A 2024 bioRxiv preprint (Kellman, Sandoval, ... Lewis; revised February 2026) reported that glycosite-proximal protein sequence and three-dimensional structure are statistically associated with the glycan structures found at each glycosylation site across human glycoproteins, quantified these associations as a substitution matrix ('glycoimpact', assembled into the BLAMO matrix), and argued that glycosylation is a 'genetically encoded bounded biosynthesis' rather than a purely cell-determined, non-templated process. The authors engineered Rituximab variants whose glycoprofiles shifted as predicted, but left the causal and mechanistic core open: the associations remain correlational, the proposed physical mechanisms (aromatic stacking, hydrogen bonding, steric exclusion) are undemonstrated, prospective validation of predicted disease variants such as tyrosinase A355V in patient-derived cells is unperformed, the SARS-CoV-2 spike predictions were tested only on the monomeric S1 subunit, and the human/mammalian-derived trends are untested in other organisms. The open problem is to establish the mechanism and scope of protein-structure encoding of glycosylation: demonstrate the structural-chemical basis causally, validate the framework's specific predictions prospectively, and test its generality across expression systems and organisms.

Background

Glycosylation — the enzymatic attachment of glycan (sugar-chain) modifications to proteins — is the most common post-translational modification of secreted and membrane proteins. Unlike DNA, RNA, and protein synthesis, which read off templates, N-linked glycan biosynthesis is traditionally described as non-templated: after a preassembled oligosaccharide is transferred to an asparagine within a sequence motif (the sequon, N-X-S/T), a cascade of glycosidases and glycosyltransferases (enzymes that remove and add sugar units, respectively) in the endoplasmic reticulum (ER) and Golgi trims and extends it, and which structures survive is thought to be governed mainly by the cellular environment — enzyme expression levels, nucleotide-sugar donor availability, and organelle residence times. Yet each glycosylation site on a given protein reproducibly carries a characteristic distribution of glycan structures (site-specific microheterogeneity), which pure cell-state determinism struggles to explain.

Evidence has accumulated that the protein itself contributes. Beyond the founding sequon, an aromatic residue two positions before the glycosylated asparagine (position N-2; the 'enhanced aromatic sequon') raises glycosylation efficiency (Huang et al. 2017); large-scale site-specific glycoproteomics — mass-spectrometric identification of the glycan structures at each attachment site — showed that protein structure dictates N-glycan type, core fucosylation and branching (Thaysen-Andersen and Packer 2012); steric accessibility of individual sites controls SARS-CoV-2 spike glycan processing (Allen et al. 2021); and structural context modulates the fucosyltransferase FUT8, a Golgi enzyme that attaches fucose to the core of N-glycans (Garcia-Garcia et al. 2021).

The source study (Kellman et al., bioRxiv 2024, revised Feb 2026) systematized this line into a quantitative framework. Building PGES-DB, a curated database pairing glycosite-proximal protein structural annotations with measured site-specific glycoprofiles on 111 human glycoproteins, the authors quantified 'intra-molecular relations' (IMR) between glycosite-proximal amino-acid features and glycan substructures (e.g., the oligomannose-versus-complex N-glycan class distinction — the two major N-glycan structural classes — as well as sialylation and branching of the sugar chains), and defined glycoimpact: the expected change in a site's glycoprofile caused by substituting one amino acid for another near the glycosite, computed as a z-score-normalized distance between expected glycoprofiles and assembled into a 20-by-20 substitution matrix they call BLAMO ('BLOSUM-PAM Orthology'), by analogy to the classical BLOSUM and PAM amino-acid substitution matrices. They reported that high-glycoimpact residues co-evolve with glycosites, that glycoimpact correlates with divergence between conservation-based and structure-based pathogenicity predictions, and that thousands of disease-associated variants in ClinVar, the NIH clinical human-variant database, that lie near glycosites are high-glycoimpact — proposing, for example, that the variant A355V of tyrosinase (the key melanin-biosynthesis enzyme, whose substitution at this position is a known sufficient cause of oculocutaneous albinism type 1; Simeonov et al. 2013, Spritz et al. 1997) may act through aberrant glycosylation. As causal support they designed five Rituximab (an anti-CD20 therapeutic antibody) variants, four of which shifted glycoprofiles as predicted, and showed IMR-consistent differential glycosylation on HIV gp160 (the HIV envelope glycoprotein), the IgG3 Fc region (an antibody constant-domain fragment), and the SARS-CoV-2 spike S1 subunit (the receptor-binding half of the spike, examined as a monomer) across variants of concern, measured by site-specific mass spectrometry. A companion preprint (Kellman et al. 2024) supplies a deep-learning predictor (recurrent plus graph neural network) of site-specific glycosylation from glycosite-flanking structure. On this basis the authors claim glycosylation is a 'bounded biosynthesis' — constrained by the genetically encoded protein structure — while acknowledging that the analysis covers only human glycoproteins from mammalian expression systems, that ER quality-control machinery and chaperones were not modeled, that dynamic cellular parameters (nucleotide-sugar donor concentrations, glycosyltransferase expression, organelle residence times) are absent from the framework, and that extension to other post-translational modifications is untested.

