ProteinMPNN

What is ProteinMPNN?

ProteinMPNN solves the inverse folding problem: given a protein backbone structure, what amino acid sequences will fold into that shape? This reverses the structure prediction question; instead of asking what structure a sequence adopts, it asks what sequences can adopt a given structure.

Developed at the Institute for Protein Design and published in Science (2022), ProteinMPNN achieves 52.4% native sequence recovery on test backbones, compared to 32.9% for the previous state-of-the-art Rosetta design software. Beyond accuracy, it runs in ~1 second per protein versus ~4 minutes for Rosetta.

Experimental validation has been extensive. Crystal structures and cryo-EM reconstructions confirm that designed sequences fold to their intended structures. The method has successfully rescued previously failed designs and enabled new applications from nanomaterials to target-binding proteins.

How does ProteinMPNN work?

ProteinMPNN represents protein structures as graphs where residues are nodes and edges connect spatially proximate residues (the 32 to 48 nearest Cα neighbors). The neural network learns from this geometric representation without requiring evolutionary information or sequence alignments.

Encoding structure

The encoder (3 layers, 128 hidden dimensions) processes pairwise distances between backbone atoms: N, Cα, C, O, and a virtual Cβ. These interatomic distances capture inter-residue geometry more effectively than dihedral angles or coordinate frames. Message passing between nodes and edges propagates structural information throughout the graph.

Decoding sequences

Rather than generating amino acids sequentially from N- to C-terminus, ProteinMPNN uses order-agnostic autoregressive decoding. During training, the model learns to predict amino acids in random order. At inference, each position is decoded conditioned on the structural encoding and any previously decoded positions.

This flexibility enables practical design scenarios: fixing certain residues while redesigning others, enforcing identical sequences across homo-oligomer chains, or biasing toward specific amino acid compositions.

How to use ProteinMPNN online

ProteinMPNN runs on ProteinIQ's managed CPU infrastructure without local installation. The page supports sequence design and scoring of the uploaded native sequence, with official or compatible custom sequence weights. Side-chain packing is available after design with official or custom packing weights.

Pricing

Jobs start at 25 credits. The exact quote is calculated before submission from the parsed protein size and the design, scoring or packing workload.

The examples below assume 150 parsed residues, all chains included, and Homo-oligomer disabled. Packing uses four packs per design, three denoising steps and 16 side-chain samples.

Number of sequences	Design without packing	Design with packing
1	25 credits	25 credits
8	25 credits	40 credits
12	33 credits	52 credits
48	81 credits	Not available with four packs per design

Packing is limited to 48 structures in total. For scoring the same 150-residue input, one score per pass and one pass cost 169 credits with Single-position conditional probabilities, or 25 credits with Autoregressive score.

Larger structures, more sequences, more score samples, additional packing work and Homo-oligomer can raise the quote. Single-position scoring scales more steeply with protein length than autoregressive scoring. Official versus custom checkpoint selection, sampling temperature and random seed do not directly change the calculator's quote.

Inputs

Input	Format and limit	Requirement
Protein	One PDB or .ent file up to 50 MiB, or an RCSB PDB ID such as 5L33.	Required for design and scoring. Complete N, Cα, C and O coordinates are recommended. Incomplete backbones and modified HETATM residues produce warnings; the native parser determines whether the selected structure is usable.
Model checkpoint	One .pt file up to 50 MiB.	Required when Sequence checkpoint is Upload custom model, for either design or scoring. Leave this input empty when using Official model.
Packer checkpoint	One .pt file up to 50 MiB.	Required only for design with Pack side chains enabled and Side-chain checkpoint set to Upload custom packing model. Leave this input empty for official packing, disabled packing or scoring.

Custom sequence checkpoints must contain integer num_edges and dense model_state_dict tensors compatible with the native ProteinMPNN architecture: 128-dimensional sequence features, three encoder layers and three decoder layers. Custom packing checkpoints must contain dense model_state_dict tensors compatible with the native side-chain packer, including 16 context atoms and 32 neighbors.

Model tensor values, shapes and dtypes are preserved. Training and optimizer records do not participate in inference. Custom Python model classes, sparse tensors and quantized tensors are unsupported. Decoded tensor data is limited to 256 MiB per checkpoint. A failed or incompatible upload is not replaced with official weights.

