Summary
We're using AF3 locally to model mucin-type O-glycosylated peptides (α-O-GalNAc / Tn antigen) by specifying GalNAc as separate ligand chains with bondedAtomPairs. The covalent linkage is placed, but the glycosidic φ dihedral (O5–C1–OG–CB) is unconstrained across seeds — no enforcement of the exo-anomeric effect that strictly defines α-O-GalNAc geometry in solution.
What we observe across 25 models (5 seeds × 5 samples) of a 25-residue O-glycopeptide with 14 GalNAc sites:
[Metric
φ O5–C1–OG–CB
C1–OG bond length
% models in correct α-range
Expected (α-O-GalNAc)
−40° to −80° (tight, exo-anomeric)
~1.43 Å (C–O ether)
N/A
AF3 output
−7° to −56° mean, std 36–110° per site
1.26–1.42 Å
0/14 sites ≥80%]
Input approach used:
{
"bondedAtomPairs": [
[["A", 12, "OG"], ["J", 1, "C1"]],
[["A", 15, "OG1"], ["L", 1, "C1"]]
]
}Questions:
Is there a recommended input strategy for O-linked glycosylation beyond bondedAtomPairs that better captures anomeric geometry?
Are O-glycan torsion potentials planned for a future release?
We're currently working around this with post-hoc GlycoShape refinement, but native support would be valuable for the glycobiology community.
Happy to share the full input JSON or benchmark data if useful.
Summary
We're using AF3 locally to model mucin-type O-glycosylated peptides (α-O-GalNAc / Tn antigen) by specifying GalNAc as separate ligand chains with bondedAtomPairs. The covalent linkage is placed, but the glycosidic φ dihedral (O5–C1–OG–CB) is unconstrained across seeds — no enforcement of the exo-anomeric effect that strictly defines α-O-GalNAc geometry in solution.
What we observe across 25 models (5 seeds × 5 samples) of a 25-residue O-glycopeptide with 14 GalNAc sites:
[Metric
φ O5–C1–OG–CB
C1–OG bond length
% models in correct α-range
Expected (α-O-GalNAc)
−40° to −80° (tight, exo-anomeric)
~1.43 Å (C–O ether)
N/A
AF3 output
−7° to −56° mean, std 36–110° per site
1.26–1.42 Å
0/14 sites ≥80%]
Input approach used:
{ "bondedAtomPairs": [ [["A", 12, "OG"], ["J", 1, "C1"]], [["A", 15, "OG1"], ["L", 1, "C1"]] ] }Questions:
Is there a recommended input strategy for O-linked glycosylation beyond bondedAtomPairs that better captures anomeric geometry?
Are O-glycan torsion potentials planned for a future release?
We're currently working around this with post-hoc GlycoShape refinement, but native support would be valuable for the glycobiology community.
Happy to share the full input JSON or benchmark data if useful.