Viruses hide their most conserved, antibody-vulnerable surfaces under coats of sugar trees called glycans. These glycans are not encoded by the virus at all: the viral genome specifies only where sugars attach, while the host cell's own biosynthetic machinery builds and decorates the trees. In influenza, specific glycan structures on hemagglutinin physically block broadly neutralizing antibodies from reaching the conserved stem region, explaining why structurally similar lineages neutralize so differently (Xiao et al.), and in SARS-CoV-2, mutations and deletions that create new glycan attachment sites measurably erode vaccine efficacy, estimated at 56% against moderate-to-severe/critical COVID-19 in the single-dose Ad26.COV2.S ENSEMBLE analysis of 484 vaccine and 1,067 placebo recipients (Magaret et al., with the structural mechanism quantified by Magaret et al.). Yet every existing model of glycan shielding treats the shield as a fixed property of the viral strain — one glycoform per sequence.
This is wrong twice over, and both errors point the same direction. First, different host cell types express very different repertoires of glycosyltransferases, the enzymes that assemble glycans, so the same viral gene product built in two cell types can carry two different shields. Second, even within one cell type, mRNA levels of these enzymes are unreliable guides to enzyme abundance: across the human proteome, mRNA abundance correlates with protein abundance only 40–60% of the time ($R^2$ of 0.16–0.36), because translational efficiency, rRNA modification, and differential turnover intervene (Adhikari et al.; one concrete mechanism is 18S rRNA methylation boosting CCND3 translation without any transcript change, per METTL5-mediated m6A modification of 18S rRNA). And bulk tissue measurements average these differences away, as when bulk lung profiling could not identify which cell lineages harbored active SARS-CoV-2 replication (Delorey et al.), or when rare stem-like clones driving tamoxifen resistance were washed out into an apparent homogeneous response (Lin et al.).
The proposal in one sentence: build a generative pipeline that takes a viral surface protein sequence plus a single-cell (or deconvolved bulk) transcriptomic profile of the host producer cell, and outputs a distribution of site-specific glycan trees — graph-structured, isomer-resolved — together with the resulting cell-type-specific antibody-accessibility of every epitope, then test whether this cell-conditioned view of the shield predicts neutralization and vaccine breakthrough better than strain-level models.
Core Claim
The claim is that the glycan shield is a property of the (host cell, viral sequence) pair, not the viral sequence alone, and that this coupling is computationally tractable through three known components that nobody has connected: a discordance-aware translator from single-cell RNA to enzyme activity, a biosynthesis-constrained generator of glycan graphs, and a structural occlusion model conditioned on those graphs. The intersection is genuinely empty. scChord and scTranslator map single-cell transcriptomes to protein abundances (scTranslator improves cosine similarity by 10.7% over sciPENN and cTP-net and holds similarity above 0.83 under shuffled input genes), but both predict only linear peptide abundance and are explicitly blind to glycoform architecture. DeepGP shows graph networks (GCN/GIN over branched glycan topologies) are the right tool for glycan chemistry, but it operates on bulk LC-MS/MS digests with no notion of a producing cell. Harvey et al. show structural priors dramatically sharpen genotype-to-antigenic-phenotype mapping, but their priors (receptor-binding-site distance, static accessibility scores) assume one frozen structure and homogeneous host glycosylation. Meanwhile the high-resolution SARS-CoV-2 molecular mapping study demonstrates the cost of that assumption: recombinant antigens produced in cells lacking the right post-translational machinery display artificially exposed surfaces, yielding false antibody footprints that are fully occluded in natural infection (SARS-CoV-2 molecular mapping). The static prior that helps Harvey on influenza is the same assumption that produces misleading epitope maps in SARS-CoV-2; the fix is not to discard structure but to condition it on host-cell state.
The core mechanism is compact. Glycan biosynthesis is an ordered walk through the Golgi reaction network, where each step is catalyzed by a specific glycosyltransferase. If $a_{r,c}$ is the activity of the enzyme catalyzing reaction $r$ in host cell type $c$, then the probability of glycan $g$ at attachment site $s$ is modeled as
$$P(g \mid s, c) \;\propto\; \prod_{r \in \mathcal{R}(g)} a_{r, c},$$
where $\mathcal{R}(g)$ is the ordered reaction sequence that builds $g$. Summing over the per-site glycoform distribution gives a per-residue occlusion probability $\pi_i(c)$ — the chance that residue $i$ of the viral protein is sterically covered by glycans in a virion produced by cell type $c$. The prediction target is therefore not one structure but a cell-type-indexed distribution of accessibility profiles.
Research Program
Three modules, each reusing a method the file identifies as proven and each extending it in exactly one dimension.
