Structure prediction
Predicts a 3D structure directly from a sequence (OmegaFold) — no template, no MSA, no reference structure needed. Capped at 274 residues.
Structure prediction
What it does
Takes an amino-acid sequence and predicts a 3D structure for it, with no template, no multiple-sequence alignment, and no reference structure required. Use it as a starting point wherever a project has a sequence but no experimental or predicted structure yet — before docking, before contact mapping, before Molecular dynamics, or before any structure-based tool that needs coordinates to work with.
How the prediction is made
The method is OmegaFold. Unlike template-based prediction, OmegaFold works directly from the sequence using a large protein language model, so it needs no external database search and no multiple-sequence alignment step — one sequence in, one structure out.
Reading the output
The result is a single PDB file. Per-residue confidence rides along in the file's own B-factor column, the same place every other structure tool in RxCubed expects to find it — readable directly by anything that reads a B-factor column, including Local structure agreement if you want to compare this prediction against a reference structure.
Input
One sequence, saved as a FASTA file (or plain text). Only the first record is used if a file contains more than one sequence — this tool predicts one structure per run.
Important limitations
Capped at 274 residues. Runtime grows steeply with length, and this is the longest target this deployment has actually measured — a sequence over the cap is rejected up front with a clear message, not silently truncated. This comfortably covers a single structural domain or the kind of designed binder RxCubed's own sequence-design tools produce.
One chain, not a complex. This predicts a single folded chain from a single sequence. For a bound complex between two known chains, use Protein–protein docking instead — a different question, not a bigger version of this one.
A prediction, not a measurement. Like any computational structure prediction, this is a model's best estimate, not an experimentally determined structure — treat it accordingly, especially on larger or multi-domain targets, where accuracy drops off (a real measurement against a known kinase structure found ~5.5 Å Cα-RMSD, compared to under 1 Å on a small single-domain protein).