RNA Structure Prediction is Hard. How Much Does That Matter? Author: Abhishaike Mahajan Date: Sep 26, 2025 Reading time: 22 minutes (~4.8k words) --- Introduction The author set out to write a primer on RNA structure modeling, initially thinking Alphafold3 might have solved the problem. A conversation with Connor Stephens from Atomic AI revealed that RNA structure prediction remains very challenging and might be among the hardest problems. Alphafold3 improves some aspects but does not solve RNA prediction outright. --- Why is RNA Structure So Hard to Model? Experimental RNA 3D structures are scarce and often low-quality compared to protein structures (7,759 RNA structures vs. 216,212 protein structures in PDB as of mid-2024). RNA molecules have many more conformational degrees of freedom than proteins because of flexible backbones and lack of bulky side chains. RNA is not unstructured but exists in multiple preferred conformations, flipping between them frequently. Traditional experimental methods face obstacles: X-ray crystallography: RNA flexibility makes crystallization hard. Cryo-EM: Challenging for small, flexible RNAs with heterogeneity. NMR: Limited to very small RNAs (<50 nucleotides). Two main RNA structure types can be physically characterized: Artificially stabilized RNAs (e.g., metal ions, modifications). Evolutionarily constrained RNAs with dominant stable conformations (less common). Alphafold3 performs variably: Better than some ab initio methods in physical plausibility. Does not always have the best atomic alignment. Tailored specialized methods outperform it on certain RNA classes. Predictions are more accurate for RNA structures similar to those in training datasets; unseen or divergent RNAs remain challenging. --- Why Even Predict RNA Structure? Proteins: 3D structure closely tied to function; tertiary structure is crucial. RNA: The situation is different; the importance splits between secondary and tertiary structures. Secondary structure (base pairing and stems) often dictates biological activity. Tertiary structure is more flexible and less consistently critical therapeutically. RNA secondary structure is easier to predict than tertiary structure and is often sufficient. Therapeutic Contexts: mRNA therapies: Secondary structure matters (prevent hairpins), but tertiary structure less so as mRNA's role is to be translated. Antisense oligonucleotides (ASOs) and siRNAs: Target binding depends on base-pairing accessibility at the secondary structure level. Aptamers and ribozymes: Rare approved drugs; tertiary structure important here as they form precise 3D shapes to bind molecules or catalyze reactions. RNA as targets: microRNAs: small, tertiary structure less relevant. long non-coding RNAs (lncRNAs): debated if they have stable tertiary structures; modulation depends on short motifs rather than global folding. Ribosomal RNA (rRNA): Tertiary structure crucial for enzymatic activity (e.g., peptidyl transferase center) and antibiotic targeting. --- Potential Gains from Improved RNA Structure Prediction Immediate benefits likely in antibiotics targeting rRNA and RNA aptamers. Antibiotic development hindered more by economic models than scientific challenges. Viral RNA targeting is promising but still early-stage. Aptamers have advantages over antibodies: easier production, stability, lower immunogenicity, but face competition from other modalities. Circular RNAs (circRNAs) depend on tertiary structure for translation initiation (via IRES elements), making tertiary RNA prediction important for this emerging therapeutic class. mRNA vaccines use modified nucleotides (e.g., m1Ψ) which alter secondary structure predictions, complicating modeling efforts. Overall, some areas benefit strongly from tertiary structure