ColabFold

ColabFold is a fast no-install online structure prediction tool you can run through tools.ranomics.com on a dedicated GPU. The quick pass while you are still iterating: no Colab notebook to babysit, no MMseqs2 round trip on your own laptop, and no queue to sit in between attempts.

Example run — Not a natural protein — a 101-residue sequence ProteinMPNN wrote for a backbone, folded here as an independent check. See what came back →

Fold inputs

Preset

Standalone, one FASTA Paste a FASTA. Multi-record FASTA is treated as one multimer fold (chains joined). Up to 600 aa total. Runs on A100-40GB. Batch for many fold targets Fold up to 200 independent targets in one job. Each record can be a monomer or a multimer (use : inside a record to break chains). Per-design results stream into the job page as folds complete.

FASTA

Paste one or more chains in FASTA format. Monomer or multimer (up to 600 aa total). Canonical 20 amino acids plus X for unknown. No MSA will be computed. Pair with AF2 (D2) if you need the full MSA-backed fold.

Fold targets (FASTA or one per line)

Up to 200 independent fold targets per submit. Each record can be up to 600 aa total across chains. For a multimer record, separate chains with : on the sequence line. Lines without a > header are auto-named. Detected records: 0.

Number of recycles

How many times AF2 cycles predictions back through the model. Default 1 for speed; up to 5 for a tighter prediction.

Group

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

You pay for the GPU time your run uses. Cancelling does not refund GPU time already used.

About ColabFold

What it is

Predicts a structure from a sequence by skipping the search for related natural sequences that full AlphaFold2 runs. Same AlphaFold2 weights, less evidence to work from — so a little less accurate and a lot faster. Useful for triaging a batch of sequences or getting a quick look at a well-behaved fold. ColabFold, Mirdita et al., Nature Methods 2022.

When to use

You want a structure quickly and can live with slightly less accuracy than full AlphaFold2.
You are folding a batch of sequences one after another and throughput matters more than the last few points of confidence.
Your target is a well-behaved single chain or small complex with no unusual chemistry.

Prerequisites

Single-letter FASTA sequence(s).
Targets with deep evolutionary signal. Multi-domain or low-information sequences underperform without MSA.

Input glossary

Sequence. Paste FASTA. Use : as a chain separator for multimers.
Recycles. Model recycles. ColabFold default is 3; reduce for speed if your target's fold is well-known.

Predicted PDB with per-residue pLDDT and PAE, plus pTM and ipTM on a multimer. An ipTM of 0.6 or more is a plausible interface and 0.75 or more is strong. On a run of your own the structure downloads from the results page. The PAE matrix downloads only from a standalone run — a batch returns no PAE file.

References

A past run we actually did

The target

Not a natural protein — a 101-residue sequence ProteinMPNN wrote for a backbone, folded here as an independent check.

This is the step that catches design failures before they cost anything at the bench. ProteinMPNN scores its own output, but those numbers only say the sequence suits the backbone it was handed — they are not a prediction that it folds. Re-folding it here, with a model that never saw that backbone, is an independent check. This run is one that did not come back clean.

What we put in the form

Preset — Standalone, one FASTA: One design, folded on its own. The batch preset takes a whole set of MPNN outputs in a single submission, which is the usual way to run this once the loop is worth having.
FASTA — 101 aa, one chain: One MPNN sequence for that backbone, pasted as FASTA.
Number of recycles — 2: How many times the model refines its own answer. More recycles let a borderline fold settle and use proportionally more GPU time; 2 is enough to see whether a design is in trouble.

What came back

A mean pLDDT of 61.05 and a pTM of 0.62 — under any threshold you would set, a fail. But the mean is the wrong summary here, and opening Per-residue pLDDT shows why. The first 22 residues are red, every one below 50. The other 79 average 67.3 and climb to 89.9 at residue 50, with 27 residues at or above 70. This is not a design that failed everywhere. It is a folded core with a 22-residue tail hanging off the front.

Fold result standalone

mean pLDDT

61.05

pTM

0.620

length

101 aa

Recycles: 2 · Templates: off · MSA: single_sequence

Per-residue pLDDT (101 residues)

blue > 90 · green > 70 · amber > 50 · red ≤ 50

How to read that

One number for a whole chain averages together the parts the model is sure of and the parts it is not, and the two failures that hides need opposite responses. Uniformly low across the whole strip means the sequence does not fold and the design is dead. Low at one end and high through the middle — this run — means most of the design is fine and a specific, locatable piece of it is not. The mean cannot tell those apart; the strip can. Read the trace before deciding the fate of a design, because the difference between them is a redesign versus a 22-residue edit. It is still a fail as submitted: a floppy 22-residue terminus is a protease target and it will not behave in an assay. But you now know which 22 residues, which is a different morning's work.

What happened next

The move a result like this points to is to trim the tail and re-fold. If the remaining 79 residues hold up alone, the core was real and the terminus was the whole problem. If the core collapses without it, the fold depended on that tail after all and it is the backbone that needs revisiting, not the sequence. Either way it is one more cheap fold before anything gets ordered, which is the reason this step is in the loop at all.

Recorded from that one past run. That is what it did, not a prediction for yours.

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 — the guide for the tool you ran states its one. 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 ColabFold 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 ESMFold online: You have one protein sequence and want its 3D shape, fast. 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.
Run ProteinMPNN online: You have a backbone — a 3D shape with no sequence decided yet — and need amino-acid sequences that will fold into it. You get ranked candidates to pick from. To generate the backbone in the first place, run a binder design tool and feed its PDB in here.