LigandMPNN
v1.0.1Design protein sequences around ligands, metals, and nucleotides for enzyme engineering and binding-site optimization.
LigandMPNN online on ProteinIQ designs protein sequences for a fixed backbone while taking the ligands, metal ions, nucleotides and cofactors in the same PDB file into account. Upload a protein-ligand structure, choose how many sequences to sample, and the job returns designed sequences with confidence and recovery statistics, FASTA and PDB files, and optional full-atom side-chain packing. A second mode scores the sequence already in the PDB instead of designing a new one.
LigandMPNN is a protein inverse-folding model from the Baker lab, published by Dauparas, Lee and colleagues in Nature Methods in 2025. It extends ProteinMPNN with atomic context from non-protein molecules. It changes amino acids only: the backbone stays as submitted, ligand coordinates must already be in place, and nothing is docked or generated.
A LigandMPNN job costs at least 50 credits. The quote grows with the number of parsed residues and the amount of work requested, and the exact figure appears before submission.
Design jobs without side-chain packing:
| Sequences | 100 residues | 150 residues | 300 residues | 500 residues | 1,000 residues |
|---|---|---|---|---|---|
| 1 | 50 | 50 | 50 | 50 | 50 |
| 8 | 50 | 50 | 79 | 110 | 172 |
| 16 | 61 | 79 | 124 | 172 | 270 |
| 48 | 124 | 161 | 252 | 351 | 550 |
Design jobs with Pack side chains on and the default packing settings (4 packs per design, 3 denoising steps, 16 samples):
| Sequences | 100 residues | 150 residues | 300 residues | 500 residues | 1,000 residues |
|---|---|---|---|---|---|
| 1 | 50 | 50 | 50 | 50 | 70 |
| 8 | 61 | 79 | 124 | 172 | 270 |
| 16 | 95 | 124 | 194 | 270 | 423 |
Each cell comes from the current calculator for one protein chain with the listed number of residues. Residues are counted over the chains that are parsed, so Parse chains only changes the quote. Homo-oligomer adds 10% to the work before scaling.
Scoring jobs are priced differently. A single pass with Autoregressive score costs 50 credits at every length in the table. A single pass with Single-position conditional probabilities costs 199, 337, 828, 1,608 and 3,959 credits for 100, 150, 300, 500 and 1,000 residues, because that method repeats the model once per residue. More scoring passes or a larger Scores per pass raise either figure.
The seed, temperature, checkpoint, atom context and position controls do not change the quote.
| Input | Accepted formats | Notes |
|---|---|---|
Protein | .pdb, .ent, or a PDB ID such as 1EVE | One structure up to 50 MB. Backbone atoms N, CA, C and O must be complete. Ligands must be present in the same file |
Protein and ligand files cannot be submitted separately. A ligand prepared elsewhere has to be placed in the intended binding geometry and merged into the protein PDB first. A pose from AutoDock Vina or GNINA can be merged this way, and a complex predicted by Boltz-2 with PDB output can be used directly.
| Parameter | Type | Default | Description |
|---|---|---|---|
Operation | enum | Design sequences | Design sequences samples new amino acids. Score native sequence returns per-position probabilities for the sequence in the PDB |
Model checkpoint | enum | 010 | Backbone noise used in training: 005 (0.05 Å), 010 (0.10 Å), 020 (0.20 Å) or 030 (0.30 Å) |
Number of sequences | integer | 1 | Independent designs to sample, from 1 to 48 |
Sampling temperature | number | 0.1 | Lower values stay close to the most probable residue, higher values add diversity |
Use atom context | boolean | on | Includes parsed ligand, metal and nucleotide atoms. Turning it off masks them but still runs the LigandMPNN checkpoint |
Use side-chain context | boolean | off | Includes atoms from fixed protein side chains |
Random seed | integer | 111 | From 0 to 99999 |
These apply only when Operation is Score native sequence.
| Parameter | Type | Default | Description |
|---|---|---|---|
Scoring method | enum | Single-position conditional probabilities | Scores each site given the others, or follows sampled decoding orders (Autoregressive score) |
Use native sequence context | boolean | on | Includes the submitted sequence while scoring |
Scores per pass | integer | 1 | Batch size for each pass, from 1 to 16 |
Scoring passes | integer | 1 | Number of passes with new decoding orders, from 1 to 16 |
Scoring cutoff (Å) | number | 8 | Residues within this distance of a non-protein atom are scored with ligand context |
Include zero-occupancy atoms | boolean | off | Keeps atoms whose occupancy is 0 in the crystal structure |
Scoring returns the mean and standard deviation for all 20 amino acids at every parsed residue. It does not generate sequences.
| Parameter | Type | Default | Description |
|---|---|---|---|
Chains to design | string | optional | Comma-separated chain IDs to redesign. Other parsed chains stay fixed |
Fixed positions | string | optional | PDB residue IDs, ranges or whole chains to keep, such as A15,A20-25,B |
Redesigned positions | string | optional | Residues to redesign while everything else is fixed. Cannot be combined with Fixed positions |
Parse chains only | string | optional | Comma-separated chains to read from the structure |
Homo-oligomer | boolean | off | Ties equivalent positions across equal-length chains |
Symmetric residue groups | string | optional | Ties arbitrary positions, for example A12,B12 | A13,B13 |
Symmetry weights | string | optional | Weights for each symmetric group |
Exclude amino acids | string | optional | One-letter codes to omit everywhere, such as CW |
Amino acid biases | boolean | off | Shows 20 numeric fields, one per amino acid (default 0, step 0.1), that shift sampling globally |
Per-position amino acid biases | JSON | optional | Residue-specific biases, such as {"A12":{"G":1.1}} |
Per-position amino acid exclusions | JSON | optional | Residue-specific exclusions, such as {"A12":"PG"} |
Residue IDs and insertion codes follow the submitted PDB.
