Skip to content

comp-034 — Physics ΔΔG + structure-gated cleavage on the inter-lobe linker

Two orthogonal methods added to the comp-034 concordance gate, unblocked by the PyRosetta install of 2026-05-29 (UW academic-license update). They answer a question the original five sequence/proxy metrics could not see: what does each linker redesign do to the actual fold, and is the WT linker even cleavable given its structure?

The five original metrics (ESM2 pseudo-pLDDT, sequence cleavage score, CAI, loop-pLDDT, WT-similarity) are all sequence- or proxy-based. Neither a physics ΔΔG on the 3-D structure nor a real solvent-accessibility/secondary-structure gate on the cleavage score existed in the gate. Both are added here.

TL;DR

NEEEQQQEEEQ (the MPNN-native arm) wins on both axes and is the wet-lab pick. The proline-rigidification arms are self-defeating. One variable explains the whole result: inter-lobe helix retention. The WT linker (UniProt 353–363, SEEEVAARRAR) is a structured α-helix (9 of 11 residues helical). Keep the helix and you keep fold stability and protease resistance (proteases require an extended substrate; a helix resists). Break it with proline and you lose both.

Candidate linker Cartesian ΔΔG (REU) structure-gated cleavage vs WT helix verdict
WT SEEEVAARRAR 0.00 (ref) — 0.818 reference
MPNN NEEEQQQEEEQ 8 subs +0.23 (neutral) −66% 0.818 wins both
MPNN NEEEQEEQDQQ 8 subs +2.39 (neutral) −66% 0.818 wins both
S353E+V357P 2 subs +21.26 (destabilizing) −17% 0.727 loses both
V357P 1 sub +20.11 (destabilizing) −9% 0.727 loses both
multi-proline EEEEPAAPPAP 6 subs +57.48 (severe) −3% 0.364 loses both, worst

Method 1 — Cartesian ΔΔG (fold stability)

Relaxed-neighborhood ΔΔG: thread the linker mutations onto the AlphaFold structure (AF-P02788-F1-model_v6.pdb, full 710-residue monomer), FastRelax the linker neighborhood (residues within 8 Å of 353–363; 23 movable residues) in Cartesian space (ref2015_cart + cart_bonded, LBFGS), coordinate- constrained to the start coords. 10 independent-seed trajectories per candidate; minimum-energy trajectory is the estimator (correct for a stochastic relax search); ΔΔG = min(mutant) − min(WT). Script: rosetta_ddg.py. This follows the Park et al. 2016 Cartesian-ddG method already cited in protein-engineering-strategy.md §5.2.

Result: both MPNN charge/polar arms are statistically indistinguishable from WT (ΔΔG +0.23 / +2.39 REU, helix retained at 0.818). Every proline arm is destabilizing: +20 REU (single V357P), +21 REU (S353E+V357P), +57 REU (multi-proline, helix collapsed to 0.364).

Estimator note (load-bearing). Per-trajectory spread is large (sd ≈ 30 REU): by the mean, NEEEQQQEEEQ even scores "more stable than WT" (−8.31), which is sampling noise. By the minimum it is a clean +0.23. The WT min-energy basin (−736.8 REU) was independently re-found on trajectories 1, 4, and 8 of 10, so min-of-10 is converged — more trajectories would not move it. (Cross-method caveat: a first-pass torsion-space, mean-based run was noisy enough to mis-rank NEEEQEEQDQQ above V357P; it motivated the upgrade to Cartesian-min and is retained only as rosetta_ddg_results_torsion3.json for genealogy. It is not an independent confirmation.)

Method 2 — Structure-gated cleavage (protease axis)

The existing comp-005/034 cleavage model (lib/protease_stability.py) uses pLDDT as its accessibility proxy: classify_accessibility(mean_plddt) → buried (≥80, 0.1×) / partial / exposed. The inter-lobe linker has pLDDT 93–98, so all 11 residues are scored "buried" → 0.1× protection. But pLDDT is a confidence score, not burial — a confidently-predicted solvent-exposed helix scores 95. Real SASA (Tien 2013 relative) on the structure: 8 of 11 linker residues are solvent-exposed, 3 partial, 0 buried. The proxy mis-classifies an exposed helix as buried, suppressing the linker's sequence-driven cleavage risk ~10×.

