conformations and rapid interconversion among conformers. AFM and SEM provide morphological information at the nano- and microscale, but higher-resolution techniques are required for atomic-level insight [
12]. Current approaches for glycan 3D structural analysis include crystallography, NMR, cryo-electron microscopy (cryo-EM), electrospray ionization scanning tunneling microscopy (ESI-STM), and molecular dynamics (MD) simulations. Crystallography has resolved lectin-glycan complexes, glycopeptides, and polysaccharide crystals, but it often fails to preserve solution-state features and is limited by crystal heterogeneity and resolution [
13]. NMR remains a major tool for glycan conformational analysis, although signal overlap is still problematic for oligosaccharides and polysaccharides with repeating units [
14]. ESI-STM enables single-molecule visualization without ensemble averaging [
12], but its dependence on surface states restricts the detection of low-abundance solution conformations. Cryo-EM has revolutionized structural biology [
15], yet glycan analysis remains challenging because rapid vitrification captures multiple conformations of flexible sugar chains and associated protein regions, complicating image alignment and limiting glycan resolution [
16,
17]. Computational chemistry, including quantum chemical calculations and molecular mechanics methods such as GLYCAM06, complements experimental techniques [
18]. For example, conformational libraries generated by MD simulations can be integrated with cryo-EM data to accelerate glycoprotein structure modeling [
16]. However, the computational cost rises sharply with atom number and polymerization degree, limiting the large-scale application of these methods in glycoscience.
In summary, the substantial diversity of polysaccharide sequences and three-dimensional conformations imposes inherent limitations on current methods for sequence determination and structural elucidation. Even when multiple complementary analytical techniques are combined, complete and unambiguous characterization of polysaccharide structures remains difficult. Moreover, experimental errors and differences in sample preparation, instrumental conditions, data-processing procedures, and analytical standards reduce the comparability and reproducibility of results across studies, thereby limiting data integration and the establishment of reliable molecular-level structural models. Inspired by the transformative success of AlphaFold in protein structure prediction [
8], ML has shown strong potential in the study of bioactive macromolecules. Therefore, introducing ML-based strategies into glycan structural elucidation is urgently needed to overcome the limitations of conventional methods and promote advances in polysaccharide science and its applications.
2.2 Efficiency constraints
Due to the lack of molecular-level theoretical support, traditional research methods face massive combinatorial search spaces generated during hypothesis formation and validation. This not only dramatically increases time and economic costs but also falls far short in modeling capabilities for high-dimensional non-linear systems, rendering the entire process beyond the effective scope of conventional analytical methods. This dilemma is particularly prominent in polysaccharide research (Fig. 2). On one hand, polysaccharide extraction processes involve core parameters such as solid–liquid ratio, temperature, pH, and extraction time. The introduction of various auxiliary technologies—such as enzymatic digestion, acid-base treatments, ultrasound, supercritical CO
2 extraction, deep eutectic solvents, ionic liquids, and three-phase solvent systems [
19–
26]—further exponentially increases the dimensionality and complexity of the parameters affecting extraction efficiency. The preparation of glycan-based materials requires balancing multiple factors, including principle selection, component ratios, and processing conditions, while simultaneously meeting diverse optimization objectives such as mechanical properties, functional characteristics, and biological activities [
3,
27–
29]. Traditional response surface methodology generally identifies local optima within a restricted experimental region rather than guaranteeing global optimality. Moreover, as the number of experimental factors increases, the required number of design points grows substantially, making global optimization increasingly challenging [
26]. On the other hand, traditional quality testing of food and agricultural products relies heavily on chemical analysis. These methods universally suffer from tedious sample preparation, destructive testing processes, high reliance on professional operators, expensive instruments and consumables, and low testing efficiency. They are also susceptible to environmental interference and sample contamination, leading to poor result stability. Furthermore, they can only acquire single-point holistic data for a sample, failing to visually characterize physical features, spatial distribution, and local defects, making them unsuitable for the high-throughput, rapid quality testing demands of the food industry [
30]. Furthermore, the evaluation of glycan bioactivity still depends largely on cell- and animal-based experiments. In the absence of critical molecular-level information, precise activity characterization and mechanistic interpretation remain difficult to achieve using conventional approaches. In addition, glycan biomarker discovery is often challenged by the massive volume, high dimensionality, and structural heterogeneity of glycomics data, which are difficult to analyze efficiently using traditional methods. Given these limitations, ML offers a powerful alternative because of its strong capability in high-dimensional data analysis and nonlinear relationship modeling.
3 Model Training
3.1 Data preparation
The foundational step in constructing ML models is the assembly and curation of datasets for model training. The data types covered in this review encompass component ratios, process parameters, spectral data, imaging data, glycan structural data, macroscopic physicochemical properties, and biological activity data. The second step in model building is typically data preprocessing, which is crucial for enhancing model quality and reliability; this primarily includes data cleaning, feature selection, feature scaling, and dataset partitioning. Currently, datasets used for model training are mostly derived from experimental measurements. However, due to inconsistencies in experimental methods, testing platforms, and data standards, datasets broadly suffer from significant heterogeneity, chaotic formatting, small scale, and inconsistent annotations. Consequently, this paper only summarizes and compiles large-scale datasets that have undergone standardized curation (Table 2).
3.2 ML algorithms
ML can generally be categorized into supervised, unsupervised, and semi-supervised or reinforcement learning. Supervised learning uses labeled data to build predictive models, whereas unsupervised learning aims to uncover latent patterns from unlabeled data. As shown in Table 3, the algorithms discussed in this review are broadly classified into traditional ML algorithms, DL algorithms, and optimization algorithms. Traditional ML includes Random Forest (RF), Support Vector Regression (SVR), Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), Extreme Gradient Boosting (XGBoost), Partial Least Squares Regression (PLSR), and k-Nearest Neighbors (KNN). DL includes Artificial Neural Networks (ANN), Probabilistic Neural Networks (PNN), Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNN), Graph Convolutional Networks (GCN), Transformers, Dilated Residual Networks (DRN), and Convolutional Neural Networks (CNN). Optimization algorithms include Genetic Algorithm (GA), Non-dominated Sorting Genetic Algorithm II (NSGA-II), Particle Swarm Optimization (PSO), Differential Evolution (DE), Starfish Optimization Algorithm (SFOA), Whale Optimization Algorithm (WOA), and Sparrow Search Algorithm (SSA).
