Can I run ProteinMPNN online without a local GPU?

Yes. Upload a backbone PDB and Ranomics Tools runs ProteinMPNN on a dedicated GPU in seconds. You get ranked sequence redesigns plus per-position recovery, with no install and no CUDA setup.

How much does one ProteinMPNN job cost?

Billing is by the second. A typical ProteinMPNN job costs a fraction of a cent because the model finishes in under a minute on most backbones. New accounts start with a $20 wallet balance, which is enough for thousands of runs.

Does ProteinMPNN need an MSA?

No. ProteinMPNN is a structure-conditioned sequence model. It takes the backbone coordinates directly and never queries an MSA. That makes it the fastest route from a designed backbone to a sequence you can synthesise.

ProteinMPNN

ProteinMPNN is a no-install online sequence design tool you can run through tools.ranomics.com on a dedicated GPU. The step between a designed shape and something you can actually order: give it any backbone, from any other tool, and it proposes the sequences most likely to fold into it.

Design inputs

Backbone PDB

Upload the backbone as .pdb, .cif, or .mmcif — a 3D shape whose sequence you have not decided yet, or one whose sequence you want to replace. Only the backbone atoms are read, so side chains and any sequence already in the file are ignored: the shape is the input, not the sequence.

Chains to design

Which chains get a new sequence — usually just your binder. Space- or comma-separated (A, A B, H L). Chains you leave out keep their existing sequence and are still seen by the model, so a target chain left out here shapes the design without being changed by it.

Number of sequences

How many candidate sequences to propose (1 to 1000). They are all designed for the same backbone; more candidates means more to choose from when you fold or order, at no extra runtime worth worrying about. 50 is a reasonable default.

Sampling temperature

How much the model is allowed to stray from its single best guess at each position. Low (0.1, the default) gives near-identical sequences that are the safest bet for folding back into your backbone; raising it toward 0.3 buys diversity at the cost of confidence, which is worth doing only once a low-temperature batch has failed to give you anything.

Fixed positions (optional)

Positions to hold fixed — everything else in the designed chain(s) gets redesigned. Leave blank to redesign the whole chain. Positions are 1-indexed within their chain (position 1 = first residue), not author residue numbers. Use CHAIN:list groups separated by spaces, with single positions or lo-hi ranges: A:1-44,46-66 B:5,7. The chain prefix may be omitted when only one chain is being designed. To redesign a few liability positions, list the complement — freezing all but 45, 67 and 89 of a 113-mer is A:1-44,46-66,68-88,90-113. Chains must be numbered from 1 with no gaps or insertion codes, or a position stops landing on the residue you named. That is checked against the real structure once the job starts, so a chain that fails it comes back as a failed run rather than an error on this page.

Upload your backbone PDB and pick chain(s) to redesign. ~30 to 60 s on A10G-24GB.

Campaign

Group related runs on your jobs page. Leave blank to file under Uncategorized.

Estimated cost $0.00 based on parameters and historical compute time

Billed by actual GPU consumed, up to the estimate. Surplus returns to your wallet automatically. Cancellation does not refund GPU time already used.

About ProteinMPNN

What it is

You give it a backbone — a 3D protein shape with no sequence decided yet — and it proposes an amino acid for every position, chosen so the sequence should fold back into that exact shape. It reads only the backbone atoms, so side chains in your file are ignored. Each candidate comes back with a score and, on a natural backbone, the fraction of the real sequence it recovered. ProteinMPNN, Dauparas et al., Science 2022.

When to use

You have a backbone and need sequences for it.
You want to re-sequence a binder another tool here designed, before folding or ordering it.
You want several alternative sequences threaded through a curated structure so you can choose between them.

Prerequisites

Backbone PDB or mmCIF (only Cα and backbone atoms are used).
Chain ID(s) of the region(s) to redesign. Other chains stay fixed as context.

Input glossary

Chains to design. Which chains in the PDB MPNN should redesign (e.g. A, A B, H L). Other chains are held fixed as context.
Number of sequences. How many independent samples to draw (1 to 1000). Each sample is independent; rank by score and ProteinMPNN recovery rate.
Sampling temperature. Lower means more conservative (closer to argmax); higher means more diverse. Defaults to 0.1 per the upstream README.
Fixed positions (optional). Positions to hold fixed inside a designed chain; everything else in that chain is redesigned. Written as CHAIN:list groups with single positions or ranges — A:1-44,46-66 B:5,7. Positions are 1-indexed within their chain, not author residue numbers, so the chain must be numbered from 1 with no gaps or insertion codes. Use this to redesign a liability patch while leaving a binding interface untouched: list the complement of the patch. Leave blank to redesign the whole chain.

Runtime

Preset	Typical min
standalone	~1 min

Ranked candidate sequences with per-position score and overall ProteinMPNN recovery, downloadable as FASTA from a run of your own. Pair downstream with AlphaFold2 or ColabFold to confirm the predicted fold.

References

Starter check: 8 sequences

Confirm your backbone file parses and you named the right chains, before sampling a full batch.

What you need

A backbone structure file (.pdb or .cif) and the chain ID(s) you want new sequences for. Every other chain stays fixed as context.

