Papex

Open problem · Chemistry · Open

RNA folding problem

Can the three-dimensional structure of an RNA molecule be predicted from its sequence?

Summary

Predicting the secondary, tertiary and quaternary structure of an RNA molecule accurately from its sequence and environment remains an open problem.

Why it is open

Ignacio Tinoco and Carlos Bustamante described RNA folding as hierarchical: secondary structure, made of helices, loops and bulges, is much more stable than tertiary folding, and the two levels can be separated experimentally by adding or removing magnesium ions [3]. They argued that secondary structure can be predicted from measured thermodynamic data, but that solving the folding problem also needs thermodynamic data on tertiary interactions and a characterisation of metal-ion binding sites [3].

The problem is therefore several problems: predicting base pairs, predicting the three-dimensional fold, predicting complexes with other molecules, and predicting molecules that adopt more than one structure [7, 14]. Each has its own difficulties, from pseudoknots, base pairs that cross rather than nest, which make exact secondary-structure prediction computationally hard [4, 14], to the small amount of structural data available for training [8].

No body defines a resolution. Progress is measured in blind tests, in which groups predict RNA structures before the experimental structures are published [11]: RNA-Puzzles since 2010 [8] and the Critical Assessment of Structure Prediction (CASP) since 2022 [7]. A 2023 perspective by Bohdan Schneider and colleagues states that protein structure prediction "has been solved for many types of proteins by AlphaFold" and argues that data limits stand in the way of the same for RNA [8]. After CASP16, held in 2024 [13], the assessors wrote that their results "confirm a continuing gap in accuracy between nucleic acid structure prediction and the near-experimental accuracy that is routinely achieved in protein structure prediction" [14]. They single out two shortfalls: automated methods trail the best human experts, and accuracy depends on closely related known structures [14]. Youhan Lee and colleagues call automatic prediction of RNA three-dimensional structure from sequence "an unsolved challenge" [13].

What is known

Secondary structure from thermodynamics (1981–2000). In 1981 Michael Zuker and Patrick Stiegler used dynamic programming, a technique from applied mathematics, to find the conformation of minimum free energy from published stacking and destabilising energies; they folded a 459-nucleotide messenger RNA fragment and showed how to add chemical and enzymatic data [1]. With expanded thermodynamic parameters, David Mathews and colleagues reported in 1999 that the predicted lowest-free-energy structure contained on average 73% of known base pairs for domains under 700 nucleotides, against 64% for earlier versions, and that a set of 750 alternative structures contained one with 86% on average [2]. In 2000 Rune Lyngsø and Christian Pedersen noted that secondary structure without pseudoknots can be predicted in polynomial time, but proved that the general problem with pseudoknots is NP-complete, a class of problems for which no efficient algorithm is known, for a large class of reasonable energy models [4]; see the Papex page P versus NP.

Learned secondary structure (2022). Marcell Szikszai and colleagues found that deep-learning models report impressive results when tested on RNA families seen in training, and provided "strong evidence that many existing learning-based models have not generalized inter-family" [6].

Blind tests of three-dimensional structure (2012–2025). The first RNA-Puzzles report, in 2012, covered three puzzles [5]; Schneider and colleagues counted 38 challenges in its first 12 years [8]. Round V, published in December 2024, assessed 18 groups on 23 RNAs and named the steps still to be mastered: identifying helix-forming pairs and non-Watson–Crick modules, stacking helices correctly and avoiding entanglements; only one of the 18 approaches used deep learning [11]. CASP15, with targets released between April and July 2022, was the first CASP to include RNA: 42 groups modelled 12 targets, two independent assessments ranked the same groups first to third, and none of these used deep learning, while deep-learning predictions were "significantly worse" [7]. The global fold was predicted correctly for all targets except two RNA–protein complexes, but non-canonical pairs, choosing among submitted models and multiple structures remained difficult [7]. In CASP16, whose nucleic-acid assessment was published in October 2025, 65 groups from 46 laboratories predicted 42 nucleic-acid targets; "performance on nucleic acids was generally poor", and no prediction of a previously unseen natural RNA reached a TM-score above 0.8 [14]. The TM-score is a global similarity measure on which values above 0.45 count as an acceptable fold [14]. Groups reached that level for 21 of 36 targets, but only two of the 21 lacked a related known structure, a template [14]. All top-ranked groups were human experts (Vfold, GuangzhouRNA-human and KiharaLab), many teams outperformed the AlphaFold 3 server entered as a baseline, and, allowing for templates, accuracy had not notably improved since earlier blind challenges [14].

Data. Schneider and colleagues count about 25 times more protein than RNA entries in the Protein Data Bank, and RNA nucleotides make up only 2% of the residues in structures solved to better than 3.0 Å [8].

