Protein · RNA · DNA structure prediction
Five structure engines, one input — and what they disagree about
Predict protein, RNA and DNA structures and complexes online, with no install. Every analysis runs five engines independently — OpenFold3, Boltz-2, Chai-1, OpenDDE and RhoFold+.
Protein and RNA chains are folded with a multiple sequence alignment, searched on our own infrastructure across 82 million sequences. Which engines use it →
You get every structure, where the methods agree and where they differ per residue, detected pockets and interfaces, ligand docking, and the provenance behind every result — including how close your target sits to what the models were trained on.
Built especially for nucleic acids, where structural assumptions and model limitations matter — and where tools tuned on proteins quietly stop working.
Free to try · no installation · five analyses a month
Complex analysis
HIV-1 TAR · Tat peptide · argininamide
3 methods
Checks
- Chai-16 steric clashes · 0.47 Å closest
- Boltz-2no clashes
- OpenFold3no clashes
HIV-1 TAR ↔ Tat peptide
16–25 residues in contact
depending on which method you ask
See the whole analysis →What do you want to analyse?
Predict a structure
RNA · DNA · protein
Start →Predict a complex
RNA · DNA · protein · small molecules
Start →Compare structures
Against an experimental or predicted structure
Start →Inspect a structure
Pockets · interfaces · structural checks
Open a structure →More workflows
Dock a ligand
Protein pockets · available on request
Batch submission
Many sequences in one run · paid
Custom analysis
Something not listed here · get in touch
Premium
Need more compute, or something built for your target?
The free tier is five analyses a month on public sequences. Paid customers get dedicated GPU capacity, batch submission, an EU-only data path with zero retention, API access, and development work on analyses that do not exist yet. Pricing depends on what you actually need, so the first step is a conversation rather than a plan you pick from a table.
Email info@rnafold.comRun on real targets
Every example below is a full report you can read without an account — including the ones where the answer was not the one we wanted.
Protein · kinase
CDK2
Pocket at rank 1 · 22/22 residues
298 aa, from sequence. The canonical ATP-competitive site, every residue of it, top-ranked.
Protein · GTPase
KRAS G12C
Rank 1 · 6/21 residues
The switch-II cryptic pocket. Found and top-ranked, but only a third of the site — which is the honest number, not the flattering one.
Viral RNA
HCV IRES domain IIa
Binding site at rank 1 · 80% of residues
55 nt element, from sequence alone. Eight of the ten real binding-site residues.
Complex · RNA + peptide + ligand
HIV-1 TAR · Tat · argininamide
16–25 residues in contact
Three methods, three answers about the interface, and one of them with six steric clashes. The spread is the output.
More than a prediction
Getting a structure is the easy part now. Knowing what it does and does not support is the work.
Multiple methods
Several structure-prediction methods run on every analysis, and you see where their outputs agree and where they diverge — per residue, not as a summary.
Structural checks
Contradictions between what you declared and what the coordinates contain, missing defining features, steric clashes, absent interfaces — reported before you interpret anything built on top of them.
Evidence and provenance
The structures, the measurements, the method versions and parameters, and how close your target sits to what those methods were trained on, where the cutoff is published.
Especially useful for RNA and other nucleic acids, where protein-focused assumptions do not always transfer — and where we can show you exactly which ones did not.
How an analysis reads
01
Check
Contradictions between what you declared and what the coordinates contain. Steric clashes. Whether an interface exists at all.
02
Compare
Where the methods agree and where they differ, per residue, as a measurement rather than a summary.
03
Inspect
One viewer, every method, the disagreeing residues highlighted.
04
Continue
Pockets from a structure, interfaces from a complex, docking from a pocket. Each analysis runs on the last one’s output.
What we don’t claim
Better here than discovered halfway through an analysis.
- Whether a molecule binds. Nothing here measures binding, and the metrics that looked like they did were tested against real binders and decoys and separated them at chance.
- Which method is right. We report where they differ; choosing between them is yours, with the provenance in front of you.
- A druggability or quality score for a pocket. The published one is a protein-trained model that scores real nucleic-acid sites at zero.
- Whether a fold is novel — only how close it sits to what the methods were trained on, and only where the training cutoff is published.
Five analyses a month, free
Sign in and paste a sequence. No call, no pitch. If it turns out to be useful for proprietary work there is an EU-only data path with zero retention — a later conversation, not a gate on trying it.
Start an analysis →