Protein Design · Sequence prediction

LigandMPNN

Protein sequence prediction, to conform with backbone coordinates. Takes external molecules into account; to some degree a superset of ProteinMPNN, but is a different model. A useful step after RFDiffusion in a protein design pipeline, and before validation with structure prediction.

A deep learning-based protein sequence design method that explicitly models all non-protein components of biomolecular systems. LigandMPNN generates not only sequences but also sidechain conformations to allow detailed evaluation of binding interactions. Experimental characterization demonstrates that LigandMPNN can generate small molecule and DNA-binding proteins with high affinity and specificity. It allows explicit modeling of small molecule, nucleotide, metal, and other atomic contexts.

Documentation ↗ Input parameters ↗ Examples ↗ Paper ↗ Source code ↗ License ↗

No jobs before this; it will run immediately.

Not logged in. Register an account to keep track of your jobs and results, and add presets and bookmarks. Email results to this address when complete:

Ready

Example preset

Choose an example to fill in its settings and input structures.

Inputs

Official documentation ↗

Job name A label for this run and its results.

Structure (PDB) Upload a protein or protein–ligand PDB, or choose a bundled official example. Required unless multiple structures are supplied. In this form: PDB contents are mapped to --pdb_path. Preserve HETATM records for ligand, nucleotide and metal context. This upstream parser reads PDB, not mmCIF.

Multiple structures (JSON) Object mapping unique names to PDB contents or bundled references. Leave the single-structure field blank. In this form: Example: {"first": "bio-tools://ligandmpnn/inputs/1BC8.pdb", "second": "bio-tools://ligandmpnn/inputs/4GYT.pdb"}. The service creates a native --pdb_path_multi file with job-local paths; server paths are not accepted.

Chains to design or score Comma-separated chain IDs, e.g. A,B. Other chains remain fixed context. Blank selects all parsed protein chains.

Parse only these chains Comma-separated chain IDs. Excludes every other chain, including its ligand atoms, from parsing and output; blank keeps all chains.

Include atoms with zero occupancy Upstream normally discards atoms whose occupancy is zero.

Model

Official documentation ↗

Model The upstream default is ProteinMPNN. Select LigandMPNN to condition sequence design or scoring on non-protein atoms.

ProteinMPNN checkpoint Official checkpoint filename in the configured model directory. Noise is the training backbone noise in angstroms.

LigandMPNN checkpoint Official checkpoint filename in the configured model directory. Noise is the training backbone noise in angstroms.

SolubleMPNN checkpoint Official checkpoint filename in the configured model directory. Noise is the training backbone noise in angstroms.

Membrane: per-residue labels checkpoint Official checkpoint filename in the configured model directory. Noise is the training backbone noise in angstroms.

Membrane: global label checkpoint Official checkpoint filename in the configured model directory. Noise is the training backbone noise in angstroms.

Sampling and scoring

Official documentation ↗

Batch size Samples evaluated together; reduce to 1 for a small CPU run.

Number of batches Total samples = batch size × number of batches. For scoring, upstream recommends at least 10 batches to average decoding-order effects.

Sampling temperature Positive temperature; higher values give more sequence diversity.

Random seed Zero chooses a random seed. A nonzero seed is reproducible within the same runtime.

Condition scores on sequence Use the sequence in the PDB as well as the backbone. Disable for backbone-only probabilities. Scoring writes a .pt dictionary of per-residue probabilities and statistics.

Residue constraints

Official documentation ↗

Fixed residues Space-separated PDB residue IDs, e.g. C1 C2 C3. These residues retain their input amino acids. In this form: Uses chain + author residue number + optional insertion code, e.g. B82A. Unlike original ProteinMPNN, numbering gaps are not filled with X.

Fixed residues per structure (JSON) Object keyed by the names in pdb_path_multi, with a space-separated residue list per structure. Overrides the common list for that structure.

Redesign only these residues Space-separated PDB residue IDs; every other residue is fixed. Choose this or fixed_residues. In this form: Uses chain + author residue number + optional insertion code, e.g. B82A. Unlike original ProteinMPNN, numbering gaps are not filled with X.

