Protein folding got solved, or close enough that AlphaFold’s developers, Demis Hassabis and John Jumper, shared half of the 2024 Nobel Prize in Chemistry for it. RNA folding did not, and the reason is not that nobody has tried. A Kaggle competition backed by Stanford closed in 2025 with the first automated models ever to match human experts on blind RNA structure predictions, a genuine milestone. The researchers running the follow-up assessment, though, still concluded that deep learning has produced only limited progress on the harder cases. RNA is often treated as protein structure prediction’s smaller sibling, and it is turning out to be a fundamentally harder problem, not just a smaller version of the same one.
Predicting a protein’s 3D shape from its amino acid sequence went from grand challenge to solved-enough-to-productize in under five years once AlphaFold2 arrived, and that same acceleration is what pushed protein language models from a research curiosity into the default way computational biology approaches sequence-to-structure problems. Predicting an RNA molecule’s 3D shape from its nucleotide sequence remains, in the words of the Kaggle competition that has become the field’s proving ground, one of biology’s remaining grand challenges. Understanding why requires looking past the surface similarity, sequences folding into shapes, and at what actually makes RNA’s version of the problem structurally different.
Why This Isn’t Just Smaller AlphaFold
AlphaFold2’s breakthrough leaned on two resources that took decades to build: a large database of experimentally solved protein structures in the Protein Data Bank, and deep multiple sequence alignments, meaning thousands of evolutionarily related protein sequences that reveal which positions co-evolve and therefore likely sit close together in the folded structure. Both resources exist for RNA in a much thinner form. Solved RNA 3D structures in the PDB number in the low thousands, versus roughly two hundred thousand protein structures. Deep alignments are harder to build for RNA because functional RNA sequences tolerate far more variation without losing their fold, which weakens the co-evolution signal that made protein contact prediction work.
The chemistry compounds the data problem. RNA is single-stranded, which means it folds back on itself using a much wider range of interactions than the double-helix logic most people associate with nucleic acids. Beyond standard Watson-Crick base pairing, RNA structures depend heavily on non-Watson-Crick pairs, base triples, and tertiary contacts that fold distant parts of the sequence into direct proximity. Structural biologists Quentin Vicens and Jeffrey Kieft have made the point directly: these non-canonical interactions are chronically underrated in how people think and talk about RNA structure, despite being critical to how any given RNA actually folds into its functional shape. A model trained mostly on the clean logic of base pairing misses a large share of what determines real RNA geometry.
How the Field Actually Tried to Fix the Data Problem
The most direct response to RNA’s data scarcity was not a new model architecture. It was a new data-generation strategy. The Ribonanza project, led by Rhiju Das’s lab at Stanford in collaboration with the citizen-science platform Eterna, ran chemical mapping experiments, DMS and SHAPE probing, across two million diverse RNA sequences, producing nucleotide-resolution measurements sensitive to local structure at a scale nothing in the field had approached before. That dataset became the foundation for a Kaggle competition that solicited and evaluated a wide range of deep learning approaches, then distilled the best ideas from the competition into a single, self-contained model called RibonanzaNet.
RibonanzaNet’s real contribution was not a single benchmark score. It was showing that a model trained primarily on this chemical-mapping data could transfer to a genuinely different downstream task with no RNA 3D coordinate labels involved at all: when fine-tuned for RNA hydrolytic degradation, a property directly relevant to designing more stable mRNA vaccines, the model reached state-of-the-art performance without ever training on that specific problem from scratch. RibonanzaNet-derived secondary-structure predictions were subsequently tested as an input feature for existing tertiary-structure pipelines such as trRosettaRNA, a comparison that let researchers isolate how much a better secondary-structure input actually improves a full 3D prediction downstream.
The Competition That Actually Tested Whether This Works
Chemical mapping data proves a model learned something about local structure. It does not prove the model can predict a full 3D fold, which is the test that actually matters for drug design and RNA biology. That is the gap the Stanford RNA 3D Folding competition on Kaggle was built to close, and its scoring mechanism is worth understanding because it was deliberately designed to resist the kind of shortcut that makes benchmark results misleading.
Competitors submitted code, not predictions, which ran directly on Kaggle’s servers against hidden targets. For every RNA sequence in the test set, a submission had to generate five candidate 3D structures, each scored by TM-score against the true experimental structure, a metric where scores above roughly 0.45 correspond to a genuinely correct global fold rather than a partial or lucky alignment. The final score for each target took the best of the five predictions, and only residues that aligned correctly by sequence numbering counted toward the score, specifically to prevent models from gaming the metric through template reuse or partial structural matches that looked better than they were. The organizers also split the evaluation temporally: training targets came from structures released before a fixed cutoff, and the final blind evaluation used twenty entirely new targets whose structures only became public in the months after the competition’s training phase closed, a design that makes it much harder for a model to have implicitly memorized the answer.
One specific structural motif explains a large share of why RNA 3D prediction resists the kind of secondary-structure shortcuts that work reasonably well for simpler cases. Pseudoknots, where a loop in the RNA sequence pairs with a stretch of bases outside its own stem-loop structure, create crossing base pairs that standard secondary-structure prediction algorithms are mathematically unable to represent using the simple nested-bracket notation that works for ordinary hairpin structures. A model can get the easy, nested part of an RNA’s secondary structure right and still miss the pseudoknot that determines the molecule’s actual 3D fold and function, because the two problems require genuinely different computational approaches rather than a straightforward extension of the same method. Ribozymes, riboswitches, and viral RNA elements rely on pseudoknots specifically because the crossing interaction lets a short sequence encode a more complex, switch-like structural behavior than nested pairing alone could produce, which means the RNA molecules most interesting to structural biologists are disproportionately the ones hardest for current models to get right.
