An open RNA model has learned to read shape, not just sequence

Most coverage of a new biological foundation model begins with parameter counts and benchmark tables. The structRFM paper, published in Nature on 5 November by Shaohua Zhou and Peng Xiong's group at the University of Science and Technology of China, deserves a different opening: it changes what the model is asked to learn. Instead of treating an RNA molecule as a string of four letters, the pretraining objective forces the network to account for the shape that string folds into. That is a different claim about the object being modelled, and it is why the release matters beyond one more leaderboard entry.

RNA is routinely introduced as a messenger between genetic information and protein, which is a convenient simplification and a misleading one. The molecule participates in protein synthesis, gene expression control and splicing, and its behaviour depends on both its nucleotide order and the secondary and tertiary structures it adopts. An identical stretch of bases can produce different outcomes in different structural states. A language-model-style encoder trained only on A, U, C and G therefore starts with half the relevant information missing, and no amount of extra sequence data can fully compensate for a variable that was never shown to it.

The field's default answer has been to borrow the protein playbook: self-supervised pretraining on large sequence collections, then a separate fine-tune for every downstream task. Structure prediction, meanwhile, grew up as its own toolchain with its own assumptions and formats. The result is a fragmented workflow in which each new question requires a freshly engineered feature set, and tasks with small labelled datasets suffer most because they cannot reuse what the structure tools already know. Anyone who has tried to move between splice-site annotation and structure inference in a single project will recognise the friction.

structRFM begins with a dataset rather than an architecture. The team assembled roughly 21 million computationally annotated sequence–secondary-structure pairs, binding the question of which bases are present to the question of which bases pair with which. Computationally annotated is an important qualifier, and we will return to it; the immediate point is that structural information enters pretraining as a training signal instead of being bolted on afterwards. Pairing at this scale also makes the supervision dense: every position in a sequence carries a structural relationship, not just a single label at the end.

The objective is consequently two-sided, learning sequence context and structural dependency together, with a masking scheme designed to make both necessary. The paper's ablation work separates the contributions of the structural branch and the masking strategy, and finds that each matters. That detail is easy to skim past, yet it answers the most useful question a follow-up group can ask: which component actually produced the gain? A result that improves simply because the corpus got bigger is far less actionable than one whose mechanism has been isolated.

From a single input sequence the trained model emits representations at three granularities: whole-sequence, per-nucleotide, and nucleotide-pair. This is a deliberately practical design. Secondary structure is fundamentally a pairing problem, splice-site detection behaves like local labelling, and homology classification needs a global view. Covering all three with one encoder means downstream projects can start from a representation instead of building a bespoke encoder for each task, which is where most of the engineering cost in computational biology tends to sit.

The evaluation list is correspondingly varied: secondary structure prediction, zero-shot homology classification, splice-site prediction, internal ribosome entry site recognition, and non-coding RNA classification. The variety is the argument. If the model has genuinely internalised structural priors, these loosely related tasks should improve together rather than one metric being tuned at the expense of the rest. A model that only wins on the benchmark its authors cared about tells you very little about whether the representation generalises.

Two further results are worth more attention than the headline numbers. Feeding structRFM features into an existing three-dimensional prediction pipeline improved both the 3D output and its associated secondary structure, which is physically sensible: two-dimensional pairing constraints narrow the space of plausible folds. And the model held up on newly released RNA families, with a segmentation strategy extending it to long non-coding RNAs. Biological databases acquire new families continuously, so robustness to families unseen during training is closer to real utility than a marginal gain on a fixed test set.

The release includes the full paired dataset, model weights and source code. For computational biology the data is often the harder artifact to reproduce, because annotation pipelines carry parameters, versions and filtering rules that silently change results; publishing weights without them can leave a paper technically open but practically unrepeatable. Combined with the decision to share the paired data, the open release gives other groups a real chance to test where the approach breaks — including the gap between computational annotation and experimental measurement, and the question of how the model behaves on folds with no detectable homologues. That is a starting point for the community, not a finished answer.