Redesign only these residues per structure (JSON) Object keyed by the names in pdb_path_multi, with a space-separated residue list per structure. Overrides the common list for that structure.

Tied residue groups Comma-separated residues within each group, | between groups: C1,C2,C3|C4,C5|C6,C7. Groups cannot overlap.

Tied residue weights Matching weights for every group: 0.33,0.33,0.33|0.5,0.5|0.5,0.5. Required with explicit design symmetry; finite negative weights are supported.

Tie all parsed protein chains Upstream ties matching PDB residue numbers and insertion codes across ALL parsed chains with equal weights. Chains must have identical residue IDs. Use parse-only chains to restrict the assembly.

Sequence preferences

Official documentation ↗

Global amino-acid bias Comma-separated AA:value pairs, e.g. W:3.0,P:3.0,C:3.0,A:-3.0. Positive values favor an amino acid.

Globally omitted amino acids One-letter codes to exclude, e.g. CDFGHILMNPQRSTVWY leaves A, E and K. LigandMPNN samples the 20 standard amino acids.

Per-residue amino-acid bias JSON keyed by PDB residue IDs, e.g. {"C1": {"G": -0.3, "P": 10.8}}.

Per-residue amino-acid bias per structure JSON keyed by pdb_path_multi names; each value is a per-residue object. Overrides common guidance for that structure.

Per-residue amino-acid exclusions JSON keyed by PDB residue IDs, e.g. {"C1": "ACDEFGHIKLMNPQRSTVW"} leaves Y at C1.

Per-residue amino-acid exclusions per structure JSON keyed by pdb_path_multi names; each value is a per-residue object. Overrides common guidance for that structure.

Atomic and membrane context

Official documentation ↗

Use ligand atom context LigandMPNN only: condition on nearby non-protein atoms, including small molecules, nucleotides and metals.

Use fixed side chains as context LigandMPNN only: use side-chain atoms of fixed residues as additional context.

Ligand score cutoff (Å) Selects residues near context atoms for the reported ligand confidence. It does not set the model’s atom-context neighborhood.

Buried membrane residues Per-residue membrane model only: space-separated PDB residue IDs, e.g. C1 C2 C3 C11. Class 2 (hydrophobic).

Membrane interface residues Per-residue membrane model only: e.g. C4 C5 C6 C22. Class 1 (polar); unlisted residues are class 0. Lists must not overlap.

Global membrane label Global membrane model only: 0 = soluble, 1 = transmembrane.

Side-chain packing

Official documentation ↗

Pack designed side chains Run the official side-chain packing model after sequence design.

Packs per sequence Independent packed structures per generated sequence. Use 1 for one pack; upstream run.py produces no packed files at 0 despite the README’s fast-example text.

Packing denoising steps Number of recycling/denoising steps per pack.

Packing mixture samples Samples drawn from each mixture distribution; the highest-likelihood sample is used.

Pack with ligand context Consider ligand, DNA and other context atoms during packing.

Repack fixed residues too Off preserves side chains at fixed positions and uses them as packing context.

Write packed context atoms as HETATM Force context atoms in packed output PDBs to use HETATM records.

Packed PDB suffix Suffix appended to packed PDB filenames; letters, digits, underscores, dots and hyphens only.

Output

Official documentation ↗

Save design statistics Write .pt statistics including sequences, probabilities, decoding order, masks, seed and temperature.

FASTA chain separator Separator between chains in output FASTA. PDB output preserves chain IDs and residue numbering.

Output filename ending Optional ending such as _xyz; letters, digits, underscores, dots and hyphens only.

Number output designs from zero Start designed PDB file numbering at 0 instead of 1.

Print progress Include upstream progress messages in the run log.

These local CLI options are not applicable to Bio Web and are omitted from its inputs:

out_folder: The service creates a separate output directory for each job and archives the results.
checkpoint_path_sc: Side-chain packing uses the installed official ligandmpnn_sc_v_32_002_16.pt checkpoint.

Ready