3.3 Evaluation
Before applying models to predictive tasks, standard evaluation protocols must guide model optimization. Hold-out validation, cross-validation, and bootstrapping are mainstream evaluation schemes. For classification models, multidimensional metrics are typically used to assess performance comprehensively, including accuracy, recall, specificity, precision, F1-score, receiver operating characteristic (ROC) curve, and area under the curve (AUC). The ROC curve and AUC primarily measure the overall discriminative ability of the model; an ROC curve closer to the top-left corner and an AUC value closer to 1 indicate superior overall performance. For regression models, common evaluation metrics include mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), ratio of performance to deviation (RPD) and the coefficient of determination (R2).
4 From Preparation and Structural Elucidation to Rational Reverse Design
This section summarizes the current research status of ML in the investigation of polysaccharide molecular structures and highlights the novel paradigm enabled by ML for this research field. As shown in Fig. 3, the review covers polysaccharide extraction and synthesis, sequence analysis, 3D structures prediction, interaction prediction and de novo design.
4.1 ML-optimized glycan extraction and synthesis
4.1.1 Extraction
Polysaccharides are generally classified into starch and non-starch polysaccharides, the latter of which can be further divided into microbial and non-microbial types. ML algorithms can optimize the extraction processes of all these polysaccharides and enable single- or multi-objective process control guided by target requirements such as glycan structure and biological activity.
In process optimization, extraction yield is usually the primary goal. Researchers typically begin by identifying key variables affecting extraction efficiency using methods such as Plackett–Burman (PB) design, gray relational analysis (GRA), or Least Absolute Shrinkage and Selection Operator (LASSO) regression [
22,
63,
64]. Based on response surface designs, experimental data are then used to establish predictive models linking process parameters to extraction performance, with algorithms such as RF, XGBoost, ANN, and SVR widely applied [
22,
64,
65]. Global optimization methods, particularly genetic algorithms (GA), are subsequently employed to determine the optimal conditions for maximum yield [
26]. Through this workflow, the extraction yields of both starch and non-starch polysaccharides can be significantly improved [
22,
66–
68]. ML has also been applied to optimize green extraction technologies, including enzyme-assisted extraction [
22], ultrasound-assisted extraction [
69], supercritical CO
2 extraction [
19], deep eutectic solvent extraction [
70], ionic liquid extraction [
24], and three-phase partitioning extraction [
26].
However, the optimal conditions identified by these models are often specific to the raw material and experimental batch used for model development. Natural feedstocks can vary substantially with geographical origin, harvest period, cultivar, storage history, and batch, which may change their composition and extraction response. Therefore, conditions optimized from a single batch should not be assumed to transfer directly to practical applications; external validation across independent sources, harvest periods, and batches is needed to assess robustness and generalizability.
Given that extraction methods impact polysaccharide structure, and that structural integrity and biological activity are widespread concerns, ML can also optimize processes and screen extraction methods based on structural requirements and activity orientation. For instance, using parameters like the polygalacturonase/substrate ratio, cellulase/polygalacturonase activity ratio, and reaction time as inputs, single-output prediction models have been constructed for 10 indicators (e.g., solid yield of sugar beet pulp enzymatic hydrolysates, polysaccharide-to-mono/oligosaccharide conversion efficiency), enabling structure-oriented precise process optimization [
71]. Other studies have used structural descriptors, such as molecular weight, monosaccharide composition, and branching degree, to distinguish different extraction methods and support intelligent method selection [
72]. In addition, neural network models linking process parameters with antioxidant activity and extraction yield have enabled activity-guided targeted optimization [
20].
Notably, polysaccharide extraction is inherently a multi-objective optimization problem requiring the simultaneous consideration of purity, structure, function, bioactivity, production cost, and extraction efficiency. Unlike single-objective optimization, such tasks are typically addressed through Pareto-optimal solution sets, which can be effectively optimized using algorithms such as NSGA-II. For instance, multi-output models combined with NSGA-II have been used to simultaneously improve total sugar content, dry extract yield, and the yields of specific active compounds [
63]. Overall, ML enables requirement-driven process control, intelligent extraction method selection, and efficient multi-objective optimization in polysaccharide extraction. With continued advances in glycomics, future studies may achieve more precise optimization targeting glycan sequence features and ultimately establish more generalizable models for diverse polysaccharide sources.
4.1.2 Synthesis
The chemical synthesis of polysaccharides is a critical foundation for conformational analysis and the development of carbohydrate-based functional materials, serving as a key step in programmable glycan assembly. Although automated glycan assembly technology has successfully synthesized a 1080-mer glycan (the longest homogeneous polysaccharide reported to date), precisely controlling the anomeric carbon stereochemistry during glycosylation remains a core challenge hindering efficient synthesis [
12].
Recently, the application of ML in reaction prediction and enzyme design has offered novel strategies to overcome the anomeric stereoselectivity challenge. Researchers screened and integrated 10,434 single-step glycosylation reactions from the CAS database (2010–2015) to build GlycoPredictor the largest, top-performing, and only publicly accessible ML prediction model for glycosylation reactions globally. This model accurately predicts the dominant anomeric product and its ratio based on key parameters like glycosyl donor, acceptor, solvent, activator, and reaction temperature [
73]. Concurrently, the rapid advancement of ML in enzyme design [
74] opens new pathways for the efficient and precise regulation of anomeric selectivity.
4.2 ML-driven glycan sequence identification
As previously noted, glycan sequence structures are highly complex, encompassing monosaccharide composition, anomeric configurations, branching sites, and chain sequences, usually requiring a combination of MS, NMR, and methylation analysis for complete sequencing. To achieve efficient and accurate polysaccharide sequencing, researchers have adopted end-to-end strategies, utilizing ML algorithms to intelligently identify and elucidate glycan sequences based on NMR data, nanopore current signals, and MS data.
4.2.1 Based on NMR and Nanopores
ML-assisted NMR data analysis has realized the sequence identification of di- and tri-saccharides [
75]. However, NMR-based structural elucidation of polysaccharides continues to face challenges: severe signal overlap (especially in high-field regions) and spectral complexity caused by conformational flexibility. The analysis of larger glycan structures remains limited; nevertheless, considering the learning capabilities of LSTM and RNN models on multi-dimensional NMR spectra, breakthroughs are anticipated [
4].
Nanopores can reversibly bind with glycan molecules, generating specific, stable, and distinguishable characteristic ionic current signals. However, the current signals produced as glycans traverse nanopores are typically weak, complex, and highly overlapping, making manual discrimination exceptionally difficult. ML technology can automatically extract key features such as current blockade amplitude, dwell time, and standard deviation, enabling high-resolution identification at the monosaccharide level [
11]. For example, ML can effectively distinguish maltose oligosaccharides with degrees of polymerization ranging from 1 to 8 [
76], as well as glycosaminoglycan disaccharides with varying monosaccharide compositions [
77]. Based on the three major nanopore glycan sequencing strategies currently proposed (strand sequencing, sequential hydrolysis sequencing, and assembly sequencing) [
11], nanopore technology is poised to achieve high-precision structural elucidation and sequencing of large, complex glycans.
