Protein–Nucleic Acid co-folding and Specificity Prediction on Vecura

For nucleic acids in particular, co-folding is the stronger method, and the reason is specific. Classical docking holds both partners rigid and searches for the best way to place them against each other, which assumes each one already has the shape it will take once bound. Nucleic acids break that assumption harder than almost anything else: DNA bends and kinks sharply around the proteins that read it, and RNA routinely refolds on binding. Co-folding makes no such assumption. It predicts the conformation of each chain and the interface between them at the same time, so induced fit is part of the prediction rather than an error the scoring function has to absorb, and the models doing it learned recognition from experimental complexes rather than from a generic energy term.

What co-folding needs, and what it gives back

The one thing to settle before you start is your input. Co-folding runs from sequence, so if your starting material is a pair of experimental structures you will be supplying their sequences and letting the model rebuild the complex rather than posing the structures themselves. For most protein–nucleic acid work that is the right trade, the bound conformation is usually not the one in your unbound structures anyway, but it is worth knowing in advance rather than discovering mid-run.

Choosing an operation

Protein–DNA co-folding

使用案例回放Protein-RNA Co-folding查看此使用案例

Predicts the structure of a protein bound to a DNA partner from the two sequences. This is the operation to use if you came looking for protein–DNA docking: same deliverable, different input.

Use it when you need a structural model of a protein–DNA interaction and do not already have one, a transcription factor on its recognition site, a repair enzyme on damaged DNA, a nuclease on its substrate. It answers where the partners sit. Once you have that complex, specificity prediction answers which sequences the protein actually prefers there.

输入

Protein sequence; DNA sequence (ATCG)

输出

Predicted protein–DNA complex structure

模型Boltz-2.1ProtenixOpenDDEOpenFold3Chai-1IntelliFold

Which model

Seven models predict protein–nucleic acid complexes. They divide by how you hand them the partners, and that is the practical basis for choosing:

Reach for Chai-1 or IntelliFold when the assembly is one protein plus one nucleic acid, the separate fields make that explicit and hard to get wrong. Reach for the assembly-style models when the complex is larger or mixed: several chains, or a nucleic acid together with a ligand, where you need control over what goes in and in what order. Note that OpenFold3 exposes this under predict_structure rather than generate_biomolecule_structure, which matters if you are chaining operations by name.

Protein–RNA co-folding

使用案例回放Protein-RNA Co-folding查看此使用案例

Predicts the structure of a protein bound to an RNA partner — the route for anyone searching protein–RNA docking.

A DNA duplex is close to a regular B-form helix, so the model has comparatively little to decide about the partner's own shape. An RNA partner has a tertiary fold that the model is predicting at the same time as the interface, which means the fold itself is part of what you are trusting. Treat an RNA complex with more scepticism than a DNA one at the same confidence score, and check the RNA fold against anything you know about it independently.

输入

Protein sequence; RNA sequence (AUGC)

输出

Predicted protein–RNA complex structure

模型AlphaFold3Boltz-2.1ProtenixOpenDDEOpenFold3Chai-1IntelliFold

Which model

The same seven models, chosen on the same basis as for protein–DNA co-folding above. Where the RNA chain goes:

RNAPro is not among them despite being an RNA structure model: it takes RNA sequences only, with no protein input, so it predicts RNA structure alone rather than protein–RNA complexes.

Protein–DNA specificity prediction

使用案例回放Protein-DNA Specificity Prediction查看此使用案例

Predicts which DNA bases a protein prefers at each position of a bound nucleic acid, given the complex structure. This is also what "binding specificity prediction" refers to on the platform — one operation, taking one complex PDB, not two separate capabilities.

Use it when the geometry is settled and the open question is recognition: which bases are read out, and how tolerant each position is. It does not produce or refine a structure. If the complex you feed it came from co-folding rather than experiment, the prediction inherits whatever error is in that structure, which is worth stating in any result built that way.

输入

Protein–DNA complex structure (PDB)

输出

Per-position nucleotide preferences

模型Na-MPNN

Which model

NA-MPNN (Baker Lab) is the only model exposing this operation, so there is no neighbour to choose against. It is a message-passing neural network for backbone-conditioned nucleic acid sequence design and protein–DNA binding specificity prediction. Given a complex structure it auto-detects the chains present and returns predicted sequence preferences conditioned on that fixed geometry. It scores preference from structural context — it does not compute binding free energy and is not a substitute for physics-based simulation. When a protein chain is present its sequence is held fixed; NA-MPNN designs the nucleic acid, not the protein.

Next step

A co-folded complex is a prediction, not a measurement, so the natural next group is Molecular Dynamics / Trajectory / Ensemble Analysis: running a short simulation on the predicted complex shows whether the interface holds together or comes apart, which is the cheapest available check on whether the structure is worth building further work on. For a complex you intend to design against rather than just inspect, that stability check belongs before any downstream design step.