Settings

Core settings

Parameter	Type	Default	Description
Operation	enum	Design sequences	Choose Design sequences (design) or Score native sequence (score) to evaluate the sequence in the uploaded PDB.
Sequence checkpoint	enum	Official model	Choose Official model (official) or Upload custom model (custom); the latter requires the sequence checkpoint input. Available in both operations.
Model checkpoint	enum	0.20 Å backbone noise	Official models trained with 0.02, 0.10, 0.20 or 0.30 Å backbone-coordinate noise (002, 010, 020, 030). Available only with Official model.
Number of sequences	integer	1	Design only. Generate 1 to 48 sequence variants, subject to the packed-structure limit when packing is enabled.
Sampling temperature	number	0.1	Design only. Must be positive; no upper bound is imposed. Lower values are more conservative and higher values increase diversity.
Random seed	integer	0	Accepted range: 0 to 99999. Zero chooses a random seed, recorded in the native FASTA or score artifact. A positive seed supports reproducible runs with identical inputs, settings and runtime. Previously admitted jobs retain their recorded seed.

The built-in Quick test (1 sequence) example loads RCSB structure 5L33, a designed protein from the ProteinMPNN examples, selects the official 0.20 Å model, generates one sequence at temperature 0.1 and uses fixed seed 111.

Temperature interpretation

Temperature	Behavior
0.05 to 0.1	Conservative designs with high predicted fitness; useful for maximizing sequence recovery.
0.2 to 0.3	Moderate diversity while maintaining recovery; useful for variant libraries.
0.4 to 1.0	High diversity at the cost of recovery; useful when exploring novel sequences.
Above 1.0	Increasingly random sampling; resulting sequences require careful validation.

Scoring settings

These settings apply only to Score native sequence; no new sequences or packed structures are generated.

Parameter	Type	Default	Description
Scoring method	enum	Single-position conditional probabilities	single_aa evaluates each position conditioned on all other sequence positions; autoregressive follows sampled decoding orders.
Use native sequence context	boolean	true	Use the uploaded sequence as context. Disable for backbone-only probabilities.
Scores per pass	integer	1	Batch size for each pass, from 1 to 16.
Scoring passes	integer	1	Repeat with new decoding orders, from 1 to 16 passes. The product of scores per pass and passes determines the sample count, up to 256.

Design options and amino acid constraints

Chain and residue selections use identifiers in the uploaded PDB, including insertion codes. Amino acid bias and exclusion controls apply to design only.

Parameter	Type	Default	Description
Chains to design	string	All parsed chains	Select chains such as A,B; other parsed chains stay fixed.
Homo-oligomer	boolean	false	Tie corresponding residues across chains with matching residue identifiers.
Fixed positions	string	optional	Preserve residues such as A15,A19,A1-10,B1-20. Cannot be combined with Redesigned positions.
Redesigned positions	string	optional	Redesign only the listed residues, using the same syntax as fixed positions; other positions stay fixed.
Parse chains only	string	All chains	Restrict parsing to chains such as A,B; omitted chains are not included in the run.
Include zero-occupancy atoms	boolean	false	Include atoms with zero occupancy; otherwise they are excluded.
Symmetric residue groups	string	optional	Tie arbitrary positions, for example A12,B12 | A13,B13.
Symmetry weights	string	optional	Design only. Matching per-group weights, for example 0.5,0.5 | 0.5,0.5.
Exclude amino acids	string	optional	Design only. Exclude uppercase one-letter codes without separators, such as C or CW.
Per-position amino acid biases	string	optional	Design only. JSON such as {"A12":{"G":1.1}} sets residue-specific logit biases.
Per-position amino acid exclusions	string	optional	Design only. JSON such as {"A12":"PG"} excludes amino acids at specific residues.
Amino acid biases	boolean	false	Enable global per-amino-acid bias controls for design.
Amino acid controls, Alanine (A) through Valine (V)	number	0	Available when Amino acid biases is enabled. Each of the 20 controls sets a logit bias; positive values favor and negative values disfavor that amino acid, with no imposed numeric range.

Side-chain packing

Packing applies only to Design sequences. Each design can produce multiple full-atom PDB structures, with at most 48 packed structures per job. The design count multiplied by packs per design must not exceed 48.