4.2.2 Based on MS
Early glycan sequencing methods relied on searching and comparing MS data against databases, which was restricted by missing structural signature ions and the inability to handle structural heterogeneity [
78]. Recently, MS data-driven ML prediction algorithms aim to correct high-throughput glycan sequencing. ML methods have massive data requirements; inevitably, related reports are skewed toward the most abundantly generated glycan data, which are often N-/O-linked glycans and those associated with glycan-binding proteins.
Current ML-based glycan sequencing research mainly follows two directions. The first predicts MS data from glycan structures. DeepGlyco uses glycan structure encoding, tree-LSTM feature extraction, fragmentation-site attention, and ion intensity prediction to achieve highly accurate MS peak prediction, with spectral angle loss below 0.11 and dot product above 0.98, while also distinguishing structural isomers [
79]. DeepGP, a hybrid framework combining Transformers and GCNs, predicts intact N-glycopeptide MS/MS spectra and retention times, improves glycopeptide identification sensitivity, and enables structure identification even without diagnostic ions, with performance exceeding 90% across multiple datasets [
78]. Nevertheless, both methods still have limited applicability, with DeepGlyco restricted to glycans composed of five specific monosaccharides [
79], while DeepGP is designed only for N-glycopeptides and does not account for the influence of anomeric isomers on MS behavior [
78].
The second direction infers glycan structures directly from MS data. GlycanFinder performs
de novo sequencing of N-linked glycans by combining dynamic programming, Graphormer-based GNNs, and ion-matching neural networks, but its reported accuracy remains relatively low at 32% [
80]. In contrast, CandyCrunch represents a major advance. Built on the largest LC–MS/MS dataset for eukaryotic glycans reported to date, containing 489,103 validated spectra, CandyCrunch integrates a dilated residual neural network with biosynthetic priors and GNN-based topology modeling to enable fragment-ion assignment, sequence inference, and biosynthetic pathway deduction in a unified framework. It can rapidly predict and classify 3391 known glycans from raw LC–MS/MS data, achieving an overall structural recognition accuracy of 90.3% (F1-score = 0.897) [
10].
Earlier ML-assisted glycan sequencing studies were generally unable to resolve anomeric configuration, which limited glycan classification, monosaccharide discrimination, and overall sequencing accuracy [
78,
80]. CandyCrunch addressed several of these limitations by achieving anomeric configuration prediction, expanding the scale of training data, and introducing biosynthetic network constraints to reduce biologically implausible false positives [
10]. Despite these advances, high-throughput and high-precision prediction of glycans from a broader range of biological sources still requires further investigation.
4.3 Three-dimensional structure
The 3D structure of biomolecules fundamentally determines their physicochemical properties and biological functions, which ultimately arise from specific intermolecular interactions. Owing to the limitations of traditional analytical techniques, progress in glycan 3D structural elucidation and molecular interaction studies has lagged far behind that of proteins and nucleic acids, despite glycans being among the most widely distributed biological macromolecules in nature [
12]. ML-based structural prediction methods, exemplified by AlphaFold3, have greatly advanced protein structure determination, functional annotation, and interaction analysis [
8]. With the accumulation of crystallographic data and the continued development of MD simulations, glycan research is now entering a new stage in which ML offers promising opportunities for 3D conformation prediction, molecular recognition analysis, and structure-function investigation.
Recent ML models for glycan 3D structure prediction are largely built on MD-derived conformational databases such as GlycoShape [
81]. For glycans with low degrees of polymerization, all-atom MD can effectively characterize structural and dynamic features because their glycosidic bonds and conformational degrees of freedom are relatively manageable [
82,
83]. GlycoShape has established a database of 534 glycans based on a cumulative 1 ms of MD simulations, with representative conformations identified by clustering analysis [
84]. Based on GlycoShape, Luc Thomès et al. [
85] developed a deep learning framework that takes glycan sequences as input and combines Graph Isomorphism Networks with von Mises mixture distributions to model the multimodal distributions of glycosidic torsion angles. The model predicts torsion-angle distributions, monosaccharide-level solvent-accessible surface area, and flexibility values directly from sequence. For previously unseen glycans, it achieved prediction errors of only φ ± 7°, ψ ± 9°, and ω ± 10°, with a Pearson correlation coefficient of 0.94 for solvent-accessible surface area. This work represents the first deep learning approach capable of predicting glycan 3D structural properties directly from sequence alone. Overall, ML-based glycan 3D structure prediction has progressed rapidly in recent years, but accurate prediction for large glycans and polysaccharides remains challenging. Given the high conformational similarity among repeating units in polymeric glycans [
14], inferring global conformations from representative repeating-unit structures may provide a promising route to overcome this limitation.
4.4 Interactions
Current studies of glycan interactions with other substances are focused mainly on glycan-protein interactions, with the major tasks being binding affinity prediction and binding site prediction. For binding affinity prediction, most models use glycan sequence information as the core input to quantitatively estimate glycan-protein binding strength. A representative example is a dual-encoder framework that combines LSTM-based viral protein sequence features with SweetNet-derived glycan graph features to predict binding affinity
Z-scores, achieving an optimal MSE of 0.735 on 120,000 interaction records and successfully identifying glycan receptors for influenza viruses, coronaviruses, and rotaviruses [
86]. GlyNet encodes glycan sequences into 272-dimensional q-gram features to predict continuous RFU values for 1257 protein samples, showing strong performance (MSE = 0.12,
R2 = 0.78, AUC = 0.912) and enabling extrapolation to 4160 novel glycans [
87]. However, its predictive scope is restricted to a fixed set of proteins. More recent studies have shown that model performance depends strongly on glycan training set coverage, molecular representation, and structural feature incorporation. MCNet, for example, uses chirality-aware Morgan fingerprints and atomic q-grams, and is trained on both Consortium for Functional Glycomics glycan array data and glycomimetic datasets [
88]. It achieved the first successful modeling of cross-chiral glycan recognition and experimentally confirmed the binding of L-glucose to fucose-binding lectins. In addition, incorporating 3D descriptors, including solvent-accessible surface area, flexibility, and torsion angles, into SweetNet improved affinity prediction accuracy by 7.4% [
85]. These advances indicate that glycan-protein affinity prediction has evolved from sequence-based modeling toward integrated frameworks that combine sequence, 3D structure, and chirality-related information.
