Refolding

Refolding re-predicts a design with an independent structure prediction model and checks whether it agrees with what RFD4-Proteina built. This page covers what refolding is for, what it produces, and how a design can end up with more than one refold. For turning it on, setting up the AF3/OF3 oracle, and the full tag listing, see the refolding guide; for the complete metric and tag reference, see the metrics reference.

Why refold a design

The metrics concept guide splits metrics into four families by what they look at. Refolding sits apart from the other three:

Family	Looks at	Answers
Design	the design alone	Is this a physically sensible structure?
Adherence	design against your input	Did it build what you asked for?
Refolding	the refolded structure alone	Does a folding model believe it?
Cross comparison	input, design and refold together	Does the whole story hold up?

A refold is the output of handing the design’s sequence (and, for templated oracles, part of its structure) to a structure prediction model and asking it to fold that sequence from scratch. If the refold’s confidence is low, or its backbone lands far from the design’s own backbone, that is evidence the sequence does not actually encode the structure RFD4-Proteina drew — independent of whether the design looks reasonable on its own.

What runs underneath a refolding tag

Turning on refolding activates three things together, in order:

1. An oracle (reference) — oracle, oracle_ppi, or oracle_dna — which actually calls AF3/OF3 and reports its own confidence outputs (overall_plddt, ptm_coi, min_ipae_coi, and friends).

2. Refolding metrics (reference) that compare the design against that refold, such as refolded_backbone_rmsd (self-consistency RMSD) or refolded_motif_allatom_rmsd.

3. Pass criteria (reference) that threshold those oracle and refolding outputs into a single {criterion}.pass flag, like ppi.pass or sm_binder.pass.

Refolding is also the most expensive step in the pipeline. A pre-oracle filter (reference) can reject an obviously broken design before it ever reaches the oracle — your biggest lever on cost per success, except when benchmarking per example, where filters should be turned off entirely so you measure the model rather than the filter.

A design can have more than one refold

A pre-oracle transform can fan a single oracle call out into several refolds of the same design — for example, folding the designed chain on its own, separately from the normal “untransformed” refold of the whole complex. When a design has more than one refold, every refolding metric and pass criterion needs to say which refold it reads, so RFD4-Proteina gives each one an address (refold.apo, refold.oracle_ppi, …) and derives is_refold:/exists_refold:/oracle: tags from it automatically — you never pass those on the CLI yourself.

This is what lets a metric like refolded_apo_holo_rmsd ask “does the binder hold its fold once the target is gone?” by comparing an apo refold against the whole-complex one. Supplying that second refold, and pairing the right tags so the comparison doesn’t silently drop, is covered in the refolding guide.

See the output format reference for the full addressing scheme, and Comparing two refolds for writing a custom metric that reads more than one refold.