glycoengineering – Other skill for AI agents

Glycoengineering

When to use

Use for canonical N-glycosylation candidate analysis, O-GalNAc candidate triage,

antibody glycoform comparison, or structural glycan shielding. Keep four kinds of

result separate: sequence motif, prediction, experimentally supported occupancy,

and glycan structure/composition. None alone establishes all the others.

This workflow focuses on eukaryotic secretory-pathway N-glycosylation and mucin-type

O-GalNAc. O-GlcNAc, O-mannose, O-fucose and other O-linked modifications require

appropriate evidence and predictors; S/T enrichment does not identify the type.

Workflow

1. Record the protein accession, isoform/version, exact sequence, expression host,

signal peptide/transmembrane topology and construct boundaries. Preserve a

mapping from submitted sequence coordinates to mature protein, PDB chain/residue

identifiers and antibody numbering where relevant.

2. Scan canonical N-X-[S/T] candidates with X not Pro. Retain overlaps: NNST

has candidates starting at 1 and 2. A missing canonical motif does not rule out

unusual N-glycosylation; a present motif does not establish occupancy.

3. Prioritize secreted/luminal/extracellular regions using topology and curated

evidence. Sequence-only results in a cytoplasmic region are not evidence of

secretory-pathway glycosylation.

4. For O-GalNAc, use S/T density only as a descriptive feature, then inspect a

suitable predictor and cell/transferase context. Do not exclude SP/TP motifs:

adjacent Pro can be favorable, depending on the GalNAc-transferase and context.

See [the GalNAc-T substrate study](https://doi.org/10.1093/glycob/cwu089).

5. Propose explicitly numbered mutations, list every changed residue and rescan

the full product for lost/created overlapping motifs. Evaluate structural and

expression effects independently of glycosylation.

6. Validate occupancy and glycoforms experimentally, with localization evidence,

glycan composition/structure confidence, and the biological assay required by

the engineering objective. For biosimilar comparisons, match sample handling,

analytical coverage and quantification before comparing glycan percentages.

Local sequence analysis

The standard-library helper [scripts/glycoengineering_tools.py](scripts/glycoengineering_tools.py)

replaces copied snippets. Import it with this skill's scripts/ directory on

PYTHONPATH, or run Python from that directory. It accepts raw canonical

20-amino-acid sequences, normalizes whitespace/case, and rejects FASTA headers,

gaps, unknown residues and empty input. Parse FASTA records first; do not remove

unknown residues because that changes coordinates. Coordinates are 1-based in the

normalized submitted sequence.

from glycoengineering_tools import (
    normalize_sequence, find_n_glycosylation_sequons,
    eliminate_glycosite, add_glycosite, find_st_rich_sites,
)

sequence = normalize_sequence("NNST")
sites = find_n_glycosylation_sequons(sequence)
assert [site["position"] for site in sites] == [1, 2]

mutant = eliminate_glycosite(sequence, 1, "Q")
assert mutant == "QNST"  # the overlapping sequon at position 2 remains
assert [site["position"] for site in find_n_glycosylation_sequons(mutant)] == [2]

# Three intended changes: A1N, P2A, A3T. P->A must be explicit.
parent = "APA"
product = add_glycosite(parent, 1, "T", allow_proline_substitution=True)
changes = [(i, before, after) for i, (before, after) in
           enumerate(zip(parent, product), 1) if before != after]
assert product == "NAT" and len(changes) == 3

# Density is a fraction of residues, not an O-glycosylation probability.
o_candidates = find_st_rich_sites("STPST", window=7, min_st_fraction=0.4)
assert [site["position"] for site in o_candidates] == [1, 2, 4, 5]

eliminate_glycosite requires a complete canonical sequon and a one-residue

replacement other than N. add_glycosite can alter up to three residues; it retains

an existing S/T at +2. Neither function predicts a mutant's fold, glycan occupancy,

function or tolerability. An N-to-Q substitution can change protein behavior even

without a glycan effect.

The S/T density window must be a positive odd integer. Terminal windows shorten,

and their denominator is the actual window length. Dense regions may help triage

mucin-like sequence; isolated real O-GalNAc sites can be missed.