First, a discordance-aware enzyme-activity module. We take single-cell RNA profiles and infer latent glycosyltransferase activities rather than assuming transcript equals enzyme. The method supplier is scChord's architecture — a zero-inflated negative-binomial VAE that separates technical noise from biology, followed by optimal-transport conditional flow matching that maps RNA to full protein distributions rather than conditional means, which is what lets it capture bimodal CD4 and heavy-tailed CD45RA expression (scChord). We reuse that machinery but retarget it from 13–228 antibody-derived surface tags to enzyme activities, and add a physics-style consistency constraint: predicted activities must reproduce observed site-specific glycoproteomic outputs under the biosynthesis equation above. Annotating the enzyme repertoire itself is feasible: the glycosyltransferase fold-landscape CNN classifies GT folds at 96% accuracy and families at 77%, including previously uncharacterized families, so unannotated enzymes in a scRNA profile can be typed computationally (GT fold landscape).
Second, a glycoform graph generator. From predicted activities we generate site-specific glycan trees as graphs, using the biosynthesis network as the constraint (the product form above makes the walk a cell-conditioned Markov chain) and DeepGP-style graph encoders as the scoring machinery, since DeepGP's GCN/GIN components already resolve glycan structural isomers that share mass but differ in branch topology and hence steric reach (DeepGP). Isomer-level resolution matters here precisely because shielding is geometric: two isomers of identical composition shield differently.
Third, a cell-conditioned occlusion layer. Generated glycoform trees are grafted at their attachment sites onto the protein's 3D coordinates, and per-residue accessibility is computed over the glycoform distribution, replacing Harvey's static priors with $\pi_i(c)$ that varies by producer cell type (Harvey et al.). For bulk tissues — a nasopharyngeal swab, a tumor sample — the host cell mixture is recovered by deconvolution, where Count Bridges (a Poisson birth-death bridge over integer count space that outperformed CIBERSORTx, MuSiC, and STDeconvolve) and POPPY (Cell Ontology alignment that improved cross-cohort silhouette scores by up to 73% and composition-correlation by up to 714%) are the strongest available tools; we feed their deconvolved cell-type proportions into the occlusion layer to obtain a population-level mixture of shielded and vulnerable virions.
An honest data constraint shapes the program: no paired single-cell RNA + site-specific glycoproteomics dataset appears in the file, so year one is largely about assembling the training signal — cell-type-sorted bulk glycoproteomics (the pattern is set by the aging mouse ovary N-glycoproteome study, which showed site-specific glycosylation shifting independently of gross protein abundance, at bulk resolution that obscures which follicular or stromal cell types changed (ovary glycoproteome)), matched scRNA references for the same tissues, and simulated glycoproteomes generated from the known reaction network with controlled enzyme profiles.
Baselines
The do-nothing reference on the antigenic-mapping question is the structurally naive Bayesian variable selection of Harvey et al., which achieves 59.8% decisive inclusion/exclusion of hemagglutinin positions and places selected residues a mean 6.86 Å from experimentally determined antigenic sites (Harvey et al.). The strongest existing competitor is the same authors' structurally aware version — receptor-binding-site proximity and BEpro accessibility as hierarchical priors — which raises decisiveness to 72.4%, cuts the mean distance to 2.55 Å, and predicts haemagglutination inhibition titers for uncharacterized viruses with a mean absolute error of 0.52 log₂ units in cross-validation over 24 years of data (more than 38,000 HI titers paired with 1,737 HA1 sequences). Our priors replace their static scores with cell-conditioned occlusion probabilities; the comparison is fair by construction — identical data, identical Bayesian machinery, identical compute and tuning budget, changing only the prior. Beating it means MAE below 0.52 log₂ units on the same cross-validation folds, and we consider the claim supported only if the gain holds when the conditioning cell mixture is inferred rather than given.
On the glycoform-prediction side, the strongest competitor is a two-stage pipeline built from existing parts: scChord or scTranslator predicting glycosyltransferase protein abundance from single-cell RNA (scChord reports the lowest RMSE and highest Pearson correlation against totalVI, sciPENN, and scTranslator on GSE100866, though no reported result found on Lacuna for the exact RMSE or PCC values), feeding a rule-based canonical-glycan assignment. That baseline isolates exactly our missing ingredient in the literature — abundance models that never output a glycan graph — so if it matches our graph-generative model on glycoproteomic recapitulation, the graph machinery is not earning its keep. On bulk deconvolution, CIBERSORTx, MuSiC, and STDeconvolve serve as the established reference points; Count Bridges reports outperforming them on the Onek1k cohort and MERFISH data, but no reported result found on Lacuna for the exact numerical error values, so we will re-run all of them ourselves on identical inputs.
Finally, four ablations isolate the proposal's ingredients: replacing the flow-matching activity model with a conditional-mean regression (tests whether distributions matter); collapsing all host cell types into one averaged enzyme profile (tests whether heterogeneity matters); using raw GT mRNA as activity, i.e., assuming transcript equals protein (tests whether discordance modeling matters); and grafting one static canonical glycoform per site from a solved structure (tests whether cell-conditioned generation matters).
Why not just…?
