A recent review in npj Digital Medicine lays out how AI-driven virtual cell models are moving from computational curiosity to a working engine for preclinical research. These models integrate single-cell transcriptomics, spatial transcriptomics, proteomics, and other high-throughput readouts with deep generative networks, graph neural networks, and physics-informed architectures to predict how a cell will respond to a drug, a genetic edit, or a disease perturbation. For the glycobiology community, the timing is notable. Virtual cells are only as good as the molecular state they are trained and validated on, and glycosylation is one of the most information-dense, functionally decisive layers of that state. In this article we examine what a virtual cell model is, why glycan and glycoprotein data belong inside it, and how rigorous Glycomics Analysis and glycoproteomics services supply the evidence loop that makes such models trustworthy.
Fig. 1 Comprehensive overview of AI-driven virtual cell models in preclinical research. (Ma, et al. 2026)
What an AI Virtual Cell Model Actually Is
A virtual cell is a computational representation of a cell's functional states, signaling networks, and dynamics under perturbation. Unlike a classic kinetic model built around a handful of hand-picked reactions, an AI virtual cell learns latent structure from very large, multimodal datasets and uses that structure to extrapolate to conditions it has not seen. The review describes three complementary technical routes. Deep generative models, including variational autoencoders, flow-matching, and diffusion frameworks, sculpt the cell-state space and can synthesize expression profiles for in silico perturbation experiments. Graph neural networks treat each cell as a node and capture the communication topologies that drive tissue behavior. Physics-informed neural networks embed known biophysical laws so that predictions remain biologically plausible rather than statistically plausible alone. Foundation models pretrained on tens of millions of cells then transfer that knowledge from the gene and cell level up to organ and patient levels.
For glycobiologists, the important point is that these models are fundamentally data-hungry and modality-agnostic. Proteomics sits explicitly among the training modalities, and protein function is inseparable from its N- and O-linked glycans. A virtual cell that predicts drug response but ignores the glycosylation layer is, in effect, predicting the behavior of a protein it has never fully measured.
Why Glycosylation Belongs Inside the Virtual Cell
Glycosylation is not a decorative appendage on a protein. It shapes folding, stability, trafficking, receptor engagement, and immune recognition. In biologics especially, the fine structure of a glycan can flip a therapeutic from effective to inert or from safe to immunogenic. Fucosylation of an antibody's Fc N-glycan tunes antibody-dependent cellular cytotoxicity; terminal sialic acid modulates anti-inflammatory signaling; branched bisecting GlcNAc alters clearance and effector function. Small-molecule programs are no exception: cell-surface and secreted glycans change with proliferation, stress, and differentiation, and those changes are themselves druggable or diagnostic.
A virtual cell model that simulates drug response in a cancer or immune cell therefore needs the glycosylation state as an input and as a readout. Routine Glycan Profiling Data provides the carbohydrate-level census of a cell population, while site-resolved glycoproteomics tells the model which proteins carry glycans, where, and with what occupancy. Without that layer, predictions about monoclonal antibody efficacy, glycan-binding protein engagement, or lectin-mediated uptake rest on an incomplete molecular picture.
The reverse also holds. Glycosylation is notoriously context dependent: the same gene produces different glycoforms in different tissues, nutrient states, and cell lines. Virtual cell models, trained on multimodal omics that include glycomics, are uniquely positioned to learn those context rules and to predict how a glycoform distribution will shift under a perturbation. That capability is exactly what preclinical teams need when they screen biologics for consistent, desirable glycosylation across manufacturing clones and patient subgroups.
From Glycan Data to Training and Closed-Loop Validation
The review is emphatic that virtual cell predictions are only as credible as the validation loop behind them. The workflow pairs computational evaluation, measured by distributional concordance and uncertainty quantification, with experimental verification using CRISPR assays, organoids, and organ-on-chip systems. Glycomics fits neatly into both halves of that loop. High-quality, reproducible glycan measurements are training data; they are also the ground truth against which a model's predicted glycosylation shift is checked.