For batch analysis, keep identifiers and normalized sequence lengths:

sequences = {"overlap": "NNST", "control": "APAA", "mucin_like": "STPST"}
rows = []
for name, raw_sequence in sequences.items():
    seq = normalize_sequence(raw_sequence)
    positions = [site["position"] for site in find_n_glycosylation_sequons(seq)]
    rows.append({"protein": name, "length": len(seq),
                 "n_sequon_positions": positions,
                 "n_sequons_per_100_residues": 100 * len(positions) / len(seq),
                 "st_rich_positions": [site["position"] for site in find_st_rich_sites(seq)]})

Prediction services

Use the official forms, which accept FASTA and return web results; this skill does

not provide a stable submission API or job-polling endpoint. Saving a CGI URL is

not a submitted job. Preserve the output, input and service version together.

| Service | Current documented scope and interpretation |

| --- | --- |

| [NetNGlyc 1.0](https://services.healthtech.dtu.dk/services/NetNGlyc-1.0/) | Human N-glycosylation; reports network potential and jury agreement. Default threshold 0.5; not a calibrated occupancy probability. Up to 2,000 sequences / 200,000 residues total / 4,000 per sequence. SignalP runs, but extracellular topology is not checked. The server may score N-P-S/T; exclude these from the canonical candidate set unless independent evidence warrants review. |

| [NetOGlyc 4.0](https://services.healthtech.dtu.dk/services/NetOGlyc-4.0/) | Mammalian mucin-type O-GalNAc; outputs GFF2 confidence scores, with scores greater than 0.5 marked positive. Up to 50 sequences / 200,000 residues total / 4,000 per sequence. Prefer full sequence including signal peptide; isolated sites need 15 residues of flanking context on both sides. A positive supports regional likelihood, not guaranteed site occupancy or glycan type beyond this model's O-GalNAc scope. |

The service documentation and sample output were reviewed; new prediction jobs

were not submitted during this refresh.

Structural shielding with GlycoSHIELD

[GlycoSHIELD](https://gitlab.mpcdf.mpg.de/dioscuri-biophysics/glycoshield-md/)

grafts pre-simulated glycan conformers onto protein structures and filters steric

clashes. It does not predict which sequons are occupied or run fresh molecular

dynamics for each query. See [Tsai et al., 2024](https://doi.org/10.1016/j.cell.2024.01.034).

Read [references/glycoshield.md](references/glycoshield.md) for the pinned source,

installation, direct Python API, input mapping and SASA analysis. **The reviewed

upstream CLI silently ignores several parsed options, including --mode,

--threshold and --dryrun; use the documented direct API workaround.** Outputs

include per-site PDB/XTC ensembles and shielding encoded in a PDB B-factor column;

those values are not crystallographic temperature factors.

Database evidence and glycan notation

Read [references/glycan_databases.md](references/glycan_databases.md) for the live

GlyTouCan SPARQL lookup, GlyConnect access limitations, curated resources and

notation. Query exact accessions and preserve dataset/source dates, species,

tissue/cell context and supporting publications. Missing records or service errors

cannot establish that a protein is unglycosylated.

A monosaccharide composition such as Hex:5 HexNAc:4 dHex:1 does not resolve

linkages, branch positions, or distinguish GlcNAc from GalNAc. A cartoon without

linkages is schematic, not IUPAC condensed notation. Use an actual sequence

(WURCS/GlycoCT/IUPAC with its uncertainty retained) and accession when available.

Core fucose attaches to the innermost GlcNAc of an N-glycan, not to core mannose.

Antibody engineering decisions

Number Fc mutations in an explicit scheme (commonly EU), then map them to the

actual construct. EU N297 is not residue 297 of an isolated Fc FASTA. Fc glycans

and any Fab glycans must be distinguished analytically.

| Objective | Candidate strategy | Required interpretation |

| --- | --- | --- |

| Increase FcγRIIIa engagement / ADCC | Reduce Fc core fucosylation while preserving the glycan | Effect size depends on antibody, receptor and assay; do not assume a universal fold gain. [Structural evidence](https://pubmed.ncbi.nlm.nih.gov/22023369/). |

| Remove canonical Fc N-glycosylation | N297Q/A/D or disrupt the +2 residue with T299A | Sequon disruption; not a guarantee of otherwise unchanged structure or effector function. Verify the actual product. |

| Alter serum persistence | Characterize glycoform-specific clearance with the relevant protein | IgG high-mannose clearance ca

… (full content is in the download package)