What it costs

About $0.03, and about 1 min of GPU time. That is an estimate from these exact settings — you are billed for the compute the job actually uses, never more than the estimate.

Then what

Raise the sequence count to 50 or more. Raise the sampling temperature too if the first eight came back near-identical to each other.

Load these settings →

A run we actually did

The target

Hen egg-white lysozyme — PDB 1HEW, chain A, 129 residues.

Its real sequence is known, so every sequence the model writes can be scored against the one nature uses. A de-novo backbone has nothing to compare against, which makes a solved structure the only way to see whether the model is behaving.

What we put in the form

Backbone PDB — 1HEW.pdb: The crystal structure, downloadable below. Only the backbone atoms are read — the sequence in the file is discarded before design.
Chains to design — A: The single protein chain. 1HEW's other chain is a sugar ligand, not protein.
Number of sequences — 2: Deliberately tiny. This run existed to prove the pipeline worked end to end, not to produce a design set.

Download the structure we used (1HEW.pdb)

What came back

Two sequences, 129 residues each, recovering 53% and 50% of the native lysozyme sequence at scores of 0.758 and 0.756. The archived payload pre-dates the fields that record the designed chain and the sampling temperature back into the result, so those two read as em-dashes in the table below; everything else is as it was returned.

Sequences (2)

Designed chain(s): —

#	Score	Recovery	Sequence
1	0.758	0.53	MKYSRCELAKKLKALGLDNYNGYPLGYYVCMAKYESGYDTGATKKQADGSTLYGLFQLNSATWCDDGTTPNATNTCNTPCSDLLKDDITTQVKCLKIAISSGNGLNIFPGYSSNCKGTDVSKYIEGCDL
2	0.756	0.50	MKYSRCELAKKLKELGMDNYNGYPLGEWVGTAYYESNYDTKYTKKQSDGSTLYGLFAIDSGTWCDDGNTPNATNTCNIKCSDLLKDDITNQVSCVKKIVSSGKGLDVFPGYKKNVKGTDASKWIEGCDL

How to read that

Recovery is the fraction of positions where the model picked the residue nature actually uses. The score legend on this page puts the usable line at 0.4 and excellent at 0.6, so about 0.5 on a real backbone is healthy — you would be suspicious of 0.95, which usually means the native sequence leaked into the run. Score runs the other way: it is the model's negative log-likelihood per residue, so lower is more confident, and below 1.0 is excellent.

What happened next

Nothing — this run existed to prove the pipeline. On a real backbone the next step is to raise the sequence count to 50 or more, then fold the best few with ESMFold or ColabFold to see which of them actually adopt the shape you designed them for.

Recorded from that one run, and took 24 seconds end to end, cold container included. That is what it did, not a quote for yours — the estimate on the form is the live number for your own inputs.

What good looks like

Each tool reports a subset of these depending on whether it does design, sequence recovery, or structure prediction.

ipTM: Predicted confidence in the contact between two chains, on a 0 to 1 scale. Higher is better: 0.65 to 0.75 depending on the tool. Individual tools set their own pass bar a little either side of that. On a multi-chain target the number may cover the target's own chain–chain interface as well as the binder's.
pLDDT: Per-residue confidence in the predicted fold. Higher means the model is more sure of that part of the structure.
i_pAE and pAE: Predicted alignment error, at the interface (i_pAE) or across the whole structure (pAE). Lower is better.
ProteinMPNN recovery: Fraction of native residues recovered when ProteinMPNN redesigns a known sequence on its native backbone — the folded shape on its own, before a sequence is chosen for it. Higher is better: > 0.4 usable; > 0.6 excellent.

Related tools on Ranomics

If you are picking between ProteinMPNN and a sibling algorithm, these run on the same hub against the same target.

Run AlphaFold2 online: You have a sequence and want the most trusted 3D prediction of it, with per-residue confidence you can act on. It searches for related natural sequences first, which is where the accuracy comes from and where the time goes. For a faster answer use ColabFold or ESMFold; to score a binder against its target use Boltz-2.
Run ColabFold online: You have a sequence and want its 3D shape in a minute or two, trading a little accuracy for speed. It skips the search for related natural sequences that full AlphaFold2 runs. Use it to triage a batch; use AlphaFold2 when the answer has to be right.
Run ESMFold online: You have one protein sequence and want its 3D shape in about 30 seconds. No search for relatives, so it works on designed or orphan sequences that have no natural family to align against. One chain only — for complexes use ColabFold or AlphaFold2.

Questions people ask about ProteinMPNN

Can I run ProteinMPNN online without a local GPU?: Yes. Upload a backbone PDB and Ranomics Tools runs ProteinMPNN on a dedicated GPU in seconds. You get ranked sequence redesigns plus per-position recovery, with no install and no CUDA setup.
How much does one ProteinMPNN job cost?: Billing is by the second. A typical ProteinMPNN job costs a fraction of a cent because the model finishes in under a minute on most backbones. New accounts start with a $20 wallet balance, which is enough for thousands of runs.
Does ProteinMPNN need an MSA?: No. ProteinMPNN is a structure-conditioned sequence model. It takes the backbone coordinates directly and never queries an MSA. That makes it the fastest route from a designed backbone to a sequence you can synthesise.