Fold №75 — Semaglutide C-Terminal Gly-31 Truncation

Verdict: DISCARDED (tool-limit failure — lipidated peptide outside current predictor resolution)

TLDR

Fold №75 was DISCARDED because the dominant structural and pharmacokinetic determinant of semaglutide — its C18 diacid lipidation at Lys-20 — cannot be modelled with chemical fidelity by Boltz-2 or Chai-1, rendering the binding interface prediction unreliable regardless of backbone confidence scores. This is a tool-limit failure, not a biological invalidation. The underlying hypothesis — that Gly-31 is dispensable for GLP-1R engagement — is strongly supported by existing pharmacological literature (endogenous GLP-1[7-36] amide is the dominant active form and naturally lacks this residue) and remains testable by wet-lab methods.

What we tried

Semaglutide is a 31-residue GLP-1 receptor agonist based on GLP-1(7-37), distinguished from the native peptide by Aib-2 (DPP-4 resistance), Arg-34 (proteolytic stability), and a C18 fatty diacid at Lys-26 via a γGlu-γGlu-miniPEG linker conferring albumin binding and a ~165-hour plasma half-life. Fold №75 asked whether the terminal Gly-31 — corresponding to Gly-37 of proglucagon — could be deleted to yield a 30-residue analog (HAEGTFTSDVSSYLEGQAAKEFIAWLVRGR) while preserving full GLP-1R engagement.

The hypothesis was motivated by two independent lines of reasoning. First, structural biology of GLP-1R agonist complexes places the ECD contact surface at positions ≤26 of GLP-1, with the GRG tail extending into solvent without direct receptor contacts. Second, and more compellingly, the dominant circulating form of endogenous GLP-1 is GLP-1(7-36) amide — a 30-residue peptide that naturally terminates at the equivalent of semaglutide's Arg-30 and is fully pharmacologically active. The C-terminal Gly-31 in semaglutide is therefore an artefact of the GLP-1(7-37) template used in drug development, not a biological necessity for receptor engagement. A secondary hypothesis proposed that exposing C-terminal Arg-30 might modestly alter carboxypeptidase susceptibility and clearance kinetics, though the albumin-bound lipidated fraction was expected to remain largely shielded.

Why it was discarded

The structural prediction produced pLDDT 0.78, pTM 0.84, and ipTM 0.93 — numerically acceptable values that would ordinarily suggest a confident complex model. However, these metrics reflect only the backbone confidence of the unlipidated peptide sequence; they do not and cannot capture the conformational geometry of semaglutide as it actually exists in solution or in its receptor-bound state. The C18 diacid lipid chain, γGlu-γGlu-miniPEG spacer, and albumin-engaged conformation collectively define the pharmacologically relevant structure of semaglutide, and none of these chemical moieties are represented in the Boltz-2/Chai-1 prediction. The output models an unlipidated peptide backbone docked to GLP-1R — a physically unrealistic scenario that cannot support conclusions about binding affinity, interface geometry, or conformational stability of the actual drug.

This is the same tool-limit failure that discarded Folds #15 (Glu-16 → homoglutamate, pLDDT 0.71), #36 (β-Ala-β-Ala spacer replacement, pLDDT 0.70), and #52 (αMe-His at position 1, pLDDT 0.72). The pattern across four semaglutide folds is consistent: the backbone prediction scores are borderline-acceptable, but the lipidation chemistry that governs the molecule's conformational ensemble and pharmacological identity falls outside the chemical space these tools can resolve. Chai-1 agreement data was not returned for this fold, removing a secondary confidence check.

What this doesn't mean

DISCARDED does not mean disproved. This verdict reflects a tool-limit failure — the current prediction infrastructure cannot model lipidated peptides of this chemical complexity with sufficient fidelity to adjudicate binding hypotheses. The biological rationale for the Gly-31 truncation is independently supported by decades of GLP-1 pharmacology: GLP-1(7-36) amide, which endogenously lacks the terminal Gly, is the dominant active circulating form and engages GLP-1R with equivalent potency to GLP-1(7-37). No evidence in the literature contradicts the core hypothesis, and no structural data places Gly-31 within the receptor binding footprint. The carboxypeptidase concern (C-terminal Arg exposure) is a real but likely modest pharmacokinetic risk in the context of the albumin-bound lipidated molecule. The hypothesis is scientifically sound; it simply requires tools and assays better suited to the chemistry than current in silico predictors.

What would answer the question

• Competitive radioligand binding assay (GLP-1R): Direct IC₅₀ comparison of the 30-mer analog vs. full-length semaglutide using [¹²⁵I]-GLP-1 or fluorescent tracer displacement at HEK293 cells overexpressing GLP-1R — the gold-standard assay for GLP-1 analog affinity, well-established in the discovery literature (Lau et al., 2015 framework).
• cAMP dose-response (functional agonism): EC₅₀ comparison via HTRF or BRET cAMP assay to confirm that Gly-31 deletion does not impair receptor activation, independent of binding affinity.
• Plasma carboxypeptidase stability assay: Incubation of the 30-mer analog in pooled human plasma at 37°C with LC-MS/MS monitoring for des-Arg-29 and des-Arg-28 cleavage products — directly tests the C-terminal Arg vulnerability hypothesis.
• Full-atom MD simulation with explicit lipidation: Force-field parameterisation of the complete lipidated 30-mer (AMBER or CHARMM with custom GAFF2 parameters for the C18 diacid and γGlu-γGlu linker) docked to the cryo-EM GLP-1R structure, with explicit solvent and optional albumin inclusion — the appropriate computational tool for this chemistry class, as distinct from generative fold predictors.

Raw metrics

Heuristic properties are sequence-based estimates only and do not account for the C18 diacid lipidation that dominates semaglutide's actual pharmacokinetic profile.

Lab context: This is the fourth semaglutide fold discarded due to lipidation-related tool limits (see also Fold #15, Fold #36, Fold #52). The pattern strongly suggests that semaglutide modifications require either wet-lab validation or purpose-built MD workflows with explicit lipid parameterisation before in silico verdicts can be trusted. The biological hypotheses across all four folds — including this truncation — remain independently plausible.