This is where measurement standardization matters. A virtual cell is only as portable as the datasets it learns from, and glycan data are unusually sensitive to sample handling, release chemistry, and instrument calibration. Isomer-resolved Glycan Separation by porous graphitized carbon separation resolves isomers that conventional methods collapse together, reducing the noise that would otherwise masquerade as biological signal. Consistent release, derivatization, and enrichment workflows keep batch effects small enough that cross-study integration is meaningful. When the underlying glycomics is rigorous, distributional-distance metrics such as Wasserstein distance or maximum mean discrepancy become honest tests of whether a model's predicted glycoform shift matches experiment. With this validation framework established, the discussion naturally turns to how glycosylation-aware virtual cells can be deployed in practical drug-discovery and development workflows.
Applications in Glycosylation-Aware Drug Discovery
Once glycosylation is represented in the model, several translational applications become accessible. The first is precision screening of biologics. Therapeutic proteins such as monoclonal antibodies require careful Glycosylation Characterization of Antibody Therapeutics, because Fc and Fab glycans govern potency, half-life, and immunogenicity. A virtual cell that carries the glycosylation layer can rank candidate clones or engineering strategies in silico for the desired effector profile before a single batch reaches the bioreactor.
The second is mechanistic inference. By introducing a candidate regulatory factor into the model and observing whether it reproduces a known glycosylation phenotype, researchers can prioritize hypotheses for wet-lab testing. This is a natural fit for teams that need bespoke glyco-constructs to confirm mechanism, who can draw on Tailored Glycoengineering Workflows to generate the exact standards or modified proteins the model calls for.
The third is the reduction of animal testing. Organoids and organ-on-chip platforms increasingly serve as human-relevant testbeds, and paired glycan analysis of those models supplies the carbohydrate readout that animal cohorts used to provide. Chemistry-driven programs such as Glycosylation-targeted Inhibitor Programs likewise benefit: a virtual cell can preview which glycosylation nodes are most exploitable in a given tumor context, focusing scarce screening capacity on the highest-yield targets. Finally, Site-resolved Glycoprotein Structural Analysis anchors the model's predictions in atomic-level evidence, closing the gap between a population-level glycomics signal and the structural change that produced it.
Fig. 2 Applications of AI-driven virtual cell models in preclinical research. (Ma, et al. 2026)
Challenges and Where Glycobiology Services Fit
The review is candid about the remaining hurdles: model interpretability, training-data bias, and an unsettled regulatory posture in which agencies treat AI outputs as supporting evidence rather than sole justification. Each of these has a glycobiology dimension. Interpretability improves when the model is constrained by real glyco-chemistry rather than left to invent it. Bias shrinks when diverse, well-annotated glycomics datasets enter the training pool instead of a few convenience cell lines. Regulatory acceptance grows when predictions are backed by reproducible, audit-ready glycan measurements.
That is the practical role for a glycobiology service provider. Virtual cell platforms will multiply, but they cannot generate the substrate-level evidence themselves. Turnkey, standardized N- and O-glycan profiling, glycoprotein enrichment, and quantitation are what convert a promising model into a validated one. As preclinical research shifts toward a simulation-first, validation-driven loop, the quality of the glycomics fed into that loop becomes a rate-limiting step for the entire field.
Conclusion
AI virtual cell models are reframing preclinical research as a closed loop between computation and experiment, and glycosylation is too central to cellular function to be left out of that loop. The opportunity for glycobiology is clear: supply the high-fidelity glycomics and glycoproteomics data that make virtual cells accurate, interpretable, and regulator-ready. For teams building or adopting virtual cell platforms, partnering early with rigorous glycosylation characterization is the difference between a model that merely fits data and one that predicts biology.
Related Services & Products
- Glycan Release Service
- Glycan Separation and Purification Service
- Glycan Derivation Service
- Glycomic Characterization Service
- Glycomic Quantitation
- N-Glycan Profiling Service
- O-Glycan Profiling Service
- Glycoprotein Enrichment
- Glycoprotein Quantification
- Sialic Acid Analysis Service
Reference
- Ma, C., et al. (2026). AI-driven virtual cell models in preclinical research: technical pathways, validation mechanisms, and clinical translation potential. npj Digital Medicine. DOI: 10.1038/s41746-025-02198-6.
