# Reproducibility Guide

## Complete Reproduction of APED Results

### Prerequisites

- Python 3.10+
- Google account (for Colab notebooks)
- ~5GB disk space

### Step 1: Environment Setup

```bash
git clone https://github.com/electricsheepafrica/APED.git
cd APED
python -m venv venv
source venv/bin/activate
pip install -r reproducibility/requirements.txt
```

### Step 2: Download AlphaFold Structures

```bash
python src/data/download_alphafold_structures.py \
    --taxonomy-ids 274,69014,2287,2261 \
    --output-dir data/structures/alphafold_db/
```

**Expected output:** 981 PDB files

### Step 3: Extract Features

```bash
python src/features/extract_features.py \
    --input-dir data/structures/alphafold_db/ \
    --output data/ml_ready/aped_ml_dataset.parquet
```

**Expected output:** Parquet file with 500 proteins, 19 features

### Step 4: Novel Protein Design (Colab)

1. Open: https://colab.research.google.com/github/sokrypton/ColabDesign/blob/main/rf/examples/diffusion.ipynb

2. Upload motif: `data/design/motifs/P56930_motif.pdb`

3. Set parameters:
   - name: `aped_design`
   - contigs: `A1-50/50-100`
   - iterations: `50`
   - num_designs: `5`

4. Run all cells

5. Download results ZIP

### Step 5: Validation

Filter designs by:
- pLDDT ≥ 90%
- RMSD < 1.0Å

```python
import pandas as pd
df = pd.read_csv('mpnn_results.csv')
best = df[(df['plddt'] >= 0.9) & (df['rmsd'] < 1.0)]
print(f"Passing designs: {len(best)}")
```

**Expected:** 3 designs passing criteria

### Checksums

Verify data integrity:

```bash
md5sum data/ml_ready/aped_ml_dataset.parquet
# Expected: [checksum will vary based on exact data]

wc -l data/design/sequences/best_designs.fasta
# Expected: 10 lines (5 sequences with headers)
```

### Software Versions

| Software | Version | Source |
|----------|---------|--------|
| RFdiffusion | 1.1.1 | ColabDesign |
| ProteinMPNN | 1.0.1 | ColabDesign |
| AlphaFold | 2.3.0 | ColabFold |
| Python | 3.10+ | - |

### Colab Notebooks Used

| Notebook | Purpose | Link |
|----------|---------|------|
| diffusion.ipynb | Backbone generation | [ColabDesign](https://colab.research.google.com/github/sokrypton/ColabDesign/blob/main/rf/examples/diffusion.ipynb) |

### Random Seeds

RFdiffusion uses stochastic sampling. Results may vary slightly between runs. Our reported results used default Colab seeds.

### Hardware

- **Dataset construction:** Any CPU (tested on Intel i7)
- **RFdiffusion/ProteinMPNN:** Google Colab T4 GPU (free tier)
- **Approximate runtime:** 
  - Structure download: ~30 min
  - Feature extraction: ~10 min
  - RFdiffusion (5 designs): ~15 min
  - ProteinMPNN + validation: ~10 min

### Troubleshooting

**Issue:** AlphaFold DB download fails
**Solution:** Check internet connection; some structures may be unavailable

**Issue:** Colab disconnects during RFdiffusion
**Solution:** Reduce num_designs or use Colab Pro

**Issue:** Different pLDDT scores
**Solution:** Minor variations expected; ensure same model weights

### Contact

For reproducibility questions: [Electric Sheep Africa](https://huggingface.co/electricsheepafrica)