De novo RNA tertiary structure prediction for targets without templates
Research-grade: blind assessments through 2025 show template reliance and frequent clashes, while data and cryo-EM RNA structures are growing quickly, so it is slightly easier than fold switching (58). Application value 24 is indirect: RNA therapeutics and RNA-targeting drugs are a growing but small slice of discovery, below the ADME and antibody problems here. Leverage 33 because RNA structure blocks one subfield, not protein-centred pipelines.
The problem
Predicting the 3D shape of an RNA molecule when no similar shape has been seen before.
- 1RNA chain folds around ions
- 2Few solved shapes to learn from
- 3Models copy known templates
- 4New folds expose the weakness
- 5Need template-free accuracy
The task. For RNAs of 50 to 500 nucleotides whose fold has no template in the PDB, predict all-atom coordinates, including non-Watson-Crick pairs and long-range contacts. Accuracy is measured by RMSD (average atom distance after superposition), TM-score adapted to RNA, and interaction network fidelity (INF), which scores whether the right base pairs and stacks are present.
Path to understanding
foundations → problemThe research entries this problem rests on, in the order to read them. Each links to its explainer.
- Start with
- Reverse-mode automatic differentiationAlgorithm
- Fast matrix multiplicationAlgorithm
- Step 2
- The transformer architectureAlgorithm
- Step 3
- Biomolecular structure prediction (AlphaFold-style)Algorithm · host
- Then
- De novo RNA tertiary structure prediction for targets without templatesThe problem
State of the art
- CASP16 (2024). The practical-outcomes analysis (Abriata and Dal Peraro, preprint 2025) reports that nucleic-acid prediction "lags considerably", relies heavily on 3D templates and expert human intervention, and that even AlphaFold 3 shows substantial limitations; Elofsson's assessment found that 72% (309 of 420) of the multimeric RNA models his group generated with AF3 had steric clashes and were discarded. The Das-lab assessment concluded that deep learning has fuelled limited progress in RNA and RNA-complex modelling.
- CASP16 ensembles. The CASP16 multi-conformation assessment (preprint) found three RNA targets and several oligomeric RNA states consistently below TM-score 0.75, naming RNA and large multimeric assemblies as the main open challenges.
- Stanford RNA 3D Folding Kaggle competition, part 1 (2025). Over 1,700 teams predicted 43 previously unreleased structures. The top three entries matched the CASP16 winners within statistical error, and the best strategy was a template-discovery pipeline rather than deep learning; a combined model, RNAPro, was reported to retrospectively outperform the individual entries (preprint, not peer reviewed). Part 2 ran January to March 2026 with hidden targets including new-to-nature folds, small-molecule and protein complexes; no peer-reviewed analysis of its results had been located as of October 2026.
- Not shown. No method has demonstrated the benchmark above on template-free targets in a blind test.
Realistic applications
Template-free RNA modelling would speed up structure-guided design of riboswitch- and RNA-targeting small molecules, mRNA therapeutics and vaccines (stability, translation), guide RNAs and aptamers, and the annotation of the large number of non-coding RNAs of unknown structure. It would also test whether physics-informed or data-driven models can capture the energetics that coevolution does not.
What counts as solved
A blinded evaluation (CASP, RNA-Puzzles or a pre-registered Kaggle-style benchmark with held-out experimental structures solved after the training cut-off) in which a method meets the median RMSD / INF-all / physical-validity criteria above on at least 30 template-free targets, with the analysis published in a refereed venue and the method openly reproducible.
How it gets verified
ReproductionA training run, experiment or computation rerun by someone else.
A method must meet the stated median RMSD, INF-all and physical-validity criteria on at least 30 template-free targets in a blinded evaluation (CASP, RNA-Puzzles or a pre-registered benchmark with structures solved after the training cutoff), published in a refereed venue and reproducible with released code.
Claims under verification
+ add a claimNo claimed results on the Verification board for this problem yet.
Where the work is
Data scarcity and measurement are the dominant limits; modelling choices matter, but there are only a few thousand non-redundant RNA structures and chemical-probing data are indirect.
Bottlenecks
ranked by votes- #1EmpiricalFew non-redundant RNA 3D structures
Training data are dominated by rRNA, tRNA and riboswitch families; new-to-nature folds in blind tests have no analogues.
Community grades
Sign in to add your own grade.
Connected problems
Progress
- Jan 7, 2026Kaggle RNA 3D folding, part 2 openssource
Hidden targets include cryo-EM structures of new-to-nature RNA folds, small-molecule and protein complexes, and assemblies up to 6,000 nt; closed 25 March 2026.
- Dec 30, 2025Kaggle RNA 3D folding, part 1 analysed (preprint)source
1,700+ teams, 43 unreleased structures; top template-based strategies matched CASP16 winners within error.
- Sep 2, 2025CASP16 ensembles: RNA among hardest targetssource
Three RNA targets and oligomeric RNA states stayed below TM-score 0.75.
- Dec 1, 2024CASP16 RNA assessmentssource
AF3-class models do not clearly surpass template-based or expert-guided approaches on RNA; clashes are common in multimer outputs.