OpenDDE co-folding

OpenDDE co-folding is a no-install online all-atom complex prediction tool you can run through tools.ranomics.com on a dedicated GPU. DNA, RNA, cofactors and small molecules are first-class parts of the input rather than something to strip out before folding, so the complex is modelled as it actually exists.

Example run — Bovine trypsin with benzamidine bound — chain A of PDB 3PTB, 223 residues, plus the ligand CCD_BEN. See what came back →

Co-folding inputs

Checkpoint

General co-folding Any mix of protein, DNA, RNA, and ligand. The default. Antibody-antigen (ABAG) The antibody-antigen specialised checkpoint. Use for antibody or nanobody complexes.

Input mode

Guided (fill in each entity) One box per entity type. We assemble the OpenDDE JSON for you. Paste an OpenDDE JSON spec Power users. Validated against the real schema and the same size limits.

Sampler

Samples / seed

Diffusion steps

Recycles

Starting seed

Seeds

Predictions returned = seeds × samples.

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 OpenDDE co-folding

What it is

Folds a whole complex at once when the complex is not all protein — protein with DNA, protein with RNA, protein with a bound small molecule, or any mix of those written into a single specification. Every atom is modelled, not just the protein backbone. It is the multi-molecule counterpart to Boltz-2, which is faster but protein-only. OpenDDE, Aureka AI Research, Apache-2.0.

When to use

Your complex has something in it other than protein: DNA, RNA, or a bound small molecule.
You are modelling an antibody or nanobody together with its antigen (use the ABAG checkpoint).
You want an AlphaFold3-style all-atom prediction without standing up the pipeline yourself.

Prerequisites

Sequences for each polymer chain (protein / DNA / RNA).
For ligands: a CCD code (e.g. CCD_ATP), a bare SMILES string, or a bundled FILE_*.sdf reference.

Input glossary

Checkpoint. General for any entity mix, or ABAG for antibody-antigen complexes.
Entities (guided). One textarea per entity type. Proteins / DNA / RNA are entered as FASTA (>id headers optional); ligands one per line. This adapter assembles the OpenDDE JSON for you.
Entities (JSON). Or paste an exact OpenDDE spec. It is validated against the real schema and re-checked against the same size limits — the JSON mode is not a way around the ceilings.
Samples / steps / recycles. Sampler settings. Samples is how many structures per seed; steps and recycles trade compute for quality.
Seeds. How many seeds to run from the starting seed. Total predictions returned is seeds × samples.

A ranked set of predicted complexes (mmCIF/PDB) with the model's own confidence ranking score per prediction. On a run of your own every row carries its structure, to download or to open in the viewer.

References

A past run we actually did

The target

Bovine trypsin with benzamidine bound — chain A of PDB 3PTB, 223 residues, plus the ligand CCD_BEN.

OpenDDE is for complexes that are not all protein, so an all-protein example would be demonstrating the one case it is not for. This is the textbook protein-plus-small-molecule pair: benzamidine sitting in the trypsin S1 pocket, deposited in 1982 and used to check docking methods ever since. It is also a deliberate correction. The first run this page ever showed folded a single ubiquitin chain, which has no second entity at all — its ipTM of 0 was arithmetically correct and told you nothing, because there was no interface to score.

What we put in the form

Protein chains (FASTA) — the 223-residue trypsin chain: One record. Worth a moment if you are preparing this yourself: 3PTB is numbered in the chymotrypsin convention, so the chain carries three insertion-coded positions — 184A, 188A and 221A. A script that reads residue numbers and ignores the insertion letter silently drops all three. Our first attempt at this example came back 220 residues instead of 223 for exactly that reason, and handed the model a different protein. Count what you extract against the deposited length before you submit it.
Ligands (one per line) — CCD_BEN: Benzamidine, by its three-letter Chemical Component Dictionary code. A bare SMILES string works too, but a CCD code names one unambiguous molecule and needs no interpretation.
Samples / seed — 1: One prediction per seed. All of the variation in this run comes from the seeds instead.
Seeds — 4: Four independent starting points, so four predictions. This is the input that matters most here, and the results below are the argument for it.
Diffusion steps — 200: The upstream default, left alone. More steps trade compute for quality; nothing about this run suggested it was step-limited.
Recycles — 10: Also the default. Worth raising on a complex that comes back geometrically incoherent, which this one did not.

What came back

Four predictions in about three and a half minutes, and they split cleanly two and two. The top pair scores 0.67 and 0.66 on the ranking; the bottom pair 0.48 and 0.47. ipTM runs 0.471 to 0.689, pTM 0.433 to 0.625, and pLDDT 42.8 to 52.9.

OpenDDE co-folding — general general

Sampler

1 sample · 200 step · 10 cycle

OpenDDE co-folds all supplied entities into one complex and ranks the predicted structures by its own confidence head. Higher ranking score is more confident.

Candidates (4) general co-folding

Returned 4 / 4

#	Ranking score ?	pTM ?	ipTM ?	pLDDT ?	3D	Structure
1 Top	0.67	0.614	0.689	51.4	—	—

2	0.66	0.625	0.670	52.9	—	—

3	0.48	0.433	0.493	42.8	—	—

4	0.47	0.462	0.471	44.4	—	—

Validating in the lab is optional Designs from your own run are yours to download and take anywhere. If you do want them made and measured, Ranomics runs yeast display, mammalian display, BLI and DMS on the same data loop.

Test binders in the lab AI Binder Sprint

How to read that

The useful thing here is where all four columns agree. Ranking, ipTM, pTM and pLDDT each draw the same line between the same two pairs: the top two are 0.689 and 0.670 on ipTM with pLDDT 51.4 and 52.9, the bottom two are 0.493 and 0.471 with pLDDT 42.8 and 44.4. Nothing crosses over. The columns measure different things — ipTM the interface, pTM the whole complex, pLDDT the per-residue confidence — and all four put the same two predictions on top. Only the split is that robust: within each pair pTM and pLDDT both invert the ranking's order. Read the split as a ranking among these four and nothing more: every one of them sits in the 40s or low 50s on pLDDT, so the split says which pair to look at first, not that any of them is a good structure. Now the part that should change how you run this tool. The form defaults to a single seed, and that seed is the one ranked last here — the bottom row, 0.47 on the ranking, ipTM 0.471, pLDDT 44.4. (The table ranks the predictions without naming their seeds; the run's own file is where that pairing lives.) Run this target once, with the defaults, and you would have concluded it does not co-fold with its ligand — when half the seeds land at 0.67 and 0.66 and put the model's confidence in the same place. One prediction on this tool is not a small version of four. It is a coin flip you cannot see the result of.

What happened next

Nothing — this run was captured for this page and stopped here. What it would take next is the one thing the table cannot show: whether the top two put the benzamidine in the same pocket. Two predictions can agree on every score and still dock a ligand in different places, so that question needs the structures themselves, which every row carries on a run of your own. There the viewer opens per row and two rows can be open at once, but each shows a single structure and nothing superposes them — the comparison is yours to make by eye. Agreement on the pocket means more seeds will sharpen the pose; disagreement means the scores were telling you about the fold and not about the binding site.

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. 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.