Protein Design Workflow

Protein Design Workflow Guide

Standard binder design pipeline

Overview

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

Phase 1: Target preparation

1.1 Obtain target structure

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

1.2 Clean and prepare

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

1.3 Select hotspots

• Choose 3-6 exposed residues
• Prefer charged/aromatic (K, R, E, D, W, Y, F)
• Check surface accessibility
• Verify residue numbering

Output: target_prepared.pdb, hotspot list

Phase 2: Backbone generation

Option A: RFdiffusion (diverse exploration)

# 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

Option B: BindCraft (end-to-end)

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

Output: 100-500 backbone PDBs

Phase 3: Sequence design

For RFdiffusion backbones

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

Output: 8 sequences per backbone (800-4000 total)

Phase 4: Structure validation

Predict complexes

# Prepare FASTA with binder + target
# binder:target format for multimer

modal run modal_alphafold.py \
  --input-fasta all_sequences.fasta \
  --out-dir predictions/

Output: AF2 predictions with pLDDT, ipTM, PAE

Phase 5: Filtering and selection

Apply standard thresholds

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']  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')

Output: 50-200 filtered candidates

Resource planning

Compute requirements

| Stage | GPU | Time (100 designs) | |-------|-----|-------------------| | RFdiffusion | A10G | 30 min | | ProteinMPNN | T4 | 15 min | | Chai / AlphaFold | A100 | 4-8 hours | | Filtering | CPU | 15 min |

Total timeline

• Small campaign (100 designs): 8-12 hours
• Medium campaign (500 designs): 24-48 hours
• Large campaign (1000+ designs): 2-5 days

Quality checkpoints

After backbone generation

• [ ] Visual inspection of diverse backbones
• [ ] Secondary structure present
• [ ] No clashes with target

After sequence design

• [ ] ESM2 PLL > 0.0 for most sequences
• [ ] No unwanted cysteines (unless intentional)
• [ ] Reasonable sequence diversity

After validation

• [ ] pLDDT > 0.85
• [ ] ipTM > 0.50
• [ ] PAE_interface < 10
• [ ] Self-consistency RMSD < 2.0 A

Final selection

• [ ] Diverse sequences (cluster if needed)
• [ ] Manufacturable (no problematic motifs)
• [ ] Reasonable molecular weight

Common issues

| Problem | Solution | |---------|----------| | Low ipTM | Check hotspots, increase designs | | Poor diversity | Higher temperature, more backbones | | High scRMSD | Backbone may be unusual | | Low pLDDT | Check design quality |

Advanced workflows

Multi-tool combination

1. RFdiffusion for initial backbones
2. Mosaic for gradient-based refinement
3. ProteinMPNN diversification
4. AF2 final validation

Iterative refinement

1. Run initial campaign
2. Analyze failures
3. Adjust hotspots/parameters
4. Repeat with insights

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: adaptyvbio
• Source: adaptyvbio/protein-design-skills
• License: MIT

Install and usage instructions live in the source repository linked above.