These apply to design jobs only. Packing runs after sequence design and can return several packed structures for every design.
| Parameter | Type | Default | Description |
|---|---|---|---|
Pack side chains | boolean | off | Runs the native side-chain packing model |
Packs per design | integer | 4 | Packed structures per sequence, from 1 to 4 |
Denoising steps | integer | 3 | From 1 to 12 |
Side-chain samples | integer | 16 | Samples per denoising step, from 1 to 64 |
Repack all residues | boolean | off | Packs fixed residues as well as redesigned ones |
Write packed atoms as HETATM | boolean | off | Uses the model's HETATM output option for packed atoms |
Use ligand context while packing | boolean | on | Includes parsed ligand and other non-protein atoms |
Projected native artifacts are capped at 100 MiB per job. For large structures, reduce the number of sequences or packs per design, or the scoring passes and scores per pass.
| Output | Description |
|---|---|
Results | Sequence viewer for the designs, with positions that differ from the input highlighted |
Sequences | FASTA text of the native output and the designed sequences |
Data | Table of per-design statistics, or per-position probabilities for scoring jobs |
Files | Native FASTA, one backbone PDB per design, optional packed PDB files and run provenance |
Scoring jobs return native PyTorch tensors and a CSV with all per-position probabilities and deviations.
| Column | Meaning |
|---|---|
Sequence ID | Identifier of the sampled design |
Confidence | Geometric mean probability the model assigned to its choices at redesigned positions, from 0 to 1 |
Ligand confidence | The same measure over redesigned positions close to the ligand |
Seq. recovery | Fraction of redesigned positions that match the submitted sequence |
Mutations | Number of positions that differ from the submitted sequence |
Identity (%) | Sequence identity to the submitted sequence across the parsed residues |
Confidence is the model's own likelihood. It is not a binding affinity, a stability estimate or a measure of activity. Higher values usually mean the pocket constrains the choice strongly, and designs from a higher temperature will have lower confidence on average.
Sequence recovery says how close a design is to the starting sequence. It is not a success score. For a native enzyme or binder, high recovery near the ligand is a sanity check, because the original residues are what evolution selected. For a de novo backbone, there is no native sequence worth recovering.
On native backbones, LigandMPNN recovered 63.3% of residues that contact small molecules, against 50.4% for Rosetta and 50.5% for ProteinMPNN. For nucleotides the figures were 50.5%, 35.2% and 34.0%, and for metals 77.5%, 36.0% and 40.6%. These are recovery rates on natural proteins, not experimental success rates.
Confidence as predicted affinity. It only reflects how sharply the model prefers its choices.Confidence between jobs run at different temperatures or checkpoints.Use atom context off equals ProteinMPNN. The LigandMPNN weights still run.Predict the structure of the best sequences with AlphaFold 2, ESMFold or Boltz-2 and check that the backbone and pocket are reproduced. Designs for enzymes and binders still need experimental validation.
ProteinMPNN sees only the protein backbone. LigandMPNN also sees the coordinates of nearby ligand, metal and nucleotide atoms, so it can choose residues that pack, hydrogen-bond or coordinate against them. Use LigandMPNN when the design depends on a positioned small molecule, cofactor, metal or nucleic acid. Use ProteinMPNN for backbone-only design, SolubleMPNN when solubility is the priority, and HyperMPNN for thermostable designs.
LigandMPNN jobs start at 50 credits, against 25 for ProteinMPNN and SolubleMPNN,.
For a new backbone, generate one first with RFdiffusion, then assign sequences with LigandMPNN.
LigandMPNN is distributed as source code at dauparas/LigandMPNN and there is no official web server. Running it locally means installing the Python environment, downloading the model weights and calling run.py with a PDB file and options.
ProteinIQ runs the same code at a pinned commit with checksum-verified weights and records the commit and checkpoint in the results. The four official LigandMPNN checkpoints and the side-chain packing model are available. Membrane-specific checkpoints are not offered. Jobs are limited to 48 sequences and a 50 MB structure, and jobs are saved to the workspace with their settings and files. Paid plans can submit jobs through the ProteinIQ API.
Design protein sequences optimized for thermal stability from backbone structures.
Design amino acid sequences for protein backbones with fixed positions, amino acid biases, and sequence diversity controls.
Design sequences with the ProteinMPNN-family model trained on structures from soluble-protein PDB IDs.
Design sequences for a structure or an aligned ensemble
Design antibody CDR loops from antibody-antigen complex structures using inverse folding.
Joint protein and coding DNA design from protein backbones
Design antibody sequences from structure with AI-powered inverse folding
Design protein sequences from 3D backbone structures with controllable sampling diversity.
Family-conditioned protein sequence generation from FASTA, A2M, or A3M input
Design binders, enzymes and ligand pockets with all-atom flow matching