Parameter	Type	Default	Description
Pack side chains	boolean	false	Run side-chain packing after sequence design and return every packed PDB.
Side-chain checkpoint	enum	Official packing model	With packing enabled, select Official packing model (official) or Upload custom packing model (custom); custom selection requires the packing checkpoint input.
Packs per design	integer	4	With packing enabled, generate 1 to 4 independently packed structures per design.
Denoising steps	integer	3	With packing enabled, use 1 to 12 side-chain denoising steps.
Side-chain samples	integer	16	With packing enabled, use 1 to 64 samples per denoising step.
Repack all residues	boolean	false	With packing enabled, repack fixed and redesigned residues; otherwise only redesigned positions are packed.
Write packed atoms as HETATM	boolean	false	With packing enabled, write packed atoms as HETATM records.
Use non-protein atom context	boolean	true	With packing enabled, use parsed non-protein atoms as packing context when present in the PDB.

The sequence, scoring and packing caps are ProteinIQ job limits. They do not describe the native model's scientific input domain. Official sequence and packing weights remain the defaults.

Results and downloadable files

Results shows designs or per-position score summaries. Sequences shows available FASTA files, Data contains the result table, and Files lists the downloadable artifacts.

Operation	File or result	Contents
Design	Native .fa	Complete native FASTA, including the input sequence, sampled designs and native headers.
Design	*_designed_sequences.fasta	Designed sequences with confidence, recovery, mutation and identity annotations.
Design	Backbone .pdb files	Every designed backbone structure.
Design with packing	Packed .pdb files	Every packed structure, with names such as input_packed_1_1.pdb.
Design	Native statistics files	Complete per-sequence statistics, including native .pt tensors.
Scoring	*_score_probabilities.csv	Every amino acid mean probability and standard deviation at each parsed position.
Scoring	Native .pt	Complete native scoring tensors, including all score samples and the recorded seed.
Both	.log and result metadata	Native command, stdout/stderr, scientific source revision, settings and checkpoint provenance. Custom uploads include the original filename/SHA-256, reconstructed checkpoint and tensor digests, and isolated-reader identity.

Checkpoints are inputs; converted model files are not separate result downloads. Existing saved jobs keep their recorded settings and inputs.

ProteinIQ caps projected native artifacts at 100 MiB per job. For large structures, reduce sequence count, packs per design or scoring samples so that all native results can be returned.

Understanding results

Each designed sequence includes:

Column	Description
Sequence ID	Identifier such as seq_1 or seq_2.
Sequence	Designed amino acid sequence.
Overall confidence	exp(−mean negative log probability) over redesigned positions. Higher values mean the model assigned greater probability to the sampled sequence; this is not a fold-success probability.
Seq recovery	Fraction of redesigned positions matching the original sequence.
Mutation count	Number of positions that differ from the input.
Identity %	Percent identity to the input sequence.

Confidence remains a numeric native score; ProteinIQ does not classify it as high, medium or low. Score means and standard deviations summarize all samples across scores per pass multiplied by scoring passes.

Applications

Inverse folding enables several protein engineering workflows:

• De novo protein design: After generating a novel backbone with tools like RFdiffusion, ProteinMPNN provides sequences likely to fold into that structure
• Sequence optimization: Generate variants of existing proteins with potentially improved expression, solubility, or stability
• Functional homolog design: Create sequence-diverse proteins that maintain a target fold, useful when avoiding immune recognition or intellectual property constraints
• Rescue failed designs: Re-sequence backbones from computationally designed proteins that failed to express or fold

Limitations

ProteinMPNN designs sequences based solely on backbone geometry. It does not consider:

• Ligand interactions: For proteins with bound small molecules, metals, or nucleotides, use LigandMPNN instead
• Membrane environment: Standard ProteinMPNN was trained on soluble proteins. For transmembrane proteins or optimizing soluble expression, consider SolubleMPNN
• Stability optimization: While designs often fold well, ProteinMPNN does not explicitly optimize thermostability. Consider ThermoMPNN for stability predictions

Experimental validation remains essential; computational metrics predict but do not guarantee foldability or function.

Sources

• LigandMPNN source and command reference at the deployed revision
• ProteinMPNN paper