Loading…
Single-pass RNA inverse folding: a graph neural network predicts a nucleotide sequence from a target 3D backbone in constant time.
No providers recorded yet. Browse all providers
Designing an RNA sequence that will fold into a target three-dimensional shape — the "inverse folding" problem — is a central challenge in synthetic biology and RNA therapeutics. Progress has been bottlenecked twice over: by the scarcity of high-resolution 3D RNA structures, which leaves data-hungry models prone to overfitting, and by the computational cost of leading methods that rely on autoregressive or iterative sampling over whole molecules.
SCRU-Seq, from Jian Wang and Nikolay V. Dokholyan at the University of Virginia School of Medicine, is the fast half of the answer proposed in their 2026 bioRxiv preprint. It is a graph neural network that reads a target 3D RNA backbone and emits the underlying nucleotide sequence in a single non-autoregressive forward pass, giving constant-time inference relative to iterative competitors. Its companion, SCRU-Diff, spends more compute per design to recover accuracy; between them a user picks a point on the speed-fidelity curve without changing frameworks or retraining.
Both rest on the same data idea. The work decomposes complex RNAs into Self-Contained RNA Units (SCRUs) — structurally autonomous modules identified through tertiary-contact clustering, each behaving as a self-stabilizing, foldable physical unit — and assembles them into SCRU-DB, a library of more than 61,000 units spanning over 8,200 unique structural clusters. That decomposition is what turns a limited pool of experimental structures into a much larger effective training signal.
The network consumes a target 3D RNA backbone as a graph and predicts the nucleotide identity at every position simultaneously. Because nothing is conditioned on previously emitted nucleotides, inference is O(1) in the number of sampling steps — the property that separates it from RiboDiffusion and from autoregressive designers, and the reason it is the member of the pair suited to bulk design.
Training runs on the SCRU-DB corpus of 61,000-plus self-contained units across 8,200-plus clusters, and evaluation is against established RNA inverse-folding systems including NA-MPNN and RiboDiffusion. On the curated, high-fidelity set112 benchmark the model reaches 63.7% native sequence recovery. That trails the 79.2% Best NSR its diffusion counterpart attains on the same benchmark, which is the trade the two-model design makes explicit: the gap is the price of the single pass. The authors credit the modular SCRU representation, rather than architecture scale, for the gains either model shows under limited 3D-structure data.
The model targets RNA nanotechnology, synthetic biology, and RNA-based therapeutics, where the task is engineering sequences that fold into specified 3D conformations — structured aptamers, ribozymes, riboswitches, or scaffolds for RNA drug development. Single-pass speed is what makes it the right choice for high-throughput screening and large design libraries, where thousands of candidate backbones must be processed and per-design fidelity can be recovered later by rescoring or by rerunning selected hits through SCRU-Diff.
Reframing RNA inverse folding around reusable self-contained units offers a route past the field's defining obstacle, the scarcity of experimental 3D RNA structures, and SCRU-DB is arguably a contribution in its own right that could support RNA modeling well beyond sequence design. The caveats are substantial and worth stating plainly: as of its 2026 preprint the work has not been peer reviewed, and no public code, weights, or database release has been located, so neither the reported recovery numbers nor the claimed plug-in usability can currently be exercised by anyone outside the lab.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.