Problem Statement

Establish whether and how protein structure causally specifies site-specific glycosylation, by converting the correlational glycoimpact framework of Kellman et al. into a mechanistically grounded and prospectively validated account of 'genetically encoded bounded biosynthesis'. Concretely, the source poses the following four linked open questions, which jointly support this one objective; each is independently refutable (a demonstrated refutation of any one resolves that sub-question as a negative result): (1) Mechanistic basis — directly demonstrate (or refute) that the proposed physical interactions between glycosite-proximal protein structure and the glycan-processing machinery (aromatic stacking, hydrogen bonding, steric exclusion) cause the observed IMR/glycoimpact associations, rather than the associations arising from co-evolution, database bias, or other confounders. (2) Prospective disease-variant validation — for predicted high-glycoimpact, glycosite-proximal pathogenic variants, beginning with the source's motivating case tyrosinase A355V in oculocutaneous albinism, directly measure altered glycosylation and its functional consequence in patient-derived or otherwise disease-relevant cells. (3) Physiological viral validation — confirm the predicted variant-dependent differential glycosylation of SARS-CoV-2 spike on the physiological trimer and in whole-virus context, where it bears on immune evasion. (4) Generality — test whether the human/mammalian-derived quantitative trends and BLAMO/glycoimpact matrices extend to glycoproteins of other organisms (plants, bacteria, lower eukaryotes) with different processing pathways. The authors also suggest, but have not tested, extending the bounded-biosynthesis concept to other post-translational modifications such as phosphorylation or acetylation; this extension is an optional further direction rather than part of the completion boundary.

