# Designed-Binder Structure Validation Protocol (GPU-conditional)

**Status:** SPECIFICATION ONLY — no GPU host available at execution (list_compute empty).
Run this pipeline once a remote GPU host is added (Customize → Compute).

## Pipeline

### Stage 1 — De novo backbone generation (Step 5 continuation)
- **Tool:** RFdiffusion (or BindCraft) targeting the mapped epitope hotspots.
- **SIGLEC6:** condition on Ig-V hotspots R122/R100/R109-114 (functional) or E28/E37 (non-competitive).
- **IL1RL1:** condition on Ig2-Ig3 core Y119/F245/L308/K22/R38 (glyco-free sub-patch).
- **CCL26:** condition on N-loop/40s-loop docking surface.
- Generate 1,000–5,000 backbones/target; ProteinMPNN sequence design (8 seqs/backbone).

### Stage 2 — Complex structure prediction / refold
- **Tool:** ESMFold2-Fast (single-sequence, fast triage) → Chai-1 or AlphaFold-Multimer (final ranking).
- Refold each designed binder–target complex.
- **GPU:** ≥24 GB VRAM (Chai-1/AF-M); ESMFold2-Fast triage runs on 16 GB.

### Stage 3 — Interface scoring & acceptance thresholds
| Metric | Accept | Strong |
|--------|--------|--------|
| ipTM (interface predicted TM) | > 0.6 | > 0.8 |
| pAE at interface (Å) | < 10 | < 5 |
| predicted DockQ | > 0.23 (acceptable) | > 0.49 (medium) |
| Binder pLDDT | > 80 | > 90 |
| Epitope recapitulation (designed contacts ∩ mapped hotspots) | ≥ 50% | ≥ 75% |
| Rosetta/PyRosetta ddG (if run) | < −30 REU | < −45 REU |

### Stage 4 — Filtering & ranking
1. Drop designs failing ipTM > 0.6 OR epitope recapitulation < 50%.
2. Rank surviving designs by composite (ipTM + epitope-recap − normalized pAE).
3. Cross-check developability on designed binder sequence (reuse Step 4 liability scan: N-glyc, deamidation, oxidation, aggregation).
4. Top 10–20 designs/target → Step 8 immunogenicity screen.

### Compute estimate
- RFdiffusion: ~5,000 backbones × 3 targets ≈ 4–8 GPU-hours.
- ProteinMPNN: minutes.
- ESMFold2-Fast triage: ~1 s/complex → thousands in <1 GPU-hour.
- Chai-1 final ranking (top 200/target): ~2–4 GPU-hours.
- **Total ≈ 1 GPU-day on a single 24 GB card (e.g. A10/L40/A100).**

### Deliverables when run
- binder_validation.csv (per-design metrics)
- top-ranked complex structures (.pdb/.cif, Mol*-viewable)
- design_ranking.png