RNA Function Follows Form — But RNA Refuses to Sit Still
At a virtual conference in November 2020, structural biology changed.
The winner of the CASP14 protein-structure-prediction challenge was announced: AlphaFold, developed by Google DeepMind. The result was not merely better than previous tools. It was dramatically better. AlphaFold showed that artificial intelligence could predict many protein structures with near-experimental accuracy, solving a problem researchers had been chasing for decades.
Protein biology had its revolution.
RNA biology is still waiting for its equivalent moment.
That is the central tension behind Diana Kwon’s Nature Technology Feature, “RNA function follows form – why is it so hard to predict?” The article captures a frustrating truth: RNA is biologically essential, structurally fascinating, and increasingly important for medicine — but predicting its shape remains far harder than many people expected.
RNA Is Not Just a Messenger
For decades, RNA was introduced in textbooks as a middleman: DNA stores genetic information, RNA carries the message, and proteins do the real work.
That explanation is now painfully incomplete.
RNA can regulate genes, catalyze reactions, guide protein complexes, control splicing, sense metabolites, organize cellular machinery, and influence disease. Ribozymes, riboswitches, long noncoding RNAs, microRNAs, guide RNAs, viral RNAs, circular RNAs, and therapeutic RNAs all remind us that RNA is not passive.
RNA does things.
But RNA does those things because it folds.
Its biological function depends on stems, loops, bulges, pseudoknots, junctions, long-range contacts, base stacking, ion coordination, and interactions with proteins or small molecules. In RNA biology, structure is not decoration. Structure is often the mechanism.
That is why the phrase “function follows form” matters.
An RNA molecule’s sequence tells us what it could become. Its structure tells us what it is actually capable of doing.
Why Proteins Were Easier for AI
AlphaFold’s success in protein structure prediction was built on several advantages.
Proteins have been studied structurally for a long time. Thousands of high-quality protein structures were already available in the Protein Data Bank. Protein sequences also carry rich evolutionary information: if two residues change together across evolution, they may physically interact in the folded protein. AlphaFold and related tools learned from these patterns at massive scale.
RNA does not offer the same easy path.
There are far fewer experimentally solved RNA structures than protein structures. Many RNAs are small, flexible, chemically sensitive, and structurally heterogeneous. RNA often depends on magnesium ions, cellular proteins, modifications, ligand binding, and environmental conditions to fold correctly. A sequence may not point to one stable structure. It may point to a shifting ensemble.
That makes RNA a much harder target for machine learning.
A protein often behaves like a molecule trying to reach a stable folded state.
RNA often behaves like a molecule negotiating among several states.
RNA’s Flexibility Is the Problem — and the Biology
The biggest mistake is to think RNA structure prediction is simply protein structure prediction with different building blocks.
RNA has its own grammar.
It has only four standard bases, but those bases can form canonical and noncanonical interactions. Its phosphate backbone is highly charged. Its folding can depend strongly on ions. It can form alternative secondary structures. It can switch conformations when binding a metabolite or protein. It can expose or hide regulatory regions depending on context.
This is not just a computational nuisance.
RNA flexibility is often exactly how RNA works.
A riboswitch must change shape to regulate gene expression. A viral RNA may remodel itself during infection. A guide RNA must fold into a form compatible with its protein partner. An mRNA may contain structural elements that affect translation, degradation, or immune recognition.
So the goal is not always to predict one “correct” RNA structure.
The goal may be to predict a population of possible structures, then understand which one matters under a specific biological condition.
That is much harder.
AlphaFold 3 Helps, But It Does Not End the RNA Problem
AlphaFold 3 expanded the modeling landscape by predicting biomolecular complexes involving proteins, DNA, RNA, small molecules, ions, and modified residues. That is a major advance because biology rarely happens molecule by molecule in isolation. Cells are crowded with interacting systems.
But RNA structure prediction is still not solved.
AlphaFold 3 can model RNA-containing complexes, but RNA-only folding and RNA conformational ensembles remain difficult. Many RNA structures depend on experimental constraints, secondary structure priors, or specialized RNA-focused modeling approaches. The RNA field is therefore not simply waiting for one universal model to solve everything.
It is building a different toolkit.
The New RNA AI Toolkit
Several AI-based methods are now pushing RNA structure prediction forward.
