Option 1: Modal (recommended)

SolubleMPNN uses the ProteinMPNN Modal wrapper with soluble model:

cd biomodals
modal run modal_proteinmpnn.py \
  --pdb-path backbone.pdb \
  --num-seq-per-target 16 \
  --sampling-temp 0.1 \
  --model-name v_48_020

GPU: T4 (16GB) | Timeout: 600s default

Option 2: Local installation

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

# Use soluble model weights
python protein_mpnn_run.py \
  --pdb_path backbone.pdb \
  --out_folder output/ \
  --num_seq_per_target 16 \
  --sampling_temp "0.1" \
  --model_name "v_48_020"  # Soluble model

Key parameters

Parameter	Default	Range	Description
`--pdb_path`	required	path	Input structure
`--num_seq_per_target`	1	1-1000	Sequences per structure
`--sampling_temp`	"0.1"	"0.0001-1.0"	Temperature (string!)
`--model_name`	v_48_020	string	Soluble model variant

Model Variants

Model	Description	Use Case
v_48_002	Standard	General design
v_48_020	Soluble-trained	E. coli expression
v_48_030	High solubility	Difficult targets

Output format

output/
├── seqs/backbone.fa
└── backbone_pdb/backbone_0001.pdb

Sample output

Successful run

$ python protein_mpnn_run.py --pdb_path backbone.pdb --model_name v_48_020 --num_seq_per_target 8
Loading soluble model weights (v_48_020)...
Designing sequences for backbone.pdb
Generated 8 sequences in 2.1 seconds

output/seqs/backbone.fa:
>backbone_0001, score=1.31, global_score=1.24, seq_recovery=0.78
MKTAYIAKQRQISFVKSHFSRQLE...
>backbone_0002, score=1.28, global_score=1.21, seq_recovery=0.81
MKTAYIAKQRQISFVKSQFSRQLD...

What good output looks like:

Score: 1.0-2.0 (lower = more confident)
Reduced hydrophobic patches compared to standard MPNN
Improved charge distribution

Decision tree

Should I use SolubleMPNN?
│
├─ What expression system?
│  ├─ E. coli → SolubleMPNN ✓
│  ├─ Mammalian → ProteinMPNN (PTMs matter more)
│  └─ Yeast → Either
│
├─ History of expression problems?
│  ├─ Yes, aggregation → SolubleMPNN ✓
│  ├─ Yes, low yield → SolubleMPNN ✓
│  └─ No → ProteinMPNN is fine
│
├─ What's in the binding site?
│  ├─ Small molecule / ligand → Use LigandMPNN
│  └─ Nothing / protein only → SolubleMPNN ✓
│
└─ Need highest solubility?
   ├─ Yes → Use v_48_030 model
   └─ Standard → Use v_48_020 model

Typical performance

Campaign Size	Time (T4)	Cost (Modal)	Notes
100 backbones × 8 seq	15-20 min	~$2	Standard
500 backbones × 8 seq	1-1.5h	~$8	Large campaign

Expected improvement: +15-30% solubility score vs standard ProteinMPNN.

Verify

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

Troubleshooting

Still insoluble: Try v_48_030 (higher solubility bias) Low diversity: Increase temperature to 0.2 Poor folding: Use standard ProteinMPNN and optimize later

Error interpretation

Error	Cause	Fix
`RuntimeError: CUDA out of memory`	Long protein or large batch	Reduce batch_size
`FileNotFoundError: v_48_020`	Missing model weights	Download soluble weights

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

Requirement	Minimum	Recommended
Python	3.8+	3.10
CUDA	11.0+	11.7+
GPU VRAM	8GB	16GB (T4)
RAM	8GB	16GB

solublempnn

SolubleMPNN Solubility-Optimized Design

Prerequisites

How to run

More from adaptyvbio/protein-design-skills

cell-free-expression

binder-design

pdb

proteinmpnn

protein-qc

alphafold