AI for Bio · Model

RhoFold+/DRfold2/trRosettaRNA

Three independent RNA 3D prediction lines from three academic groups, not mutually comparable

Structure & co-foldingAcademic labChina, Singapore, United StatesMaintained only中文

Overview

NameRhoFold+/DRfold2/trRosettaRNA · aliases: RhoFold+, DRfold2, trRosettaRNA, trRosettaRNA2
Full nameFamily entry; no single expanded name
Ambiguous names (not auto-linked)RhoFold, DRfold, trRosetta
FamilyCovers only these three RNA 3D prediction lines, including the earlier trRosettaRNA; other RNA models such as RNAPro are separate entries.
  • RhoFold+ · †
  • DRfold2 · †
  • trRosettaRNA2 · †
OrganizationCUHK / Harvard / MIT collaboration, National University of Singapore group, Shandong University (Qingdao) group
Organization typeAcademic lab
CountryChina, Singapore, United States
Model classRNA 3D structure prediction
DomainStructure & co-folding (also: Genomics & RNA)
First release† unverified · preprint
Basis: Preprint v1 of the earliest member, RhoFold+ (member and date not re-verified, unverified)
Note: Family entry: the first release is the earliest member's, not a joint release source
Other milestones
  • First public evidence: † · The earliest member's preprint v1 (unverified) source
  • Preprint v1: † · DRfold2 bioRxiv v1 source
  • Journal online: † · RhoFold+ journal publication source
  • Code release: No sourced date† · No sourced first code-release date for any of the three lines
Latest update† partly verified · journal
Basis: trRosettaRNA2's journal publication (DOI and date second-hand, Crossref rate-limited)
trRosettaRNA2 appeared in a journal; DRfold2 was published 2026-02-17 and RhoFold+ on 2024-11-21
Note: A member event inside the family, not a joint update of all three lines (unverified) source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · All three published in journals between 2024 and 2026; repositories are in maintenance
OpennessCode: Partly open
Weights: Research use only
RhoFold+ ships weights and inference code under Apache-2.0, the DRfold2 repository states no licence, and trRosettaRNA2 offers only a free web server
Benchmarks reported2 counted (board 1, retrospective 1)
Wet-lab validationNone recorded
Unverified fields† first_release.date, family.members[0].first_release.date, family.members[2].first_release.date, latest_update.date, benchmarks_reported[1]
As of2026-11-16

What it does

The three lines predict RNA 3D structure with, respectively, an RNA language model plus a structure module, a compound language model plus denoising, and a Transformer that predicts geometric restraints before energy minimisation. With only a few thousand experimental RNA structures, all three lean on sequence databases and language models, and absolute accuracy stays well below protein co-folding such as AlphaFold3.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 2 of 2 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
01 CASPOfficial leaderboard or resultsCASP16 automated-server section (trRosettaRNA2); CASP15 retrospective (RhoFold+)BoardFamily entry: trRosettaRNA2 (server) took part in CASP16; RhoFold+ is a CASP15 retrospective evaluation †partly verified · evidence
25 RNA-PuzzlesRetrospective dataset usesingle-chain RNA-Puzzles targets, retrospective (RhoFold+)Third-partyRhoFold+ retrospective on 24 RNA-Puzzles; DRfold2/trRosettaRNA2 unconfirmed; family cell †partly verified · evidence

Other evaluations (outside the 44)

  • Self-built RNA test set (self report) Self-reported: The DRfold2 preprint compares against a co-folding model on 28 self-assembled structures, excluding targets over a length cut-off (self-reported)
  • Evidence from a public challenge (third party evaluation) Third-party: In part one of an RNA 3D folding challenge the top three tied with the previous blind-test winner within statistical error, and the best strategy was template retrieval without deep learning (abstract, second-hand)
  • Review comparison (third party evaluation) Third-party: A review compares six RNA structure methods on one set of data; the ranking itself was not obtained (third party)

Developers report: Developers report each line beating contemporaneous methods on retrospective RNA sets; the protocols differ and are not comparable inside the family (self-reported).

Independent evidence: The CASP16 organisers' assessment of the automated-server section is the firmest third-party evidence, and it belongs to the trRosettaRNA2 line alone.

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesA few thousand RNA structures in the PDB plus sequence databases such as RNAcentral; the RhoFold+ language model was pretrained on about 23.7M sequences source
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: Three academic groups and no company: a CUHK collaboration with Harvard and MIT (RhoFold+), a group at the National University of Singapore (DRfold2), and a group at Shandong University in Qingdao (trRosettaRNA).
  • Funding: Academic projects; funding amounts not public
  • Products and partnerships: RhoFold+ releases weights and inference code, the DRfold2 repository states no licence, and the trRosettaRNA2 site says its server is free for all users; the three compete with each other.
  • Related models: RNAPro, AlphaFold 3

Reading notes

  1. The three lines use inconsistent RNA-Puzzles protocols and two are unconfirmed, so do not rank them on the page.
  2. The CASP16 server-section placing is the firmest evidence, but only for trRosettaRNA2.
  3. Template retrieval remains very strong on RNA; deep learning's edge on template-free targets is unproven.
  4. The family's first-release date is the earliest member's and has not been re-verified.

Caveats

  1. The first-release member and date were not re-verified and are shown as unverified.
  2. The latest update is one member's journal publication, not a joint update.
  3. Experimental RNA structures are scarce, so absolute accuracy is low under every protocol.
  4. Openness differs across the three; only one line states a clear licence.

Sources

  1. RhoFold+ journal paper (journal) · †
  2. DRfold2 preprint full text (preprint)
  3. trRosettaRNA2 preprint (preprint)
  4. DRfold2 journal publication (journal)
  5. CASP16 server-section presentation (official site)

Corrections

2 corrections
FieldWasNowBasisFound
latest_update.descriptionDRfold2 existed only as a preprintDRfold2 was published in a journal on 2026-02-17source2026-11-16
benchmarks_reported.bench-25-rnapuzzlesAll three lines had RNA-Puzzles resultsOnly the RhoFold+ retrospective holds; the other two are unconfirmed (unverified)source2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.