Suppose you want to change an RNA sequence so that a cell makes more of a protein. A model that predicts which RNA bases pair together may help you explore structure. To choose useful changes, you also need evidence that its predictions relate to protein production in your experimental system.
This is why choosing an RNA model starts with a biological question. Predicting a splice site, a base-pair map, a three-dimensional structure and protein output are different tasks.
Many RNA models first learn patterns from large collections of sequences. They turn a sequence into a numerical description that another predictor can use. If that predictor is trained using examples with known experimental outcomes, it still relies on labelled data, even when the original sequence model is left unchanged.
For translation, ask exactly what was measured. Ribosome load, the average number of ribosomes associated with an RNA, and measured protein output are related, but they are different readouts. The RNA construct, cell type and assay all matter.
The UTR-LM study provides a concrete example. Its authors trained a model on 5′ untranslated regions, the regions before the protein-coding sequence that help regulate translation. They adapted it to specific prediction tasks and tested newly designed sequences in the laboratory. Those experiments support conclusions about the conditions tested. Using the predictions in another setting requires further validation.
Structure predictions need similar care. A base-pair map shows which nucleotides pair; it does not give the positions of every atom. A single three-dimensional structure also does not capture all the shapes an RNA may adopt. Some systems use related sequences or previously solved structures as additional information, so these inputs should be reported with the model version.
The CASP16 nucleic-acid assessment illustrates why this matters. Researchers submitted predictions before the experimental structures were released. This blind assessment showed that the availability of related structures strongly affected prediction accuracy.
When evaluating designed sequences, count the whole experiment: how many candidates were generated, selected, synthesised, tested and successful? Include failures as well as the best example. A promising computer-generated sequence needs an appropriate experiment to establish the claimed improvement.
Before choosing a model, write down six things:
-
The biological object: a mature transcript, an untranslated region or another RNA?
-
The available inputs: sequence alone, related sequences or known structures?
-
The required measurement: base pairing, protein output or another defined result?
-
The training needed: can the model be used directly, or will it need labelled examples from your task?
-
The evaluation: were test examples kept separate from training, including closely related sequences where relevant?
-
The evidence needed: what comparison or laboratory result would justify your next decision?
These questions make model selection easier to discuss with collaborators and reveal where a simpler method or another experiment may be more useful.
Disclosure: I write rewire.it. This contribution is adapted from my longer RNA foundation-model guide, which links the primary studies and explains their evaluation methods.