Proteinmpnn

ProteinMPNN Sequence Design

Prerequisites

How to run

First time? See Getting started to set up Modal and biomodals.

Option 1: Local installation (recommended)

git clone https://github.com/dauparas/ProteinMPNN.git
cd ProteinMPNN

python protein_mpnn_run.py \
  --pdb_path backbone.pdb \
  --out_folder output/ \
  --num_seq_per_target 16 \
  --sampling_temp "0.1"

GPU: T4 (16GB) sufficient | Time: ~50-100 sequences/minute

Option 2: Modal (via LigandMPNN wrapper)

cd biomodals
# modal_ligandmpnn.py takes --input-pdb and forwards run.py args via --params-str
modal run modal_ligandmpnn.py \
  --input-pdb backbone.pdb \
  --params-str "--model_type protein_mpnn --number_of_batches 16 --temperature 0.1"

GPU (Modal): A10G default | Timeout: 900s default

Note: LigandMPNN includes ProteinMPNN functionality (select with --model_type protein_mpnn).

Config Schema

Core Parameters

Temperature Guide

0.1  -> Low diversity, high recovery (production)
0.2  -> Moderate diversity (default)
0.3  -> Higher diversity (exploration)
0.5+ -> Very diverse, lower quality

IMPORTANT: Temperature must be passed as a string, not float.

Common mistakes

Temperature Parameter

✅ Correct:

--sampling_temp "0.1"    # String with quotes

❌ Wrong:

--sampling_temp 0.1      # Float without quotes - may cause errors
--sampling_temp 0.1,0.2  # Multiple temps need proper format

Fixed Positions JSONL

✅ Correct:

{"A": [1, 2, 3, 10, 11], "B": [5, 6]}

❌ Wrong:

{"A": "1,2,3,10,11"}     # String instead of list
{A: [1, 2, 3]}           # Missing quotes on key
{"A": [1,2,3,]}          # Trailing comma

Chain Selection

✅ Correct:

--pdb_path_chains A,B    # No spaces

❌ Wrong:

--pdb_path_chains A, B   # Space after comma
--pdb_path_chains "A,B"  # Quotes may cause issues

Amino Acid Biases

# Bias toward certain AAs (positive = favor)
--bias_AA_jsonl '{"A": {"A": 1.5, "W": -2.0}}'

# Omit specific AAs globally
--omit_AAs "CM"  # No cysteine or methionine

# Per-position omission
--omit_AA_jsonl '{"A": {"1": "C", "2": "CM"}}'

Multi-Chain Design

# Design chains A and B together
--pdb_path_chains A,B

# Tie chains (same sequence)
--tied_positions_jsonl tied.jsonl

Variants Comparison

Output format

output/
├── seqs/
│   └── backbone.fa          # FASTA sequences
└── backbone_pdb/
    └── backbone_0001.pdb    # PDBs with designed sequence

FASTA Header Format

>backbone_0001, score=1.234, global_score=1.234, seq_recovery=0.85
MKTAYIAKQRQISFVKSHFSRQLE...

Common workflows

Binder Sequence Design

python protein_mpnn_run.py \
  --pdb_path binder_backbone.pdb \
  --out_folder output/ \
  --num_seq_per_target 16 \
  --sampling_temp "0.1" \
  --pdb_path_chains B  # Design binder chain only

Interface Redesign

# Fix core, design interface
python protein_mpnn_run.py \
  --pdb_path complex.pdb \
  --fixed_positions_jsonl core_positions.jsonl \
  --num_seq_per_target 32

Multi-State Design

# Design for multiple conformations
python protein_mpnn_run.py \
  --pdb_path_multi state1.pdb,state2.pdb \
  --num_seq_per_target 16

Sample output

Successful run

$ python protein_mpnn_run.py --pdb_path backbone.pdb --out_folder output/ --num_seq_per_target 8
Loading model weights...
Designing sequences for backbone.pdb
Generated 8 sequences in 2.3 seconds

output/seqs/backbone.fa:
>backbone_0001, score=1.234, global_score=1.189, seq_recovery=0.82
MKTAYIAKQRQISFVKSHFSRQLEERGLTKE...
>backbone_0002, score=1.198, global_score=1.156, seq_recovery=0.79
MKTAYIAKQRQISFVKSQFSRQLDERGLTKE...

What good output looks like:

• Score: 1.0-2.0 (lower = more confident)
• Seq recovery: 0.3-0.6 for de novo, 0.7-0.9 for redesign
• Diverse sequences (not all identical) when temp > 0.1

Decision tree

Should I use ProteinMPNN?
│
├─ Have a backbone structure?
│  ├─ Yes → Continue below
│  └─ No → Use RFdiffusion first
│
├─ What's in the binding site?
│  ├─ Nothing / protein only → ProteinMPNN ✓
│  ├─ Small molecule / ligand → Use LigandMPNN
│  └─ Metal / cofactor → Use LigandMPNN
│
├─ Priority?
│  ├─ Solubility/expression → Consider SolubleMPNN
│  ├─ Speed → ProteinMPNN ✓
│  └─ AF2 optimization → Consider ColabDesign
│
└─ Need fixed positions?
   ├─ Yes → Use --fixed_positions_jsonl
   └─ No → ProteinMPNN ✓ (design all)

Typical performance

Throughput: ~50-100 sequences/minute on T4 GPU.

Verify

grep -c "^>" output/seqs/*.fa  # Should match backbone_count × num_seq_per_target

Troubleshooting

Low sequence diversity: Increase sampling_temp to 0.2-0.3 Poor recovery: Decrease sampling_temp to 0.1 OOM errors: Reduce batch_size Unwanted cysteines: Use --omit_AAs "C"

Error interpretation

Next: Structure prediction for validation → protein-qc for filtering.