Binding site prediction, in contrast, relies largely on crystallographic data. With the expansion of carbohydrate-protein interaction datasets, this field has progressed from traditional machine learning methods trained on relatively small datasets [
89,
90] to high-precision deep learning models supported by large-scale structural data. DeepGlycanSite is a representative example, integrating deep equivariant graph neural networks with Transformers to predict glycan-binding pockets and specific glycan-protein binding sites. In this framework, both proteins and glycans are encoded as atomic-level molecular graphs, enabling structural feature extraction and binding site prediction in a unified architecture [
91]. Its performance surpasses that of existing tools such as GlycoTorch Vina, AutoDock Vina, DiffDock, and EquiBind. GlycanInsight further extends this framework into a web platform that supports both experimental and AlphaFold2-predicted protein structures, as well as multiple glycan ligand formats [
92].
4.5 De novo design guided by functional properties
Polysaccharides are widely used as functional matrices in food, biomedicine, and biomaterials, and their macroscopic physicochemical properties, including rheological behavior, gel strength, solubility, and texture, together with their biological activities, directly determine the performance, application value, and market acceptance of final products. Accordingly, the design of polysaccharides and polysaccharide-based composite systems guided by target properties and activities has become an important strategy for precise and functional development. In this review, AI-assisted de novo design encompasses two complementary directions: property-to-structure design and reaction-condition-to-property optimization. The latter is included because predictive models can be coupled with optimization algorithms, such as GA, to identify conditions that achieve desired properties. It should also be noted that, because glycan-specific studies remain limited, selected studies from other fields are included as methodological references for future glycan and polysaccharide design.
4.5.1 Property-to-structure design
For property-to-structure prediction, the structure of the target material must be represented as the model output. Polysaccharide structures can be encoded using molecular fingerprints, topological indices, and molecular graph- or string-based representations [
84]. For instance, carbohydrate structures encoded as strings have been used for immunogenicity prediction with accuracies above 92% [
93]. Glycan molecular graphs have also been employed to predict binding strength between glycans and viral proteins, demonstrating potential for glycan-based antiviral design [
86]. In parallel, molecular fingerprint-based approaches have been shown to effectively predict the binding strength between L-conFig.d carbohydrates and proteins [
88]. As previously demonstrated, researchers have shown that incorporating stereochemical and spatial structural information can significantly improve predictive accuracy. In addition, topological descriptors derived from molecular graphs can capture spatial structural information of molecules and have been shown to achieve strong predictive performance [
94]. Notably, the proposed topological descriptor framework enables the use of polymerization degree as an input feature to generate corresponding quantitative outputs for property prediction. Although amylose is composed of relatively simple linear glucose units linked via α-1,4-glycosidic bonds, this study provides an important methodological insight: partial structural descriptors can be effectively mapped to macroscopic physicochemical properties, offering a generalizable strategy for predicting glycan-related functions from incomplete or abstracted molecular structural information.
These structural representations provide a foundation for property-to-structure design by enabling candidate structures to be represented, generated, and evaluated in a machine-readable space. Building on these representations, the next step is to establish a conceptual framework for AI-assisted property-to-structure design. From a statistical perspective, this process can be theoretically formulated as sampling structures or sequences that satisfy predefined target properties from a conditional distribution,
x~p(
). In this framework, a general generative model, p(
x), learns the prior distribution of plausible structures, sequences, or atomic configurations, whereas a property predictor, p(
), estimates the likelihood that a candidate exhibits desired properties, such as stability, activity, or binding affinity. These two components can be combined through Bayes’ rule, p(
)∝p(
)p(
x), thereby enabling inverse design from desired properties to candidate structures or sequences. Depending on how property information is incorporated, AI-assisted property-to-structure design can generally be implemented through three strategies: baked-in conditioning, in which target properties are explicitly incorporated during model training; adaptive conditioning, which combines a general generative model with an independent property predictor; and on-the-fly guidance, in which property information dynamically guides the sampling process during generation [
22]. Thus, rather than simply identifying a single candidate with the highest predicted score, property-to-structure design aims to learn a feasible design space and guide the generative distribution toward regions associated with the desired properties, thereby producing multiple candidate structures or sequences that satisfy the predefined design criteria.
4.5.2 Reaction-condition-to-property optimization
For reaction- or formulation-condition-to-property optimization, a study established an iterative ML-assisted framework for the inverse design of underwater adhesive hydrogels. Based on an initial dataset of 180 bioinspired hydrogels, the relative proportions of six functional monomers were used as input variables and underwater adhesive strength as the target property. Gaussian process and random forest regression models were first trained to predict adhesive strength, and were subsequently coupled with batched sequential model-based optimization to propose new monomer compositions with improved performance. Through three iterative rounds of ML prediction, experimental synthesis, property measurement, and model updating, the dataset was expanded to 341 formulations and the optimal composition was progressively identified, ultimately yielding a hydrogel with an underwater adhesive strength exceeding 1 MPa [
95]. With advancements in glycan chemistry, the precise design of glycan molecules has gradually become feasible [
12]. Therefore, integrating glycan chemical synthesis with ML algorithms to design product macroscopic properties and biological activities is emerging as a core paradigm for realizing the
de novo design of glycans and glycan-based systems.
5 Applications of ML-Driven Polysaccharide Research
Polysaccharides serve as one of the essential material carriers of biological information and have been widely applied in the food industry. This section reviews the applications of ML in biomarker and bioactivity screening, quantitative identification and traceability of polysaccharides in agricultural products and processed foods, as well as material fabrication as illustrated in Fig. 4.
5.1 Biomarker discovery and biological activity screening
5.1.1 Biomarker discovery
The human glycome encompasses N-, O-, and C-linked glycans, glycosphingolipids, proteoglycans, glycosaminoglycans, and exogenous glycans. Under pathological conditions, aberrant glycan biosynthesis and glycosylation pathways serve as direct indicators of disease states [
96]. Recently, the synergy between ML and glycomics has facilitated the identification of multivariate glycan biomarkers, overcoming the limitations of traditional single-molecule markers. Adhering to the biomarkers, endpoint, and other tools (BEST) criteria framework, ML-driven approaches have identified diagnostic, predictive, prognostic, and susceptibility glycan biomarkers across diverse pathologies, including malignancies (digestive, respiratory, and urogenital systems), cardiovascular diseases, autoimmune disorders, and chronic metabolic conditions such as chronic kidney disease and fatty liver [
96–
103].