Tools such as RhoFold+ use RNA language models and deep learning to predict RNA 3D structures from sequence. RhoFold+ was trained with large-scale RNA sequence information and designed to address the data scarcity that limits RNA modeling.
Other methods, including trRosettaRNA and trRosettaRNA2, use RNA-specific structural logic, secondary structure information, and deep learning to improve 3D prediction. Recent work on trRosettaRNA2 emphasizes the value of secondary-structure-aware modeling and conformer prediction — a crucial feature because RNA often exists in multiple structural states rather than one fixed architecture.
These methods suggest an important principle:
RNA prediction will probably not be solved by sequence alone.
It will require secondary structure priors, evolutionary signals, experimental probing data, cryo-EM maps, chemical constraints, molecular simulations, and AI models working together.
Experiments Still Matter
The rise of AI does not make RNA experiments obsolete.
It makes them more important.
Chemical probing methods such as SHAPE and DMS-based approaches can reveal which nucleotides are flexible, paired, exposed, or protected. Cryo-electron microscopy can capture larger RNA-containing complexes. NMR can reveal dynamics and local structure. X-ray crystallography can still provide atomic detail when crystals are available.
Each method sees a different part of the RNA story.
AI can propose models quickly. Experiments can test whether those models are real.
This is where RNA structure biology is heading: not toward purely computational prediction, but toward integrative structure determination.
A model becomes more trustworthy when it agrees with probing data, mutational analysis, cryo-EM density, biochemical function, and evolutionary conservation.
For RNA, evidence must converge.
Why This Matters for Medicine
RNA structure prediction is not just a technical problem for structural biologists. It is becoming central to biotechnology and medicine.
mRNA vaccines, siRNA drugs, antisense oligonucleotides, CRISPR guide RNAs, aptamers, ribozymes, circular RNAs, and self-amplifying RNAs all depend on folding behavior. A therapeutic RNA may fail because it folds incorrectly, exposes the wrong region, activates unwanted immune sensors, degrades too quickly, or binds inefficiently.
RNA-targeted small-molecule drugs are another major frontier. For a drug to bind RNA selectively, the RNA must present a recognizable structural pocket or motif. Without structural knowledge, RNA drug discovery becomes guesswork.
Better RNA structure prediction could help researchers design more stable RNAs, improve guide RNA performance, identify druggable RNA motifs, engineer riboswitches, and understand viral RNA elements.
In short, RNA structure prediction is not only about seeing molecules.
It is about designing them.
The Real Lesson from AlphaFold
AlphaFold changed protein biology because it made high-quality structural models widely accessible. It did not eliminate experiments, but it changed where experiments begin.
RNA needs a similar shift.
But the RNA version of that revolution may look different. It may not be a single model that predicts one final structure from sequence. It may be a network of tools that predicts secondary structures, tertiary folds, alternative conformers, RNA–protein complexes, RNA–ligand interactions, and experimentally testable structural hypotheses.
RNA does not sit still.
So RNA structure prediction must become comfortable with motion.
The next breakthrough will not simply tell us, “Here is the RNA structure.”
It will tell us:
Here are the structures this RNA can adopt. Here is when they appear. Here is what they bind. Here is how they regulate biology. Here is how we can redesign them.
That is the future RNA biology is moving toward.
Protein structure prediction had its AlphaFold moment.
RNA’s moment may be harder, slower, and messier.
But it may also be more interesting — because RNA is not merely a molecule with a shape.
It is a molecule with possibilities.
References / Sources
Kwon, D. “RNA function follows form – why is it so hard to predict?” Nature, 2025. (Nature)
Jumper, J. et al. “Highly accurate protein structure prediction with AlphaFold.” Nature, 2021. (Nature)
Abramson, J. et al. “Accurate structure prediction of biomolecular interactions with AlphaFold 3.” Nature, 2024. (Nature)
Shen, T. et al. “Accurate RNA 3D structure prediction using a language model-based deep learning approach.” Nature Methods, 2024. (Nature)
Wang, W. et al. “The trRosettaRNA server for RNA structure prediction.” Nature Protocols, 2026. (Yang Lab)
CASP16 RNA structure prediction assessment and Yang-Server/trRosettaRNA2 reporting. (Wiley Online Library)
No comments:
Post a Comment