The real protection is conformation, not burial: proteases need an extended substrate across the active-site cleft (Tyndall et al. 2005), so a helix resists even when solvent-exposed. The re-gated score replaces the pLDDT proxy with real SASA × a secondary-structure conformation factor (helix 0.2×, strand 0.7×, loop/turn 1.0×). Computed first on the WT backbone (structure_gated_cleavage.py) and then on each mutant's own relaxed structure (refold_via_relax.py — thread → Cartesian-relax → dump pose → real per-residue SASA + DSSP), which removes the WT-backbone approximation.

Result (real mutant structures): the MPNN arms cut structure-gated cleavage ~66% vs WT (they strip preferred residues and keep the helix). The proline arms deliver only −9% (V357P), −17% (S353E+V357P), and −3% (multi-proline): breaking the protective helix re-exposes the backbone and largely cancels the sequence-preference gain. The proline-rigidification strategy is self-defeating on this target.

Two corrections vs the first-pass approximation (intellectual honesty)

The WT-backbone approximation (structure_gated_cleavage.py) overstated two things; real mutant structures (refold_via_relax.py) corrected both:

  1. Multi-proline is not worse than WT. The approximation put it at 1.624 (above WT's 1.155) because it assumed too much helix loss (0.273). On the real relaxed structure it retains more helix (0.364) and lands at 1.104 — roughly WT-equivalent, not worse. Honest reading: 6 substitutions buy ~3% net protease benefit. Damning for an "aggressive" design, but "no better than WT."
  2. NEEEQQQEEEQ's cleavage edge is ~3×, not ~9×. The approximation (0.134) borrowed the WT backbone's smaller side chains; the real Glu/Gln-loaded structure is more exposed (0.388). Still decisively best — by an honest margin.

Wet-lab implication (revises the §1.10 arm recommendation)

The original comp-034 framing treated the proline single/double mutants as the "conservative / safe regulatory story" and NEEEQQQEEEQ as the "aggressive" arm. The physics inverts that risk ordering. NEEEQQQEEEQ is the most stable (ΔΔG ≈ 0) and most protease-resistant (−66%) candidate, with the helix intact — it should be the primary wet-lab arm, not the aggressive one. The proline arms are destabilizing and deliver little protease benefit; they are the higher- risk choices and should be demoted or dropped. Recommended plate: WT control + NEEEQQQEEEQ (primary) + NEEEQEEQDQQ (sibling/backup), with a proline arm retained only if an orthogonal-mechanism control is wanted.

Caveats

  • Evidence level: Mechanistic Extrapolation (in silico). Wet-lab proteolysis
  • thermal-stability (Tm) assays are the validators.
  • Conformation-gate weights (helix 0.2×) are heuristic; the direction is robust (helix-breaking raises accessibility; MPNN arms dominate under any reasonable weighting), the absolute structure-gated numbers are not precise.
  • ΔΔG is single-AlphaFold-backbone; no backbone-conformational ensemble, no explicit solvent, no entropy term. Multi-mutant additivity is not assumed (full relax of each mutant), but ref2015 force-field error (~1–2 kcal/mol class) dominates the residual, not sampling.
  • A genuinely orthogonal ML predictor (ESMFold/AF2) was blocked on this host (openfold + deepspeed not installed; hard Apple-Silicon/CPU build). Method 2's "real structures" come from Rosetta relax, not an independent fold. An independent re-fold belongs on the GPU/cloud path (bio-ai-tools A4 backlog).

Reproduce

cd etc/experiments/comp-034-lactoferrin-linker-redesign/rosetta_concordance
python3 rosetta_ddg.py --ntraj 10 --out rosetta_ddg_results_cartesian10.json   # ~2.1 h CPU
python3 structure_gated_cleavage.py                                            # seconds
python3 refold_via_relax.py                                                    # ~38 min CPU

Requires PyRosetta (pyrosetta 2026.3, installed 2026-05-29 under the Rosetta & PyRosetta Non-Commercial License). See provenance.md for tool versions, inputs, and the pre-commit verification record.

Artifacts

  • rosetta_ddg_results_cartesian10.json — gold-standard ΔΔG (min/mean/sd/all-traj)
  • rosetta_ddg_results_torsion3.json — noisy first-pass (genealogy only)
  • structure_gated_cleavage_results.json — WT-backbone structure-gated cleavage
  • refold_via_relax_results.json — structure-gated cleavage on real mutant structures
  • relaxed_mutant_poses/*.pdb — 6 dumped relaxed mutant structures