Loading…
Diffusion-based RNA inverse folding, denoising toward a nucleotide sequence conditioned on a target 3D backbone for higher native sequence recovery.
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 bottlenecked by the scarcity of high-resolution 3D RNA structures, which leaves data-hungry models prone to overfitting on a small pool of experimental examples.
SCRU-Diff, from Jian Wang and Nikolay V. Dokholyan at the University of Virginia School of Medicine, is the high-fidelity half of the pair proposed in their 2026 bioRxiv preprint. It is a generative diffusion model that conditions on a target 3D backbone and denoises toward a nucleotide sequence over multiple steps. Its companion, SCRU-Seq, answers the same question in one forward pass; this model spends the extra compute and recovers accuracy for it, reaching 79.2% best native sequence recovery on set112 against the single-pass model's 63.7%.
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, and the authors credit it, rather than architecture scale, for the gains either model shows.
The model conditions on a target 3D RNA backbone and denoises toward a nucleotide sequence over a series of steps, recovering accuracy at the cost of compute proportional to the step count. This is the standard trade against non-autoregressive designers, and here it is quantified directly: on the curated, high-fidelity set112 benchmark the model attains a best native sequence recovery of 79.2%, against 63.7% for the single-pass graph network trained on the same corpus, with both compared to established RNA inverse-folding systems including NA-MPNN and RiboDiffusion.
Training runs on the SCRU-DB corpus of 61,000-plus self-contained units across 8,200-plus structural clusters. Because SCRUs are structurally autonomous, each contributes an independent training example rather than a fragment whose folding depends on the rest of its parent molecule — the mechanism the authors identify as making a diffusion model trainable at all on a dataset this size.
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. Its place in a workflow is wherever structural fidelity matters more than runtime: designing a small number of constructs intended for synthesis and wet-lab testing, or refining the most promising hits from a bulk screen that SCRU-Seq generated. Both models expose the same interface over the same framework, so moving a design between them costs nothing.
The paired release makes an explicit, measurable speed-accuracy curve available within one framework, which is more useful to a practitioner than either model alone would be — the 15.5-point recovery gap is the quantity a designer actually needs to decide how much compute a project can justify. 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. SCRU-DB, if released, would likely outlast either model as a contribution.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.