SolubleMPNN

What is SolubleMPNN?

SolubleMPNN is a ProteinMPNN-family inverse-folding model distributed with LigandMPNN. Its checkpoint was trained only on structures associated with soluble-protein PDB IDs. Given a protein backbone, it samples compatible amino acid sequences without changing the backbone coordinates.

The training-set choice distinguishes SolubleMPNN from the general ProteinMPNN checkpoint. It is not a solubility predictor, and its output does not guarantee soluble expression, reduced aggregation, or retained function. Those properties depend on sequence, structure, construct boundaries, expression conditions, and experimental context.

How SolubleMPNN works

SolubleMPNN uses the ProteinMPNN architecture and sequence-design procedure with its own trained weights. Backbone geometry is encoded first, then amino acids are sampled autoregressively at redesigned positions. Fixed residues, chain selection, symmetry constraints, global biases, and per-position constraints alter which sequences the model may sample.

Four official checkpoints are available:

Checkpoint	Training noise	Notes
002	0.02 Å	Lowest-noise SolubleMPNN checkpoint
010	0.10 Å	Lower coordinate-noise augmentation
020	0.20 Å	Default SolubleMPNN checkpoint
030	0.30 Å	Highest-noise SolubleMPNN checkpoint

How to use SolubleMPNN online

Submit a PDB or ENT structure, or fetch a PDB entry from RCSB, then choose sequence design or native-sequence scoring. ProteinIQ runs the selected SolubleMPNN checkpoint and returns designed sequences, confidence and recovery statistics, backbone PDBs, scoring probabilities, and optional full-atom structures from side-chain packing.

Input

Input	Description
Protein	One PDB or ENT file up to 50 MB, or an RCSB PDB ID. Every parsed standard residue must contain N, CA, C, and O backbone atoms.

SolubleMPNN does not use ligand, metal, nucleotide, or cofactor atoms during sequence design. LigandMPNN is intended for designs where those atoms should influence amino acid selection.

Core settings

Setting	Description
Operation	Design sequences samples new amino acids. Score native sequence evaluates the sequence encoded by the PDB.
Model checkpoint	Selects 002, 010, 020, or 030. The default is 020.
Number of sequences	Number of independent designs, from 1 to 48. The default is 1.
Sampling temperature	Positive sampling temperature, default 0.1. Lower values are more conservative; higher values increase diversity.
Random seed	Integer seed from 0 to 99999, default 111.

Temperature must be greater than zero. SolubleMPNN accepts values above 1.0 because the source model does not impose an artificial maximum.

Scoring settings

Setting	Description
Scoring method	Single-position conditional probabilities scores each site conditioned on the other sequence positions. Autoregressive score follows sampled decoding orders.
Use native sequence context	Includes the submitted sequence while scoring. Disable it to calculate probabilities from backbone context without native sequence context.
Scores per pass	Batch size for each scoring pass, from 1 to 16.
Scoring passes	Number of passes with new decoding orders, from 1 to 16.

The scoring operation returns mean probabilities and standard deviations for all 20 amino acids at every parsed position. It does not sample new sequences.

Position and composition controls

| Setting | Description | | ------------------------------------ | ------------------------------------------------------------------------------------------- | --------- | | Chains to design | Comma-separated chains to redesign. Other parsed chains remain fixed. | | Fixed positions | PDB residue IDs, ranges, or whole chains to keep unchanged, such as A15,A20-25,B. | | Redesigned positions | Residues to redesign while fixing all others. It cannot be combined with Fixed positions. | | Parse chains only | Comma-separated chains to include from the submitted structure. | | Homo-oligomer | Ties equivalent positions across equal-length chains. | | Symmetric residue groups | Ties arbitrary positions, such as A12,B12 \| A13,B13. | | Symmetry weights | Optional weights corresponding to the symmetric groups. | | Exclude amino acids | Uppercase one-letter amino acid codes to omit globally, such as CW. | | Amino acid biases | Adds unrestricted global sampling biases for individual amino acids. | | Per-position amino acid biases | JSON mapping PDB residue IDs to residue-specific biases, such as {"A12":{"G":1.1}}. | | Per-position amino acid exclusions | JSON mapping PDB residue IDs to excluded amino acids, such as {"A12":"PG"}. |

Residue-specific settings use the numbering and insertion codes from the PDB, not one-based sequence indices.

Side-chain packing

SolubleMPNN sequence design produces sequences on the submitted backbone. The optional LigandMPNN side-chain packer adds full-atom side-chain conformations after design.

ProteinIQ caps projected native artifacts at 100 MiB per job. For large structures, reduce the number of sequences or packs per design, or reduce scoring passes and scores per pass, so every native result can be returned safely.

Setting	Description
Pack side chains	Runs the native side-chain packing model after design. Off by default.
Packs per design	Packed structures returned for each sequence, from 1 to 4. The default is 4.
Denoising steps	Side-chain denoising steps, from 1 to 12. The default is 3.
Side-chain samples	Samples per denoising step, from 1 to 64. The default is 16.
Repack all residues	Packs fixed as well as redesigned residues.
Write packed atoms as HETATM	Uses the model's HETATM output option for packed atoms.
Use non-protein atom context while packing	Includes parsed non-protein atoms during packing. It does not add ligand context to SolubleMPNN sequence design.

Interpreting SolubleMPNN results

Design table

Column	Meaning
Sequence ID	Identifier assigned to the sampled design
Sequence	Designed sequence with chain boundaries preserved
Overall confidence	Geometric mean model probability over redesigned positions, from 0 to 1
Seq recovery	Fraction of redesigned positions that match the submitted sequence
Mutation count	Number of redesigned positions that differ from the submitted sequence
Identity %	Sequence identity across all parsed residues

SolubleMPNN results do not include a ligand-confidence metric. In the original command output, that header duplicates overall confidence for non-ligand models rather than measuring ligand context, so ProteinIQ omits it for SolubleMPNN.

Confidence ranks sequence compatibility under the model. It is not a calibrated solubility, stability, expression, or aggregation score. Sequence recovery is agreement with the input at redesigned positions, not percent identity across the whole structure and not a measure of experimental quality.

Files

• The native FASTA preserves the model's original headers and chain separators.
• The generated-design FASTA contains only sampled sequences.
• One backbone PDB is returned for every sampled design.
• Optional packed PDB files contain full-atom side-chain coordinates.
• Native statistics and run provenance remain downloadable.
• Scoring jobs return native PyTorch tensors plus a CSV containing all amino acid probabilities and deviations.

When to use SolubleMPNN vs alternatives

SolubleMPNN is a reasonable design model when the intended target is a soluble protein and the training provenance is relevant to the design choice. ProteinMPNN is the general backbone-only alternative and remains appropriate when no soluble-only training preference is needed. LigandMPNN should be used when an already positioned ligand, metal, nucleotide, or cofactor must affect sequence selection.

Candidate sequences should be evaluated with independent structure, stability, aggregation, and solubility methods as appropriate. For sequence-based solubility assessment, NetSolP 1.0 answers a different question from inverse folding.

SolubleMPNN does not redesign the backbone, model conformational changes, predict expression yield, or guarantee that a sequence will fold or remain soluble. Experimental validation remains necessary.