Machine learning (ML) approaches for de novo antibody design generate thousands of candidate sequences, but most lack a robust assessment of post-translational modification liabilities. N-linked glycosylation in variable domains can fundamentally alter antibody function, yet current penetrance estimates are derived from glycosylation-enriched datasets inflating risk estimates. We surveyed 19,265 human antibody structures from the Protein Data Bank (PDB), identifying 1,368 N-X-S/T sequons in variable domains (this study) with 7.82% observed penetrance (107/1,368; Wilson 95% CI 6.5–9.4%) – a crystallographic lower bound reflecting selection against glycosylated structures in the PDB – compared with 16.47% from a previously published glycosylation-enriched corpus. At the X-position, no glycosylation was observed at histidine (0/61; Wilson 95% upper bound 5.9%), lysine (0/26; upper bound 12.9%), or tryptophan (0/26), consistent with suppression. Glutamine showed 6.45% penetrance (2/31), indistinguishable from baseline and refuting its prior classification as a suppressor. Variable light domains showed higher aggregate penetrance than variable heavy (9.62% vs 6.79%; Fisher OR = 1.46, p = 0.075), though this contrast did not survive joint Bayesian adjustment. Regionally, FR2 was glycosylation-resistant (0/119) while FR1 (14.0%) and FR4 (12.1%) showed elevated rates. N-X-T sequons were 3.84-fold more susceptible than N-X-S in aggregate (13.64% vs 3.55%; Fisher exact p = 8.13 × 10−12, OR = 4.29). An independent PDB-overlap-free validation subset (n = 449, 33 glycosylation events, 189 nonoverlapping PDBs) reproduced the central findings. Leakage-corrected Bayesian logistic regression confirmed the N-X-T preference as the most prior-stable effect and identified strong negative coefficients for proline and aromatic residues at the +3 position immediately C-terminal to the sequon; the histidine effect at the X-position was borderline under prior sensitivity. We present a confidence-interval-aware decision tree that integrates X-position, third-position, chain type, and regional context as a sequence-only triage instrument for ML-designed antibody candidates.
Christopher L. Gaughan (Tue,) studied this question.