In clinical diagnostics and therapeutic response prediction, these novel biomarkers demonstrate superior performance. Diagnostic ML models enable early-stage screening (e.g., detection of Stage I-II colorectal cancer [
98] and Stage 0-II non-small-cell lung cancer [
97], facilitate differential diagnosis in complex pathological contexts (e.g., distinguishing hepatocellular carcinoma from cirrhosis or chronic hepatitis B [
100]), and achieve precise subtyping (e.g., lung adenocarcinoma vs. squamous cell carcinoma [
97]; infectious vs. non-infectious febrile states [
104]). Furthermore, predictive models facilitate the early identification of clinical responses to specific interventions, such as chemotherapy [
101] or surgery [
105], providing a critical foundation for individualized targeted therapy. Additionally, ML is increasingly applied to prognostic quantification and latent risk assessment. Prognostic models categorize disease trajectories, such as overall survival and recurrence-free survival in oncological patients [
102]. Regarding risk assessment, glycan biomarkers are utilized to estimate “biological age”. By quantifying organ-specific glycosylation age (e.g., renal GlyKage) and its acceleration (GlyKageDiff) [
103] researchers can digitally assess latent cardiovascular risk and organ senescence.
ML-based multimodal biomarker discovery effectively integrates multidimensional biological data to characterize disease heterogeneity and complex biological correlations, significantly enhancing predictive accuracy, stability, and generalization [
106]. For instance, studies utilizing MALDI MS imaging to simultaneously capture peptide and N-glycan profiles have identified subtype-specific markers (e.g., tubulin, histone H2A, and fucosylated N-glycans). While single-mode ML models achieved accuracies of 87.49% (peptides) and 94.16% (N-glycans), multimodal data fusion further elevated accuracy to 94.25%–96.65%, representing a 4%–5.4% improvement over unimodal approaches [
107]. In conclusion, the convergence of ML and glycomics offers a transformative pathway for biomarker discovery and clinical translation, particularly in the early diagnosis, stratification, and risk prediction of cancer. Multimodal integration significantly optimizes diagnostic efficacy. Future coupling with genomics, proteomics, transcriptomics, and metabolomics is expected to further enhance model precision. Notably, while these methods have demonstrated high accuracy in retrospective and cross-sectional studies, rigorous validation through prospective clinical trials remains essential for clinical implementation.
5.1.2 Biological activity screening
The biological activities of glycans are mainly mediated through two pathways: direct interactions with receptor proteins and indirect effects via active metabolites generated by gut microbial metabolism. Some high molecular weight polysaccharides (e.g., glucans) can be recognized in the gut by receptors such as EGFR, Dectin-1, and BMPRIA, absorbed into the bloodstream, and subsequently interact with cell-surface receptors to exert biological activities [
107]. In contrast, macromolecular polysaccharides are often consumed as dietary components or food additives, and their physiological effects are generally mediated either directly through receptor recognition and limited absorption, or indirectly via microbial fermentation and degradation into bioactive metabolites such as short-chain fatty acids [
108]. On this basis, ML-assisted glycan bioactivity screening can be grouped into three major strategies: end-to-end bioactivity prediction, glycan-protein interaction prediction, and multi-omics-based prediction of gut microbial glycosidase profiles.
The first strategy directly predicts biological activity from glycan structures without explicitly modeling the underlying mechanism. Early studies used structural descriptors of isolated polysaccharides, such as molecular weight, monosaccharide composition, and glycosidic linkages, to predict immunomodulatory activities [
109]. Although these models achieved low prediction errors, their applicability was limited by the narrow structural space of the training data. More generalizable approaches represent glycans as strings or molecular graphs. Using graph convolutional networks, glycan molecular graphs have been applied to classify immunogenicity and pathogenicity with accuracies of 0.946 and 0.919, respectively [
86]. Similarly, string-based glycan representations combined with bidirectional recurrent neural networks achieved classification accuracies of 92% and 89% [
93]. The relatively broad glycan coverage of such models suggests their potential for predicting the activities of larger polysaccharides. Notably, the training set for the SweetTalk model contains 19,299 glycans, with about 2000 having a degree of polymerization greater than 10 (up to 43), indicating that such methods can effectively predict polysaccharides with molecular weights of thousands of Daltons.
The second strategy screens based on whether glycans can interact with microbial glycoside hydrolases or receptor proteins. Glycan-protein interaction models usually represent glycans using molecular graphs or 3D structural features to predict binding strengths or sites. As mentioned in Section 4.4, existing models can predict binding strengths [
86], but restricted by the protein types in training data, they mainly apply to viral proteins and have limited generalizability. DeepGlycanSite [
91], trained on 12,023 protein-carbohydrate complex crystal structures, has broader applicability, providing an effective method for screening glycan receptor proteins and microbial glycan hydrolases in a high-throughput manner.
The third strategy integrates multi-omics data to infer the glycosidase profiles of gut microorganisms. Based on fecal metagenomic datasets from infants and adults under different dietary or prebiotic interventions, ML models have been developed to characterize polysaccharide hydrolase repertoires and distinguish the glycosidase features of different
Bifidobacterium strains [
110]. These methods provide important support for screening functional microbes capable of selectively utilizing specific polysaccharides.
5.2 Testing of polysaccharide content and crop origin
5.2.1 Rapid non-destructive quantitative analysis of polysaccharide content and composition in crops and products
Spectroscopic techniques, including near-infrared (NIR), mid-infrared (MIR), and visible-near-infrared (Vis-NIR) hyperspectral imaging, provide non-destructive tools for detecting glycan-related structural information in complex samples. By extracting informative features from spectral signals, ML enables rapid and accurate quantification of polysaccharides and their components, offering advantages of high efficiency, low cost, and non-destructive analysis. At present, these approaches have been applied mainly to the detection of polysaccharides in agricultural products and beverages.
NIR spectroscopy and Vis-NIR hyperspectral imaging combined with ML have been widely used to predict non-starch polysaccharide content in agricultural products. For example, NIR-based random forest models achieved strong predictive performance for polysaccharides in
Cistanche tubulosa, with
R2 values above 0.86 and RPD values above 3.0 [
111]. NIR combined with ML can also simultaneously predict the content of multiple polysaccharides and other substances. Portable NIR systems have also been used for simultaneous quantification of cellulose, hemicellulose, lignin, and pectin in fresh sweet potato roots [
112]. In
Ganoderma lucidum, NIR spectroscopy enabled rapid, non-destructive prediction of both total polysaccharides and moisture, while NIR hyperspectral imaging further supported polysaccharide prediction across samples from different geographical origins [
113,
114].
