Loading…
Inverse RNA folding from contact maps: an axial-attention transformer designing sequences for pseudoknots, non-canonical pairs and multiplets.
No providers recorded yet. Browse all providers
Dot-bracket notation cannot write down a pseudoknot without borrowing a second bracket type, and it has no way at all to write a nucleotide that pairs with two partners. Both are ordinary in solved RNA: across the PDB-derived test sets used here, 67–95% of structures contain a pseudoknot and 74–95% a base multiplet. Structure predictors emit adjacency matrices and represent these interactions fine; design methods read dot-bracket strings, so what the predictors find cannot be handed back as a design target.
RNAinformer removes that mismatch by making the contact matrix the input. Built in the Department of Computer Science at the University of Freiburg, with Frank Hutter also at the ELLIS Institute Tübingen, it is an encoder–decoder transformer that embeds the target as a binary L×L adjacency matrix, refines it with axial attention in a two-dimensional latent space, pools that to a one-dimensional encoding, and generates a sequence autoregressively conditioned on it. A matrix restricts neither which nucleotides pair nor how many partners each takes, so nested helices, pseudoknots, non-canonical pairs and multiplets are all entries in one array.
The design is the mirror image of RNAformer, the same group's structure predictor, which reads a sequence and returns a contact map. gRNAde inverse-folds from a 3D backbone and GenerRNA samples with no structural target; RNAinformer works at the secondary-structure level and inverts the algorithms that dominate it.
Six encoder and six decoder blocks, model dimension 256, four attention heads; training runs 50,000 steps at batch size 128 on two A40 GPUs, about 18 hours per model, with a separate model per dataset and conditioning mode and a 200-nucleotide length cap. Training pools come from Rfam 14.10: 1,000 sequences sampled per family with Infernal, split on clans. Three synthetic sets of roughly 444,000 samples each were folded differently — SynNested with RNAfold, SynPseudoknot with HotKnots 2.0, SynMultiplet with RNAformer (45% pseudoknots, 57% multiplets, 64% non-canonical pairs).
On SynNested, RNAinformer solves 91.8% of inverse-folding tasks from 20 candidates, behind the search method SAMFEO at 99.6% and ahead of libLEARNA at 77.2%; with a target GC content it solves 69.6% against libLEARNA's 59.0%. On SynPseudoknot it solves 68.5% versus antaRNA's 15.6%, and 33.7% versus 1.2% under GC conditioning. Solved rates fall sharply on the multiplet-rich PDB test sets, but the designs fold back to the target with higher F1 than the deposited sequences themselves — a gain of 0.12 to 0.19 that holds when they are refolded with SPOT-RNA, MXfold2 and UFold. Training on synthetic data alone beat fine-tuning on real RNA data by nearly double the F1. A later manuscript in the repository reports different headline numbers, among them 39.1% of pseudoknotted targets solved with one candidate against antaRNA's 15.6%.
The riboswitch experiment shows the intended shape of use: fix the theophylline aptamer and the 8-uracil stretch, mask the spacer and complementary regions, set a target GC content, and let the model fill the gaps — a formulation on which it solves over 90% of tasks. The same interface serves synthetic biologists designing aptamers, ribozymes or scaffolds around a preserved motif, and groups building RNA biosensors or therapeutics, who can pass a predictor's contact map into a design run without translating through dot-bracket notation.
RNAinformer's contribution is representational: moving the design interface from strings to matrices makes structure predictors' outputs usable as design inputs, and it is the first inverse-folding method able to target base multiplets at all. A model trained purely on synthetic Rfam folds also beat one fine-tuned on experimental data, arguing that homology-controlled synthetic data can substitute for scarce real structures here. The limits are real: the 200-nucleotide cap excludes most long non-coding RNAs, the two-dimensional latent is memory-intensive, solved rates on experimental structures with mixed interaction types stay low, and the work remains an unreviewed preprint.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.