adaptyvbio/protein-design-skills/protein-design-workflow

Target Preparation --> Backbone Generation --> Sequence Design
         |                     |                     |
         v                     v                     v
    (pdb skill)          (rfdiffusion)         (proteinmpnn)
                               |                     |
                               v                     v
                        Structure Validation --> Filtering
                               |                     |
                               v                     v
                         (alphafold/chai)      (protein-qc)

# Download from PDB
curl -o target.pdb "https://files.rcsb.org/download/XXXX.pdb"

# Extract target chain
# Remove waters, ligands if needed
# Trim to binding region + 10A buffer

# RFdiffusion runs from the official repo, not biomodals
python run_inference.py \
  inference.input_pdb=target_prepared.pdb \
  contigmap.contigs=[A1-150/0 70-100] \
  ppi.hotspot_res=[A45,A67,A89] \
  inference.num_designs=500

modal run modal_bindcraft.py \
  --input-pdb target_prepared.pdb \
  --target-hotspot-residues "45,67,89" \
  --number-of-final-designs 100

for backbone in backbones/*.pdb; do
  modal run modal_ligandmpnn.py \
    --input-pdb "$backbone" \
    --params-str "--number_of_batches 8 --temperature 0.1"
done

# Prepare FASTA with binder + target
# binder:target format for multimer
 
modal run modal_alphafold.py \
  --input-fasta all_sequences.fasta \
  --out-dir predictions/

import pandas as pd
# Load metrics
designs = pd.read_csv('all_metrics.csv')
# Filter
filtered = designs[
    (designs['pLDDT'] > 0.85) &
    (designs['ipTM'] > 0.50) &
    (designs['PAE_interface'] < 10) &
    (designs['scRMSD'] < 2.0) &
    (designs['esm2_pll'] > 0.0)
]
# Rank by composite score
filtered['score'] = (
    0.3 * filtered['pLDDT'] +
    0.3 * filtered['ipTM'] +
    0.2 * (1 - filtered['PAE_interface'] / 20) +
    0.2 * filtered['esm2_pll']
)
top_designs = filtered.nlargest(50, 'score')