Vis-NIR hyperspectral imaging and NIR spectroscopy combined with ML have begun to be applied to predict amylose and amylopectin contents across different crop varieties. In sorghum, Vis-NIR hyperspectral imaging coupled with ML achieved highly accurate prediction of amylose and amylopectin contents [
115]. Additionally, NIR-based PLSR models have successfully predicted total starch, amylose, the amylose/amylopectin ratio, neutral polysaccharides, acidic polysaccharides, and total soluble polysaccharides in yams, with all RPDs exceeding 2.5, thereby achieving simultaneous prediction for both starch and non-starch polysaccharides [
116]. Beyond agricultural products, NIR and MIR spectroscopy combined with ML have also shown promise for beverage analysis. In shiitake mushroom beverages, NIR-based models achieved highly accurate prediction of lentinan content, with
R2 values above 0.985 [
117]. In fresh orange juice, MIR spectroscopy combined with unsupervised and supervised ML methods enabled effective classification of soluble pectin levels, with accuracies of 89% and 85% for SVM and kNN models, respectively [
118]. Overall, the integration of spectroscopy and ML provides an efficient strategy for the rapid prediction, quantification, and classification of polysaccharides in complex food-related systems.
5.2.2 Crop origin and classification identification
This line of research utilizes fine structural differences in polysaccharides as “molecular fingerprints”. By combining high-efficiency separation techniques (like LC) and highly sensitive detection methods (like MS/MS, NMR) with ML algorithms for pattern recognition, it achieves precise identification of polysaccharide origins (geographical origin, cultivar, species). Researchers extracted polysaccharides from 96 batches of
Citri Reticulatae pericarpium across 4 varieties, utilizing HPLC to map acid hydrolysates and diffraction patterns to construct an HPLC fingerprint with 10 common peaks. An LDA model built upon this classified the varieties with > 0.954 accuracy [
119]. For 205 batches of
Gastrodia elata across 4 cultivars, 6 small molecular compounds and α1,4/1,6-glucan content were simultaneously measured. After using α-amylase-assisted Fenton’s initiation to degrade the polysaccharides toward defined oligosaccharide groups and precise quantification via UHPLC and LC–MS/MS, an XGBoost classification model achieved a weighted accuracy of 0.80 on the test set, with SHAP analysis confirming α1,4/1,6-glucan as the core distinguishing indicator [
120]. For 70 batches of American ginseng from 4 origins, monosaccharides and oligosaccharides were obtained via acid hydrolysis and ultrasound-assisted acid depolymerization. Monosaccharide composition/content and oligosaccharide content were acquired via HILIC-CAD, and an LDA model successfully distinguished the 4 origins with a 100% test set accuracy [
121].
Collecting 140 yams across 4 species, researchers (merging two species due to small sample sizes) applied SNV processing to NIR spectra. CNN and LSTM networks extracted features from cross-sectional and powder images, and a Multimodal Transformer predicted yam species with a test set accuracy of 97.30% [
116]. Using FTIR and FT-NIR spectroscopy on 262
Gastrodia elata samples from 3 species, a ResNet DL model achieved precise cultivar identification with 100% test set accuracy [
122].
5.3 Functional modeling and smart design of polysaccharide matrices
5.3.1 Hydrogels
Polysaccharide-based hydrogels and composite gel materials exhibit excellent biocompatibility, high water retention, structural tunability, and environmental responsiveness, leading to widespread use in the food industry, wound dressings, and 3D bioprinting. Key properties such as gel density, swelling degree, mechanical strength, water vapor transmission rate, porosity, and rheological behavior are significantly influenced by raw material ratios, preparation processes, and reaction conditions. With its powerful non-linear fitting, multi-objective optimization, and non-destructive testing capabilities, ML can efficiently optimize gel formulations, control processes, predict performance, and rapidly screen candidates, offering a novel pathway for the intelligent design and precise preparation of polysaccharide-based gels.
ML can precisely predict and regulate key gel properties based on process parameters and component concentrations. Using the yeast/glucose mass ratio, gelation temperature, and reaction time as input variables, researchers built a ML model to effectively predict gel density, maximum swelling degree, and compressive strength (
R2 = 0.9287). Coupled with GA for further optimization, structurally controllable hierarchical porous hydrogels were successfully prepared [
29]. Another study employed a QP-ANN model to optimize the ratio of PVA/chitosan/starch composite systems, significantly enhancing the water vapor transmission rate, gel content, swelling ratio, and porosity. This allowed the gel to maintain structural stability while efficiently absorbing wound exudate, keeping a moist healing environment, and providing good breathability, thereby meeting requirements for biological dressings and wound repair [
28].
In terms of gel performance prediction and smart auxiliary testing, ML enables rapid, non-destructive, and highly efficient predictions of key indicators based on component ratios, rheological properties, and spectral data, substantially increasing material screening efficiency. Based on component ratios, an RF model using hyaluronic acid methacrylate and gelatin methacrylate concentrations predicted the system’s shear rate with
R2 = 0.9739 [
123]. For rheology-based printability predictions, SVR models utilizing NIR spectra and LF-NMR data as inputs successfully predicted rheological properties [
124]. The NIR-based model achieved a prediction set
R2 up to 0.9796 (RPD = 7.0708), and the LF-NMR model reached
R2 = 0.9869 (RPD = 8.7220), facilitating rapid, non-destructive, indirect prediction of polysaccharide gel printability for 3D printing [
125].
5.3.2 Films and coatings
Polysaccharide-based films and coatings, favored for their green safety, biodegradability, and excellent biocompatibility, have become crucial materials in food packaging, preservation, and targeted delivery carriers. In practical applications, packaging films must simultaneously meet multiple key performance requirements, including good mechanical strength, barrier properties, and antibacterial activity. Coatings must fulfill specific needs, such as food preservation or targeted colonic release. ML provides efficient, precise solutions for formulation design, process control, and functional prediction of polysaccharide-based packaging materials.
In film preparation and optimization, ML effectively assists in optimizing raw material ratios and process parameters, significantly elevating overall film performance. Using tamarind seed starch, xanthan gum, and glycerol concentrations as inputs, a PNN model predicted the tensile strength and elongation at break of edible films with
R2 values of 0.9513 and 0.9287, respectively, providing a reliable basis for intelligent formulation optimization [
27]. An ANN model utilizing starch concentration, glycerol addition, preparation temperature, and chitosan content predicted film tensile strength and elongation at break with
R2 values of 0.9955 and 0.9859 [
3]. Another ANN model using ultra-high-pressure treatment pressure, essential oil concentration, and treatment time predicted the mechanical properties of potato residue edible films, achieving
R2 values of 0.9892 and 0.9866 for tensile strength and elongation at break [
126]. ML can also effectively assist in optimizing and synchronously enhancing the mechanical, barrier, and antibacterial properties of polysaccharide-based packaging. Using starch concentration and orange peel essential oil addition as inputs, an ANN accurately predicted film tensile strength, water vapor transmission rate, and inhibition zone diameter (all
R2 > 0.9999), facilitating the optimization of a highly antibacterial composite film formulation [
127].
