FOLD №78 — Humanin S14Dab/L10D Lactam Staple — DISCARDED

Verdict: DISCARDED (tool-limit failure) — Not biologically disproved. The structural prediction pipeline could not model the Dab non-canonical residue or the covalent lactam crosslink with sufficient confidence to yield interpretable results.

TLDR

FOLD №78 was DISCARDED due to a tool-limit failure. The intramolecular lactam bridge between Asp-10 and Dab-14 requires modeling a non-proteinogenic amino acid (L-2,4-diaminobutyric acid) and a covalent macrocyclic constraint that lies outside the training distribution and atom-type vocabulary of the Boltz-2/AlphaFold-class predictors used here. The resulting pLDDT of 0.44 and ipTM of 0.12 do not reflect the peptide's biological potential — they reflect the model's inability to encode the chemistry. This is a tool-limit discard, not a biological invalidation.

What we tried

Building on the two PROMISING Humanin disulfide-staple folds in this lab — Fold #22 (i,i+6 Cys-8/Cys-14, pLDDT 0.56) and Fold #66 (i,i+4 Cys-7/Cys-11, pLDDT 0.56) — this fold tested whether replacing the redox-sensitive disulfide staple with a chemically irreversible lactam bridge would sharpen the predicted helical geometry and BAX-interface confidence. The chemistry: native Leu-10 was replaced with L-Asp (to provide a β-carboxylate electrophile), and native Ser-14 was replaced with L-2,4-diaminobutyric acid / Dab (to provide a γ-amine nucleophile). A covalent amide bond between these two side chains creates an i,i+4 macrocyclic lactam spanning exactly one helical turn — the textbook helix-nucleating geometry (Felix, Houston) and the same spacing used in published SAHB analogs of BH3 peptides targeting BCL-2 family grooves.

The scientific rationale was sound: a redox-stable amide bridge cannot be reduced by the cytosolic glutathione pool that would cleave a disulfide, making this staple more appropriate for the intracellular BAX target where Folds #22 and #66 would be vulnerable. Position 14 is known to tolerate modification — the S14G substitution (HNG) is a well-documented gain-of-function analog — and position 10 is predicted solvent-facing in helical models. The prediction target was that the 7–15 segment would achieve local pLDDT > 0.65 and that the interface ipTM would match or exceed the 0.56 baseline from the disulfide folds.

Why it was discarded

The fold returned pLDDT 0.44, pTM 0.35, and ipTM 0.12 — a substantial degradation relative to the disulfide folds at all three metrics. The most likely cause is that Dab is not a proteinogenic amino acid and its γ-amine is not encodable as a canonical residue in Boltz-2's sequence vocabulary or scoring function. The covalent lactam crosslink further imposes a macrocyclic backbone constraint that the predictor cannot represent: it uses a standard open-chain atom graph. The model therefore attempts to fold the peptide as if the staple constraint does not exist, producing a physically incorrect, disordered ensemble and inflating positional uncertainty across the entire sequence. This is the same class of failure seen in Fold #59 (N-terminal myristoylation, FAILED), where a non-canonical chemical modification overwhelmed the predictor's representations.

By contrast, Folds #22 and #66 installed cysteine residues — canonical amino acids — and relied on the predictor's implicit treatment of disulfide bonds, which are encoded in training data. That approach yielded pLDDT 0.56 in both cases, sufficient for a PROMISING verdict. The lactam staple trades chemical stability for computational tractability.

An additional biological concern worth flagging: the Leu-10 → Asp substitution introduces a charged, hydrophilic side chain into what is likely a hydrophobic region of the HN–BAX interface. Even if the staple were correctly modeled, this substitution alone could disrupt native hydrophobic contacts with the BAX BH3-binding groove. The double modification (L10D + S14Dab) was necessary to install the staple but creates a compound perturbation that has not been evaluated.

What this doesn't mean

DISCARDED is not "disproved." This fold was discarded because the current prediction pipeline lacks the atom-type vocabulary and covalent crosslink representation required to model a Dab-containing lactam-stapled peptide — not because the peptide was predicted to be inactive, misfolded, or non-binding. The scientific hypothesis (that an amide staple would confer greater helical stability and redox stability than the disulfide staples from Folds #22 and #66) is well-grounded in published chemistry and remains entirely open. The 0.44 pLDDT score is an artefact of tool mismatch, not a signal about the peptide's actual helical propensity or BAX affinity. Lactam-stapled α-helical peptides targeting BCL-2 family hydrophobic grooves have repeatedly shown enhanced binding affinity, proteolytic stability, and cellular activity in published literature — the concept is sound. What is missing is the right computational tool to evaluate this specific instance.

What would answer the question

• Molecular dynamics / FEP with parameterized non-standard residues: OpenMM or GROMACS with a custom GAFF2 or CHARMM CGenFF parameter set for Dab and the Asp–Dab lactam bond could model the covalent constraint explicitly and generate helicity, stability, and binding free energy predictions far more reliably than AlphaFold-class models for this chemistry. FEP+ (Schrödinger) supports macrocyclic peptides with non-canonical residues and would be the gold-standard computational approach.
• Circular dichroism (CD) spectroscopy: Synthesizing the peptide (Fmoc SPPS with Dab Fmoc protection and on-resin cyclization) and measuring helical content in buffer ± TFE would directly confirm whether the lactam staple enforces the expected α-helical conformation — a critical preliminary validation before any binding assay.
• Surface plasmon resonance (SPR) or isothermal titration calorimetry (ITC) with recombinant BAX: Direct binding affinity measurement (K_D, ΔH, ΔS) against monomeric BAX protein would determine whether the stapled analog binds more tightly than native HN or the disulfide analogs. Guo et al. (2003) demonstrated that cell-free binding assays with isolated mitochondria are feasible and sensitive for HN/BAX interactions.
• Cellular MOMP inhibition assay: A cytochrome c release assay in a BAX-dependent apoptosis model (e.g., BAX/BAK double-knockout MEFs reconstituted with BAX) would determine whether the lactam-stapled analog retains or enhances functional anti-apoptotic activity relative to native HN and the S14G (HNG) benchmark, while also reporting on cell permeability of the stapled peptide.

Raw metrics

Note: heuristic profile values are sequence-based estimates, not derived from the 3D prediction, and do not reflect the covalent lactam constraint.