Recent claims. In May 2024 Josh Abramson and colleagues at Google DeepMind and Isomorphic Labs reported that AlphaFold 3 predicts RNA structures in recent Protein Data Bank entries more accurately than RoseTTAFold2NA; on ten CASP15 RNA targets it beat the best AI-based entry but "did not reach the performance of the best human-expert-aided CASP15 submission" [9]. An independent 2025 benchmark by Clément Bernard and colleagues found that AlphaFold 3 outperforms most existing methods, but "hardly predicts non-Watson–Crick interactions" and "does not generalize well to orphan structures" unrelated to known RNA families [12]. In November 2024 Tao Shen and colleagues reported that RhoFold+, built on a language model pretrained on about 23.7 million RNA sequences, outperformed existing methods "including human expert groups" in retrospective tests on RNA-Puzzles and CASP15 targets [10]. A Kaggle code competition launched in February 2025 and reported in a preprint, not yet peer reviewed, drew over 1,700 teams; its top three algorithms scored within statistical error of the CASP16 winners on 43 unreleased structures, and the winning strategy searched for templates without deep learning [13]. Its scoring used one atom per nucleotide and a sequence-independent alignment, unlike CASP [13].

Main approaches

Thermodynamic secondary-structure models. Base pairing is predicted by minimising free energy with measured parameters [1, 2]. Strongest evidence: 73% of known base pairs on average for domains under 700 nucleotides [2]. Main difficulty: pseudoknots make the general problem NP-complete [4], and tertiary interactions and ion binding lack the thermodynamic data the approach needs [3].

Evolutionary information from sequence alignments. Alignments of many related sequences are used to infer which parts of an RNA pair and touch [14]. Strongest evidence: in CASP16, targets without templates were predicted better when deep alignments existed, and good performance on one such target, the OLE RNA, together with accurate global secondary structure, suggested that structural information can be extracted from alignments [14]. Main difficulty: high-quality RNA alignments are hard to build and often need an expert's manual work, so far fewer exist than for proteins [8].

Expert-guided three-dimensional modelling. Human predictors guide computational modelling with insights from the literature, sequence alignments, possible templates and their own intuition [13]. Strongest evidence: expert groups ranked first in CASP15 and CASP16 [7, 14]. Main difficulty: accuracy depends on templates [14], and automated methods still trail the experts, which Lee and colleagues link to the limited use of RNA structure prediction in biology [13].

Deep learning. General models such as AlphaFold 3 and RNA-specific ones such as RhoFold+ learn structure from sequences and known structures [9, 10]. Strongest evidence: AlphaFold 3 outclassed other deep-learning approaches on every benchmark set of Bernard and colleagues [12], and RhoFold+ reported retrospective results above expert groups [10]. Main difficulty: there are few RNA structures to learn from [8], learned models generalise poorly beyond known families [6, 12], and in blind CASP16 tests human experts still led [14].

Template-based modelling. A new RNA is modelled on a known structure that it resembles in three dimensions [13]. Strongest evidence: the winning Kaggle strategy matched CASP16 winners within statistical error [13]. Main difficulty: it works only when a related structure exists, and in CASP16 only two acceptable folds lacked one [14].

Open sub-questions

  1. Orphan RNAs. Can any automated method reach a TM-score above 0.45 on most RNAs unrelated to known families, where AlphaFold 3 generalises poorly [12, 14]?
  2. Pseudoknots in practice. The general pseudoknot problem is NP-complete [4]. Do the pseudoknots of natural RNAs, such as those CASP16 predictions failed to recover consistently [14], mostly belong to simple types that an efficient algorithm could still predict?
  3. Learning across families. Under strict inter-family cross-validation, does any learned model predict secondary structure more accurately than the thermodynamic baseline on families absent from training [2, 6]?
  4. Ions and tertiary contacts. Does modelling magnesium ions explicitly improve recovery of the non-canonical pairs and tertiary motifs that CASP16 predictions still missed [3, 14]?
  5. Templates versus learning. How much of the Kaggle winners' accuracy remains when templates above a fixed structural similarity are excluded [13, 14]?
  6. Choosing the best model. Can a scoring method pick the most accurate of a group's five submitted models more often than chance on CASP15 and CASP16 targets [7, 14]?

Work on this in Papex

If you have a testable answer to one of the sub-questions above, you can draft it as a private hypothesis in Papex, or prepare the draft from Claude or ChatGPT with the Papex connector. You review the exact version, sign it and decide whether to publish it; a public record that names this problem is listed on this page. A Papex record shows who signed which version and when. It is not peer review and does not mean the hypothesis is correct.