Why not just measure bulk mRNA of glycosyltransferases? This is the most serious alternative because it is cheap and partially informative. It fails for two documented reasons: mRNA predicts protein only 40–60% of the time with $R^2$ of 0.16–0.36, so bulk GT transcript counts do not track functional enzyme abundance (Adhikari et al.), and bulk averaging obscures which cell types are actually present and active, as in the SARS-CoV-2 tissue atlas work (Delorey et al.). But the honest decisive test is ours to run: we construct synthetic bulks (in the spirit of POPPY's 2,286 simulated Kassandra profiles) in which a rare producer population with a distinctive enzyme profile generates a disproportionate share of under-shielded, immunologically vulnerable virions. If a bulk GT-mRNA regression recovers the population-level shield distribution as well as deconvolution plus our pipeline, the single-cell machinery is unnecessary for this question and we will say so plainly.
Why not just compute solvent accessibility on static structures (PDB or AlphaFold)? Because the artifact this produces has already been documented: recombinant SARS-CoV-2 antigens expressed without host-matched glycosylation expose surfaces that are fully occluded in natural infection, and assays built on them identify misleading, non-neutralizing binding footprints (SARS-CoV-2 molecular mapping). Static structures also carry no glycan isomer information, yet isomers of identical mass have different branch topologies and different steric reach, which is exactly why DeepGP needed graph networks to distinguish them (DeepGP). Our replay experiment settles it: the model should predict, without being told, that surfaces exposed on glycan-deficient production cell lines are buried on airway-cell-produced virions, matching the documented in vivo footprints.
Why not just run scTranslator or scChord and use the predicted protein levels? Because those systems predict only linear abundance — scTranslator generates total abundance for over 14,000 proteins, scChord 13–228 surface tags — and neither outputs glycoform architecture or site occupancy, a gap both papers acknowledge as their boundary (scTranslator; scChord). The fair version of this alternative is one of our ablations: scChord-predicted GT abundances feeding our glycoform generator. If it performs as well as our activity-inference module, then the biosynthesis-consistency training is superfluous and the simpler path wins; the ablation is powered to answer that directly. We note one more caution from scChord's own limits: correlation-matrix fidelity degrades past 100–200 protein targets, which our small-but-structured enzyme panel is designed to sidestep.
Decisive Experiments
Four checks, each anchored to a benchmark in the file. First, influenza antigenic mapping: rerun Harvey's Bayesian variable selection on the 24-year H3N2 corpus with cell-conditioned occlusion priors in place of static ones, targeting mean absolute error below 0.52 log₂ HI units and mean residue-to-antigenic-site distance below 2.55 Å on the same folds. Second, glycoform recapitulation: predict site-specific glycan distributions for mouse tissues and the aging ovary, score predicted glycopeptide spectra against measured LC-MS/MS using DeepGP's spectral predictor as the fixed evaluator (DeepGP itself reports median cosine similarity above 0.95 on mouse tissue test sets and 0.977 cross-instrument after fine-tuning, plus 96.2% correct resolution of synthetic isomer mixtures and 122% more peptide-spectrum matches at a 5% decoy ratio when rescoring pGlyco3 — these numbers bound the evaluator's fidelity, not our task, and no reported result found on Lacuna for any prior model's accuracy at predicting tissue glycoform distributions from RNA, which is exactly the empty cell we aim to fill). Third, the in vitro artifact replay described above, scored against the documented binding-footprint discrepancies. Fourth, vaccine breakthrough: on the ENSEMBLE cohort (484 vaccine versus 1,067 placebo infections; vaccine efficacy 56%), test whether predicted shield-coverage of neutralizing epitopes under airway-cell glycoform mixtures separates breakthrough sequences from placebo sequences better than sequence features alone (Magaret et al.); no reported result found on Lacuna for an existing predictor's discrimination on this cohort, so we establish the first number ourselves.
One scope limit, stated plainly: this proposal models glycan shielding only. Conformational masking — spikes that tuck their receptor-binding cores into closed trimer states opened transiently during entry (Song et al.; Hotzel) — is an orthogonal masking mechanism requiring an ensemble dynamics model we do not attempt here; the occlusion layer is designed so a conformational ensemble can later replace the single grafting conformation without changing the rest of the pipeline.
References
- Light chain modulates heavy chain conformation to change protection profile of monoclonal antibodies against influenza A viruses
- Quantifying how single dose Ad26.COV2.S vaccine efficacy depends on Spike sequence features
- A high-stringency blueprint of the human proteome
- METTL5‑mediated m6A modification of 18S rRNA drives oral squamous cell carcinoma progression by enhancing CCND3 translation
- COVID-19 tissue atlases reveal SARS-CoV-2 pathology and cellular targets
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- A pre-trained large generative model for translating single-cell transcriptome to proteome
- Deep Learning Prediction of Glycopeptide Tandem Mass Spectra Powers Glycoproteomics
- A Bayesian approach to incorporate structural data into the mapping of genotype to antigenic phenotype of influenza A(H3N2) viruses
- Detailed analysis of antibody responses to SARS-CoV-2 vaccination and infection in macaques
- Mapping the glycosyltransferase fold landscape using interpretable deep learning
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- A high-resolution N-glycoproteome landscape of aging mouse ovary
- Cryo-EM analysis of the HCoV-229E spike glycoprotein reveals dynamic prefusion conformational changes
- Structural Evolution of Trimeric Gamma-Type Retroviral Envelope Proteins