References

  1. Kellman BP, Sandoval D, Zaytseva OO, Brock K, Baboo S, Nachmanson D, Irvine EB, Schoffelen S, Armingol E, Mih N, Zhang Y, Jeffris M, Bartels P, Nguyen T, Tam A, Gasman S, Ilan S, Canner SW, Shamie I, Diedrich JK, Wang X, van Woudenbergh E, Altman M, Aylward A, Bao B, Castro A, Sorrentino J, Chiang AWT, Campbell M, Bartsch Y, Aguilar-Calvo P, Sigurdson C, Alter G, Lauc G, Yates JR, Voldborg BG, Marks D, Lisacek F, Lewis NE. Protein structure, a genetic constraint on glycosylation [v1 2024 titled 'Protein structure, a genetic encoding for glycosylation']. bioRxiv preprint (2024; v3 posted 2026-02-07). DOI: 10.1101/2024.05.15.594261. https://doi.org/10.1101/2024.05.15.594261
  2. Kellman BP, Mariethoz J, Zhang Y, et al. Decoding glycosylation potential from protein structure across human glycoproteins with a multi-view recurrent neural network. bioRxiv preprint (2024). DOI: 10.1101/2024.05.15.594334. https://doi.org/10.1101/2024.05.15.594334
  3. Thaysen-Andersen M, Packer NH. Site-specific glycoproteomics confirms that protein structure dictates formation of N-glycan type, core fucosylation and branching. Glycobiology 22(11):1440-1452 (2012). DOI: 10.1093/glycob/cws110. https://doi.org/10.1093/glycob/cws110
  4. Huang Y-W, Yang H-I, Wu Y-T, Hsu T-L, Lin T-W, Kelly JW, Wong C-H. Residues comprising the enhanced aromatic sequon influence protein N-glycosylation efficiency. Journal of the American Chemical Society 139(37):12947-12955 (2017). DOI: 10.1021/jacs.7b03868. https://doi.org/10.1021/jacs.7b03868
  5. Allen JD, Chawla H, Samsudin F, Zuzic L, Shivgan AT, Watanabe Y, He W, Callaghan S, Song G, Yong P, Brouwer PJM, Song Y, Cai Y, Duyvesteyn HME, Malinauskas T, Kint J, Pino P, Wurm MJ, Frank M, Chen B, Stuart DI, Sanders RW, Andrabi R, Burton DR, Li S, Bond PJ, Crispin M. Site-specific steric control of SARS-CoV-2 spike glycosylation. Biochemistry 60(27):2153-2169 (2021). DOI: 10.1021/acs.biochem.1c00279. https://doi.org/10.1021/acs.biochem.1c00279 [first posted as bioRxiv preprint 10.1101/2021.03.08.433764]
  6. Garcia-Garcia A, Serna S, Yang Z, Delso I, Taleb V, Hicks T, Artschwager R, Vakhrushev SY, Clausen H, Angulo J, Corzana F, Reichardt NC, Hurtado-Guerrero R. FUT8-directed core fucosylation of N-glycans is regulated by the glycan structure and protein environment. ACS Catalysis 11(15):9052-9065 (2021). DOI: 10.1021/acscatal.1c01698. https://doi.org/10.1021/acscatal.1c01698
  7. Simeonov DR, Wang X, Wang C, Sergeev Y, Dolinska M, Bower M, Fischer R, Winer D, Dubrovsky G, Balog JZ, et al. DNA variations in oculocutaneous albinism: an updated mutation list and current outstanding issues in molecular diagnostics. Human Mutation 34(6):827-835 (2013). DOI: 10.1002/humu.22315. https://doi.org/10.1002/humu.22315
  8. Spritz RA, Ho L, Furumura M, Hearing VJ Jr. Mutational analysis of copper binding by human tyrosinase. Journal of Investigative Dermatology 109(2):207-212 (1997). DOI: 10.1111/1523-1747.ep12319351. https://doi.org/10.1111/1523-1747.ep12319351
  9. Kellman BP, Lewis NE. Protein structure, a new constraint for glycosyltransferase specificity: Template-like association between the glycoprotein surface on glycan structure. FASEB Journal 34(S1):07467 (2020), conference abstract. DOI: 10.1096/fasebj.2020.34.s1.07467. https://doi.org/10.1096/fasebj.2020.34.s1.07467
  10. Reddy JV, Singh SK, Leibiger T, Lee KH, Ierapetritou M, Papoutsakis ET. Flux balance analysis and peptide mapping elucidate the impact of bioreactor pH on Chinese hamster ovary (CHO) cell metabolism and N-linked glycosylation in the Fab and Fc regions of the produced IgG. Metabolic Engineering 87:37-48 (2025). DOI: 10.1016/j.ymben.2024.11.005. https://doi.org/10.1016/j.ymben.2024.11.005
  11. Canner SW, Schnaar RL, Gray JJ. The proteome as a lectome: predictions from deep learning propose substantial protein-carbohydrate interplay. bioRxiv preprint (2025). DOI: 10.1101/2025.03.07.641884. https://doi.org/10.1101/2025.03.07.641884
  12. Joeres R, Bojar D. Higher-order message passing for glycan representation learning. arXiv preprint arXiv:2409.13467 (2024). DOI: 10.48550/arXiv.2409.13467. https://doi.org/10.48550/arXiv.2409.13467
  13. Pratyush P, Pokharel S, Schulze S, Bramer L, Newman RH, KC DB. Large language model (LLM)-based advances in prediction of post-translational modification sites in proteins. Methods in Molecular Biology (Springer, 'Large Language Models (LLMs) in Protein Bioinformatics'), pp. 313-355 (2025). DOI: 10.1007/978-1-0716-4623-6_19. https://doi.org/10.1007/978-1-0716-4623-6_19
  14. Gomez Toledo A, Sorrentino JT, Schoffelen S, Voldborg B, Velasquez E, Scott AM, Larson G, Lewis NE, Malmström J. Population-level analysis of glycoprotein glycoforms. mAbs 18(1):2665879 (2026). DOI: 10.1080/19420862.2026.2665879. https://doi.org/10.1080/19420862.2026.2665879
  15. Maciej-Hulme ML. Pushing frontiers for proteoglycans. Proteoglycan Research 4(2):e70046 (2026), perspective article. DOI: 10.1002/pgr2.70046. https://doi.org/10.1002/pgr2.70046
  16. Li H, Chiang W-T, Gazestani VH, Bao B, Li S, Menard P, Arnsdorf J, Dalgaard ZS, Bjorn SP, Brondum KK, Hansen AH, Schoffelen S, Voldborg BG, Lewis NE. Delineating the transcriptional and phenotypic impact from biotherapeutic glycoengineering. bioRxiv preprint (2026). DOI: 10.64898/2026.04.21.719832. https://doi.org/10.64898/2026.04.21.719832