# Gut-Restricted Protein Binder Design — Tier-2 Celiac Biology

### Protease-hardening and delivery-format engineering for the MICA-decoy (anti-NKG2D) and IL-15 binder

## Purpose

The celiac omics stream nominated two high-value Tier-2 targets on the villous-destruction axis: the **NKG2D–MIC** cytotoxic trigger and **IL-15**, the cytokine that arms intraepithelial lymphocytes. This session had already produced first-pass binders for both: a soluble **MICA-decoy** (sequesters NKG2D away from stressed enterocyte MIC) and an **IL-15 binder** (neutralizes the β/γc signaling epitope, like AMG-714 but in a small format). The gut-restricted therapeutics landscape review identified the **oral / protease-stable biologic** as the modality that would let these high-value targets reach the tolerability-sensitive celiac population, and as the hardest engineering problem in the field. This report takes the two binders through the sequence-level engineering that a gut-restricted format requires: mapping their protease liabilities, hardening them without touching the binding interface, scoring developability, and specifying the delivery-format path.

## What was done and what it produced

### 1. Protease-liability mapping

Both binders were scanned with PeptideCutter-style cleavage rules for the four major GI proteases — **pepsin, trypsin** (after K/R, not before P), **chymotrypsin** (after F/Y/W/L/M), and **elastase** (after A/V/G/S/I/L). Every predicted cut site was cross-referenced against the fixed binding-interface residues (**29** NKG2D-contact residues on the MICA-decoy, **12** IL-15-contact residues on the binder).

The binders are heavily protease-liable, as expected for sequences never selected for GI stability: the MICA-decoy carries cleavage sites at **154 of 275** positions, the IL-15 binder at **110 of 202**. Critically, a subset of cut sites falls *on* interface residues (19 for the MICA-decoy, 7 for the IL-15 binder) — these are unremovable without breaking binding, which is itself a finding: it sets a floor on achievable protease resistance and is one reason a purpose-built VHH scaffold (below) is preferable to hardening the current scaffold indefinitely.

### 2. Interface-preserving protease-hardening

Hardening was done by re-running SolubleMPNN with the binding interface held fixed and the cleavage-prone residue types forbidden outside the paratope (`omit_AAs = "KRFWYM"`, eliminating trypsin's K/R and the chymotrypsin/pepsin aromatics F/W/Y/M while leaving the aliphatic residues the fold needs, and retaining cysteines for disulfide engineering). This lets the network choose structurally-compatible protease-resistant residues rather than making blind point substitutions.

On the harden-able proteases (trypsin + chymotrypsin + pepsin), the result was a **46% reduction for the MICA-decoy (123 → 67 sites)** and a **49% reduction for the IL-15 binder (85 → 43 sites)**. Trypsin sites — the most relevant for the K/R-rich, absorption-prone liability, were nearly eliminated (**32 → 5** on the MICA-decoy, **18 → 1** on the IL-15 binder; the residual few are the locked interface K/R). All **29** and **12** interface residues were preserved exactly (confirmed at the sequence level).

The one trade-off, shown honestly in the figures: **elastase sites rise**, because forbidding basic and aromatic residues pushes the composition toward the small aliphatic residues (A/V/G/S/I/L) that elastase targets. Elastase liability is intrinsic to any folded protein chain and cannot be engineered away at the sequence level; it is managed at the format level (a compact, disulfide-locked VHH resists elastase structurally).

### 3. Developability scorecard

Native scaffold, SolubleMPNN design, and protease-hardened variant were scored on protease sites, cysteine count, net charge, aromatic content, and hydropathy. Two properties of the hardened designs matter for gut-restriction:

- **Net charge became strongly negative** (MICA-decoy −42, IL-15 binder −26). This is a *bonus* for the gut-restriction goal (a highly charged molecule does not cross an intact epithelial membrane, reinforcing non-absorption), but −42 is extreme and must be checked against folding, solubility, and interference with the (charged) binding interface.
- **Structural cysteines were retained** (5 on the MICA-decoy, 3 on the IL-15 binder) after an initial cysteine-free variant was rejected. These are the anchor points for the disulfide lock that gives a gut-format its proteolytic and thermal stability.

### 4. Delivery-format design specification

A six-step reformatting path was specified for each binder: (1) minimize to the folded core bearing the paratope; (2) **reformat into a VHH / single-domain antibody** (~13 kDa) by grafting the fixed epitope onto a stable nanobody framework, the reference gut-stable format, precedented by the oral anti-TNF VHH **V565** that reached the inflamed gut lumen intact; (3) apply this campaign's interface-preserving hardening to the framework; (4) **disulfide-lock** (the canonical VHH Cys22–Cys92 plus retained structural cysteines); (5) **mucin-anchor** via a mucus-binding module to prolong mucosal residence.

The **target-depth assessment** is the decisive design input. NKG2D–MIC operates at the **apical epithelial surface** (MIC is a stress ligand on the enterocyte surface; the killing synapse is at the surface), so a luminally-delivered decoy reaches it: **favorable** for gut-restriction. IL-15 is **trans-presented at the epithelial surface** but may also have a **submucosal pool**; the surface fraction is reachable but the deep pool is a caveat, giving a **moderate** fit. This is the same surface-versus-deep rule that governs the whole gut-restriction field.

## Validation status (honest)

This is a **CPU-only, sequence-and-structure design campaign**. What is established: the interface is preserved exactly, the protease-site reductions are real against the stated cleavage rules, and the developability shifts are computed from the actual sequences. What is **not** yet done, and requires GPU:

- **No structure prediction or co-folding.** The hardened sequences were not folded (ESMFold/AlphaFold) nor co-folded against NKG2D or IL-15 (Boltz-2/Chai-1), so binding retention is inferred from interface preservation rather than scored by ipTM.
- **The VHH graft is specified, not modeled.** The epitope residue set and grafting path are defined; the folded nanobody requires RFdiffusion/AlphaFold-multimer.
- **Protease rules are predictions, not digestion.** PeptideCutter-style rules approximate specificity; empirical stability needs an in-vitro simulated-gastric/intestinal-fluid digest.
- The **very negative net charge** aids non-absorption but is unvalidated for folding/solubility.

## GPU next-steps (in priority order)

1. Fold the hardened sequences (ESMFold2) and confirm the scaffold fold survived hardening.
2. Co-fold each hardened binder against its target (Boltz-2 / Chai-1) and filter on ipTM to confirm the interface still engages.
3. Execute the VHH graft (RFdiffusion motif-scaffolding on the fixed epitope) and co-fold the nanobody against target.
4. In-silico simulated-gastric/intestinal-fluid digestion to validate the predicted protease resistance.
5. Charge/solubility optimization pass if folding is compromised by the net-negative charge.

## Deliverables

| Artifact | Contents |
|----------|----------|
| celiac_binder_protease_map.csv | Per-position cleavage sites (4 proteases) + interface flags, both binders |
| celiac_binder_hardened.csv / .fasta | Protease-hardened sequences, interface-intact, per-protease site counts |
| celiac_binder_developability.csv / _summary.csv | Native vs design vs hardened scorecard + trade-off flags |
| celiac_binder_delivery_spec.json | Six-step VHH / mucin-anchor reformatting path + target-depth assessment |
| celiac_binder_protease_map_fig.png | Protease-liability map with interface residues |
| celiac_binder_hardening_fig.png | Before/after cleavage sites per protease |
| celiac_binder_format_schematic.png | Gut-restricted reformatting path schematic |

---
*Companion to the gut-restricted therapeutics landscape review and the Tier-2 protein design campaign. All designs are computational and require experimental validation.*