AlphaFold made protein structure look solved — but RNA is the harder sibling: floppier, stabilized by subtle non-canonical pairs, and starved of training data. AlphaFold3 extended structure prediction to nucleic-acid complexes and RhoFold+ built an RNA-specific language model on ~23.7M sequences, yet the CASP15 community verdict is blunt: RNA 3D structure prediction ‘remains an unsolved problem,’ and there deep-learning methods were significantly worse than the top groups that used none. And still — design outruns prediction. LinearDesign optimizes a spike-protein mRNA out of ~2.4×10^632 candidates in 11 minutes and lifted antibody titres up to 128× in mice; RNA language models and 5′UTR models already read and write translation. RNA is a core module of any virtual cell, its bottleneck is data scarcity — the best argument yet for open, shared, callable models — and the prove-it discipline is baked into the field’s own scorecard.