Key references

  1. Zuker, M. & Stiegler, P. (1981). Optimal computer folding of large RNA sequences using thermodynamics and auxiliary information. Nucleic Acids Research 9(1), 133–148. 10.1093/nar/9.1.133 · Checked 15 Oct 2026
  2. Mathews, D. H., Sabina, J., Zuker, M. & Turner, D. H. (1999). Expanded sequence dependence of thermodynamic parameters improves prediction of RNA secondary structure. Journal of Molecular Biology 288(5), 911–940. 10.1006/jmbi.1999.2700 · Checked 15 Oct 2026
  3. Tinoco, I. & Bustamante, C. (1999). How RNA folds. Journal of Molecular Biology 293(2), 271–281. 10.1006/jmbi.1999.3001 · Checked 15 Oct 2026
  4. Lyngsø, R. B. & Pedersen, C. N. S. (2000). RNA pseudoknot prediction in energy-based models. Journal of Computational Biology 7(3–4), 409–427. 10.1089/106652700750050862 · Checked 15 Oct 2026
  5. Cruz, J. A., Blanchet, M.-F., Boniecki, M., Bujnicki, J. M., Chen, S.-J., Cao, S. et al. (2012). RNA-Puzzles: a CASP-like evaluation of RNA three-dimensional structure prediction. RNA 18(4), 610–625. 10.1261/rna.031054.111 · Checked 15 Oct 2026
  6. Szikszai, M., Wise, M., Datta, A., Ward, M. & Mathews, D. H. (2022). Deep learning models for RNA secondary structure prediction (probably) do not generalize across families. Bioinformatics 38(16), 3892–3899. 10.1093/bioinformatics/btac415 · Checked 15 Oct 2026
  7. Das, R., Kretsch, R. C., Simpkin, A. J., Mulvaney, T., Pham, P., Rangan, R. et al. (2023). Assessment of three-dimensional RNA structure prediction in CASP15. Proteins: Structure, Function, and Bioinformatics 91(12), 1747–1770. 10.1002/prot.26602 · Checked 15 Oct 2026
  8. Schneider, B., Sweeney, B. A., Bateman, A., Cerny, J., Zok, T. & Szachniuk, M. (2023). When will RNA get its AlphaFold moment? Nucleic Acids Research 51(18), 9522–9532. 10.1093/nar/gkad726 · Checked 15 Oct 2026
  9. Abramson, J., Adler, J., Dunger, J., Evans, R., Green, T., Pritzel, A. et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630(8016), 493–500; addendum, November 2024. 10.1038/s41586-024-07487-w · Checked 15 Oct 2026
  10. Shen, T., Hu, Z., Sun, S., Liu, D., Wong, F., Wang, J. et al. (2024). Accurate RNA 3D structure prediction using a language model-based deep learning approach. Nature Methods 21(12), 2287–2298. 10.1038/s41592-024-02487-0 · Checked 15 Oct 2026
  11. Bu, F., Adam, Y., Adamiak, R. W., Antczak, M., de Aquino, B. R. H., Badepally, N. G. et al. (2025). RNA-Puzzles Round V: blind predictions of 23 RNA structures. Nature Methods 22(2), 399–411 (published online 2 December 2024). 10.1038/s41592-024-02543-9 · Checked 15 Oct 2026
  12. Bernard, C., Postic, G., Ghannay, S. & Tahi, F. (2025). Has AlphaFold 3 achieved success for RNA? Acta Crystallographica Section D: Structural Biology 81(2), 49–62. 10.1107/S2059798325000592 · Checked 15 Oct 2026
  13. Lee, Y., He, S., Oda, T., Rao, G. J., Kim, Y., Kim, R. et al. (2025). Template-based RNA structure prediction advanced through a blind code competition. Preprint, bioRxiv 2025.12.30.696949 (version 1, 30 December 2025), not peer reviewed. 10.64898/2025.12.30.696949 · Checked 15 Oct 2026
  14. Kretsch, R. C., Hummer, A. M., He, S., Yuan, R., Zhang, J., Karagianes, T., Cong, Q., Kryshtafovych, A. & Das, R. (2026). Assessment of nucleic acid structure prediction in CASP16. Proteins: Structure, Function, and Bioinformatics 94(1), 192–217 (published online 30 October 2025). 10.1002/prot.70072 · Checked 15 Oct 2026

Page reviewed 15 Oct 2026

Source

Open, as listed by Wikipedia · Unsolved problems in chemistry. Checked 29 Sep 2026.

The status is the source’s. Papex does not decide whether a problem is solved.

Records on this problem

No public record addresses this problem yet.

More in Chemistry