For coating, ML can achieve precise formulation design and performance enhancement tailored to food storage requirements. Using storage time and concentrations of sodium alginate, guar gum, and pectin as inputs, SVR models predicted physicochemical indicators of strawberries (weight loss, firmness, TSS, pH, total color difference) with
R2 > 0.97. Subsequently, combining an overall quality score objective function with DE and PSO successfully improved strawberry quality during storage [
128]. Additionally, ML coupled with spectral techniques enables the rapid prediction of packaging release behaviors, supporting targeted coating design. Utilizing Raman spectra of polysaccharides as input, a model predicted the cumulative colonic drug release rate of polysaccharide and 5-aminosalicylic acid coatings with
R2 = 0.79 [
129].
5.3.3 Emulsions
Emulsion stability is a critical parameter affecting food shelf life, texture, and quality. Natural polysaccharides are widely used as food emulsion stabilizers due to their high safety and excellent emulsifying and stabilizing effects. As systems become increasingly complex, the efficiency of traditional evaluation methods becomes limited. Leveraging its efficient prediction and pattern mining capabilities, ML can rapidly elucidate stability mechanisms and optimize formulations, providing a highly effective new approach for regulating food emulsion stability.
ML effectively supports formulation optimization, efficient screening, and intelligent stability testing of polysaccharide-based emulsions, offering novel data-driven strategies for emulsion system design and characterization. An ANN model using pectin-rich orange by-product powder (OBF) and soy protein concentrations as inputs successfully predicted apparent viscosity, d10/d50/d90 percentiles, Zeta potential, percentage of flocculation, and creaming index (
R2 ≈ 0.997). Following multi-constraint optimization, the optimized ratio produced an emulsion with fine particle size, no obvious phase separation, low flocculation, and excellent 24-hour stability [
130].
ML also shows immense potential in the efficient screening of polysaccharide-based surfactants. Using molecular weight, glycosidic linkages, and monosaccharide composition of corn fiber gum as input features, a machine learning model predicted the TurbiScan Stability Index, Zeta potential, and particle size of emulsions, with prediction
R2 values for all three indicators ≥ 0.896 [
131]. Employing MOE molecular descriptors to characterize structural features alongside an RF algorithm, a QSAR classification model achieved high-throughput screening of natural emulsifiers containing polysaccharides with a 93.33% accuracy rate [
132].
Furthermore, ML enables intelligent, dynamic testing of polysaccharide-based emulsion stability. Based on droplet size and count data acquired via a particle size analyzer over different storage times, an LSTM network achieved time-series prediction of cumulative droplet mass percentages (
R2 = 0.898), which is applicable for long-term stability assessments of gum arabic-stabilized emulsions [
133]. Utilizing Confocal Laser Scanning Microscopy (CLSM) images as input, a dual-task DL model achieved statistical droplet classification (Oil-Extreme, Oil-Large, Oil-Medium, Oil-Small, Pickering-Medium, Pickering-Small) and morphological classification (Well-distributed, Flocculation, Coalescence) for Pickering emulsions, yielding mAP@0.5 scores of 0.9315 and 0.7853, respectively [
134]. This offers an efficient method for the automated quantitative analysis of emulsion microstructure and stability.
5.3.4 Cross-study comparison and generalization
Taken together, current studies demonstrate that ML can substantially improve interpolation, screening, and optimization within experimentally defined polysaccharide formulation spaces. However, there is currently insufficient evidence that one algorithm is universally superior. ANN models provide flexible nonlinear fitting but require careful regularization and validation; RF performs well on small tabular datasets but has limited extrapolation ability; SVR is suitable for small datasets but strongly depends on kernel definition and feature scaling; LSTM is advantageous only when meaningful sequential information is present; and GA, PSO, and DE can efficiently search complex surrogate response surfaces but provide no guarantee that the predicted optimum corresponds to the global optimum of the real physical system.
A further limitation is that most reported validation is interpolation-oriented. Random train/test splitting may place highly similar formulations, polysaccharide sources, spectra, or storage states in both subsets and therefore overestimate deployment performance. More rigorous evaluation should use grouped or external validation, such as leave-one-polysaccharide-source-out, leave-one-formulation-family-out, leave-one-instrument-out, or independent-laboratory testing. Model performance should also be reported together with an applicability domain or predictive uncertainty. These practices are particularly important for polysaccharides because variations in molecular-weight distribution, branching, substitution, source, purification, and processing history can produce substantial distribution shifts even when materials share the same nominal polysaccharide name.
Accordingly, the current evidence supports ML primarily as a local surrogate and screening tool rather than a universally transferable predictor. Progress toward genuinely intelligent polysaccharide material design will require models trained on structurally diverse datasets, standardized external benchmarks, explicit uncertainty assessment, and prospective experimental validation of AI-proposed formulations.
6 Challenges and Outlook
The preceding sections demonstrate that the principal bottlenecks in AI-driven glycoscience are interconnected rather than independent. Limited structural characterization constrains the quality of training data; incomplete molecular representations restrict structure–function modeling; narrow datasets limit model generalization; and insufficiently validated predictive models, in turn, restrict rational synthesis and formulation design. Therefore, future development should not be framed as isolated advances in analytical, synthetic, and computational technologies. Instead, these components should be integrated into an experimentally validated data–model–design loop.
6.1 Data scarcity is fundamentally a problem of both quantity and standardization
The first challenge is the limited availability of standardized, machine-readable glycan data. As discussed in Sections 4.2–4.4, current datasets are strongly biased toward well-characterized N- and O-linked glycans, whereas structurally heterogeneous and high-molecular-weight polysaccharides remain underrepresented. For many polysaccharides, only coarse descriptors such as molecular weight, monosaccharide composition, and branching degree are reported, and these measurements are often obtained using different analytical protocols.
Increasing data volume alone will therefore be insufficient. Experimental metadata must also be standardized so that measurements from different laboratories can be meaningfully integrated. Existing initiatives provide an important foundation. The MIRAGE initiative establishes minimum reporting requirements for glycomics experiments, including experimental procedures, analytical conditions, and data-processing information [
135]. GlycoPOST further provides a FAIR repository for raw glycomics MS data and incorporates MIRAGE-compatible experimental metadata [
136]. Future polysaccharide datasets should extend these principles by recording source organism, extraction and purification history, molecular-weight distribution, degree of polymerization, monosaccharide composition, linkage and branching information, chemical modifications, processing conditions, analytical platform, and batch information together with the measured properties.