The first run of that competition closed with a genuine milestone. According to Das’s lab, the winning teams, credited under the usernames john, odat, and team Eigen, produced the first fully automated solutions to blind RNA structure prediction that were competitive with human expert modelers, a bar that had not been cleared before by code running without a human in the loop. Part 2 of the competition raised the difficulty deliberately: targets included RNA molecules with no available structural templates at all, RNA complexes bound to small molecules and proteins, and assemblies as large as six thousand nucleotides, developed in collaboration with the AI@HHMI initiative at the Howard Hughes Medical Institute’s Janelia Research Center and experimental structural biology groups. Its submission phase closed in March 2026.
What the Honest Assessment Actually Found
The most useful data point on where the field actually stands did not come from a leaderboard. It came from CASP16, the sixteenth Critical Assessment of Structure Prediction, the community-wide blind benchmark that has run biennially since 1994 and that gave protein folding its most credible progress signal for two decades. A CASP16 special issue assessing roughly sixty groups’ blind RNA structure predictions, led by Rachael Kretsch at Stanford, reached a conclusion that is a useful corrective to any headline implying RNA structure prediction is now solved: deep learning has fueled limited progress in modeling RNA and its complexes, and much progress is still required. That assessment also did something CASP had not done before, evaluating the water molecule ensembles surrounding an RNA structure, a level of physical detail that matters for how these molecules actually behave in solution and that current models are not yet being held to consistently.
A December 2025 analysis of the Stanford Kaggle competition’s blind evaluation phase adds a specific, structural reason the gains have been limited. The paper found that the strongest-performing submissions leaned heavily on template-based approaches, essentially finding and adapting structurally similar RNA sequences with known folds, rather than learning the underlying folding physics the way protein models increasingly do. That is a meaningfully different kind of progress than what happened in protein structure prediction, where deep learning genuinely displaced template-based homology modeling as the dominant approach. RNA structure prediction, on this evidence, has not made that same leap yet, and the distinction matters because template-based methods have a hard ceiling: they can only ever be as good as the library of previously solved structures they are searching against, which means performance on genuinely novel RNA folds, the ones with no close structural relative already solved, tells you almost nothing from a template method’s score on familiar targets.
The evaluation methodology itself is still evolving in ways that reveal how young this field’s assessment culture is compared to protein structure prediction’s. CASP has spent three decades refining exactly how to score a predicted protein structure fairly, developing metrics like GDT-TS and lDDT that the community broadly trusts. RNA assessment is still catching up: the CASP16 special issue’s decision to evaluate water molecule ensembles around predicted RNA structures for the first time reflects a field actively deciding what “correct” should even mean for a folded RNA, not just how well models hit an agreed target.
Why the Difficulty Isn’t Just an Academic Curiosity
The gap matters commercially, not only scientifically. RNA-based medicine, mRNA vaccines, RNA aptamer therapeutics, CRISPR guide RNA design, and an expanding class of RNA-targeted small-molecule drugs all depend on knowing how a candidate RNA sequence actually folds, because structure determines function, stability, and how a molecule interacts with its targets. Protein-targeted drug discovery got a genuine AI-driven acceleration once AlphaFold made structure prediction fast and cheap, the same acceleration that raised real dual-use questions about AI-designed biology once structure prediction got good enough to work in reverse, generating novel sequences rather than just explaining known ones. RNA-targeted drug discovery is still waiting for the equivalent unlock, and the data above suggests that unlock is a harder engineering and data problem than a straightforward architecture copy from the protein side would suggest.
There is also a basic biology angle competition organizers keep returning to. Vast stretches of the genome transcribe into RNA that never gets translated into protein at all, non-coding RNA whose function is determined almost entirely by the 3D shape it folds into. Structure prediction is often the only practical way to guess what a newly discovered non-coding RNA actually does, which is why Das’s lab frames the competition’s goal explicitly around illuminating the functions of natural RNA molecules, not just building a better benchmark score.
What Happens Next
Part 2 of the Stanford RNA 3D Folding competition closed its submission phase in March 2026, having raised the bar specifically to push past template reuse: its hardest targets were RNA sequences with no close structural relative in any existing database, exactly the case where template-based approaches have nothing to copy from. Whether the models that competed there actually learned generalizable folding physics, rather than a more sophisticated form of pattern matching against known folds, is a question about the results that this piece has not independently verified, and is worth checking against the competition’s own published solution write-ups directly. The organizers have also said they intend to return to CASP as predictors rather than assessors, putting their own methods directly into the same blind evaluation that just delivered the sober CASP16 verdict.
The realistic read on RNA structure prediction going into the second half of 2026 is that it is a genuinely active, well-funded, well-organized research problem with real institutional backing, not a solved one wearing AlphaFold’s old clothes. The data infrastructure that took protein folding from hard to solved, in RNA’s case a combination of Ribonanza-scale chemical mapping data, Kaggle-hosted blind competitions, and CASP-grade independent assessment, is now in place. What is missing is the equivalent of AlphaFold2’s actual architectural breakthrough, and nothing in the current evidence says that arrives on the same timeline the protein story set as a precedent.
Leave a Reply