ligandmpnn

LigandMPNN Ligand-Aware Design

Prerequisites

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

How to run

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

Option 1: Modal (recommended)

cd biomodals
# modal_ligandmpnn.py takes --input-pdb; LigandMPNN run.py args go in --params-str
modal run modal_ligandmpnn.py \
  --input-pdb protein_ligand.pdb \
  --params-str "--model_type ligand_mpnn --number_of_batches 16 --temperature 0.1"

GPU: A10G default | Timeout: 900s default

Option 2: Local installation

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

python run.py \
  --model_type ligand_mpnn \
  --pdb_path protein_ligand.pdb \
  --out_folder output/ \
  --number_of_batches 16 \
  --temperature 0.1

Key parameters (LigandMPNN run.py)

Parameter	Default	Description
--pdb_path	required	PDB with ligand
--model_type	protein_mpnn	ligand_mpnn, soluble_mpnn, etc.
--temperature	0.1	Sampling temperature
--number_of_batches	1	Batches (sequences = batch_size x batches)
--batch_size	1	Sequences per batch
--ligand_mpnn_use_side_chain_context	0	Use ligand side-chain context

Ligand Specification

In PDB File

Ligand must be present as HETATM records:

ATOM    ...protein atoms...
HETATM  1  C1  LIG A 999      x.xxx  y.yyy  z.zzz  1.00  0.00           C

Supported Ligand Types

• Small molecules (HETATM)
• Metals (Zn, Fe, Mg, Ca, etc.)
• Cofactors (NAD, FAD, ATP)
• DNA/RNA

Output format

output/
├── seqs/
│   └── protein.fa          # FASTA sequences
└── protein_pdb/
    └── protein_0001.pdb    # PDBs with designed sequence

Sample output

Successful run

$ python run.py --pdb_path enzyme_substrate.pdb --out_folder output/ --num_seq_per_target 8
Loading LigandMPNN model weights...
Processing enzyme_substrate.pdb
Found ligand: LIG (12 atoms)
Generated 8 sequences in 3.1 seconds

output/seqs/enzyme_substrate.fa:
>enzyme_substrate_0001, score=1.45, global_score=1.38
MKTAYIAKQRQISFVKSHFSRQLE...
>enzyme_substrate_0002, score=1.52, global_score=1.41
MKTAYIAKQRQISFVKSQFSRQLD...

What good output looks like:

• Score: 1.0-2.0 (lower = more confident)
• Ligand detected and incorporated in context
• Active site residues preserved or optimized

Decision tree

Should I use LigandMPNN?
│
├─ What's in your binding site?
│  ├─ Small molecule / ligand → LigandMPNN ✓
│  ├─ Metal ion (Zn, Fe, etc.) → LigandMPNN ✓
│  ├─ Cofactor (NAD, FAD, ATP) → LigandMPNN ✓
│  ├─ DNA/RNA → LigandMPNN ✓
│  └─ Nothing / protein only → Use ProteinMPNN
│
├─ What type of design?
│  ├─ Enzyme active site → LigandMPNN ✓
│  ├─ Metal binding site → LigandMPNN ✓
│  ├─ Protein-protein binder → Use ProteinMPNN
│  └─ De novo scaffold → Use ProteinMPNN
│
└─ Priority?
   ├─ Solubility/expression → Consider SolubleMPNN
   └─ Ligand context accuracy → LigandMPNN ✓

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

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

---

Verify

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

---

Troubleshooting

Ligand not recognized: Check HETATM format, verify ligand residue name Poor binding residues: Increase sampling around active site Missing contacts: Verify ligand coordinates in PDB

Error interpretation

Error	Cause	Fix
RuntimeError: CUDA out of memory	Long protein or large batch	Reduce batch_size
KeyError: 'LIG'	Ligand not found in PDB	Check HETATM records
ValueError: no ligand atoms	Empty ligand	Verify ligand has atoms in PDB

---

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