High-throughput analytical methods can increase data acquisition, but analytical throughput alone does not guarantee model generalizability. CandyCrunch demonstrates that large-scale curated LC–MS/MS datasets can support high-accuracy deep-learning-based glycan identification [
10]. However, its success is associated with both data scale and the structural domain represented in the training set. The same performance cannot yet be assumed for heterogeneous high-molecular-weight polysaccharides for which sequence-resolved reference data remain scarce. Thus, future data acquisition should increase both dataset size and structural coverage.
6.2 Molecular representation must bridge sequence, conformation, and macroscopic function
A second bottleneck is the mismatch between the structural information available for glycans and the information required for predictive modeling. Sequence- and graph-based representations can encode monosaccharide identity, connectivity, branching, and anomeric configuration, but they cannot fully represent conformational ensembles. Conversely, coarse descriptors such as molecular weight and monosaccharide ratios are readily measurable for polysaccharides but cannot uniquely define their molecular architectures.
Future models should therefore integrate information across multiple structural scales, including primary sequence and topology, stereochemistry, three-dimensional conformation, molecular-weight distribution, chemical modifications, and processing conditions. GlycoShape represents an important step in this direction by providing conformational information derived from molecular simulations for hundreds of glycans [
81]. However, its current structural coverage remains much smaller than sequence-level glycan repositories and does not capture the full heterogeneity of high-molecular-weight polysaccharides. Consequently, 3D-informed representations should complement rather than replace experimentally derived structural descriptors.
For polysaccharide matrices, molecular descriptors and formulation variables should also be explicitly distinguished. A model based only on polymer concentration may accurately describe one experimental system but cannot distinguish concentration effects from differences in molecular weight, substitution, branching, or source. Integrating molecular structure with processing and formulation variables would therefore shift ML from empirical formulation fitting toward more transferable structure–process–property relationships.
6.3 Generalization and uncertainty should become primary evaluation criteria
A third major challenge is model evaluation. As shown in Section 5.3, high R2, accuracy, or AUC values are frequently reported, but these metrics are commonly derived from random partitions of datasets generated within a single laboratory, material source, or experimental design. Such validation primarily evaluates interpolation and provides limited evidence that a model will perform reliably for previously unseen glycans, polysaccharide sources, or processing conditions.
Future benchmarks should therefore deliberately evaluate models under distribution shifts. GlycanML provides a useful example by constructing standardized benchmarks for multiple glycan prediction tasks and using motif-based dataset splitting to evaluate models on glycans containing structural patterns distinct from those encountered during training [
123]. Similar principles could be extended to polysaccharide research through leave-one-source-out, leave-one-structural-family-out, leave-one-batch-out, leave-one-instrument-out, and independent-laboratory validation. Such evaluation would provide a more realistic estimate of model performance when applied to newly characterized polysaccharides.
Predictive uncertainty should also accompany point predictions, particularly when ML models are coupled with optimization algorithms. In inverse design, an optimizer may preferentially search regions where a surrogate model predicts extreme performance even though these regions are poorly represented in the training data. This can lead to apparently optimal solutions that reflect model extrapolation errors rather than genuinely superior experimental conditions. Uncertainty-calibrated reaction prediction has demonstrated that model confidence can provide useful information on prediction reliability [
137]. Extending uncertainty estimation and applicability-domain analysis to glycan and polysaccharide models would therefore improve the reliability of candidate selection and experimental decision-making.
6.4 Connecting prediction with synthesis and experimental validation
The final challenge is to convert predictive capability into experimentally actionable glycan design. This connection is becoming increasingly feasible because advances in glycan synthesis have expanded the experimentally accessible structural space. Automated glycan assembly has enabled the synthesis of a sequence-defined 100-mer polysaccharide [
138], while automated solution-phase multiplicative synthesis has subsequently extended chemically synthesized glycan chains to 1080 monosaccharide units [
139]. These advances indicate that increasingly large and structurally defined glycans can be prepared experimentally, providing an important basis for establishing systematic sequence–structure–property datasets.
At the computational level, data-efficient approaches may be more practical than simply increasing model complexity. Pesciullesi et al. demonstrated that transfer learning from a general chemical reaction model to a smaller carbohydrate-specific reaction dataset substantially improved the prediction of regio- and stereoselective carbohydrate transformations and experimentally validated the resulting model during oligosaccharide synthesis [
140]. This study provides direct evidence that information learned from larger chemical reaction datasets can be transferred to data-scarce carbohydrate chemistry. However, the effectiveness of transfer learning will depend on the similarity between the source and target chemical domains and should therefore be evaluated rather than assumed.
The long-term objective should consequently be a closed-loop workflow rather than independent advances in analytical characterization, synthesis, and AI. High-quality structural and functional measurements should first be deposited using standardized representations and metadata. These data can then support predictive or generative models that propose candidate glycan structures, reaction conditions, or formulations. Automated or semi-automated synthesis and experimental characterization can subsequently test these candidates, while both successful and unsuccessful experiments are returned to the dataset for model updating. This framework directly connects the analytical strategies discussed in Sections 4.2–4.4, the synthesis and reaction-prediction approaches described in Section 4.1.2, the inverse-design concepts discussed in Section 4.5, and the formulation optimization strategies summarized in Section 5.3.
Accordingly, the immediate priority for AI-driven glycoscience is not necessarily the development of increasingly complex algorithms. More critical needs are structurally diverse and standardized datasets, evaluation protocols that explicitly test out-of-distribution generalization, uncertainty-aware predictions, and prospective experimental validation. Addressing these issues would allow the field to progress from high-accuracy retrospective fitting toward reproducible, transferable, and experimentally actionable glycan design.
7 Conclusions and Future Researches
In summary, machine learning is reshaping glycan research by providing powerful tools for handling high-dimensional, nonlinear, and heterogeneous data that are difficult to address using conventional experimental and statistical approaches. Across the full research pipeline, from extraction and synthesis to sequence elucidation, 3D structures prediction, interaction analysis, and property-guided molecular design, ML has demonstrated substantial potential to accelerate discovery and improve precision. In particular, the integration of ML with NMR, MS, nanopore sensing, spectroscopy, imaging, and glycomics has greatly expanded the ability to characterize complex glycans and uncover their functional relevance in food science, biomedicine, and biomaterials.
Despite the rapid progress of AI-driven glycoscience, several challenges remain, including the limited availability of large-scale standardized datasets, the structural bias of existing data toward well-studied glycan classes, and the continued reliance on coarse-grained descriptors that fail to fully capture sequence and conformational complexity. Future advances will rely on the integration of high-throughput analytical technologies, improved glycan synthesis, and next-generation AI approaches such as transfer learning, multimodal learning, and physics-informed modeling, collectively driving glycoscience from descriptive prediction toward mechanism-based understanding and rational design.
The Author(s) 2026. This article is published by Higher Education Press.