RNA inverse folding model that designs nucleotide sequences for a target 3D backbone by running discrete diffusion in hyperbolic space.
No providers recorded yet. Browse all providers
Designing a nucleotide sequence that folds into a prescribed three-dimensional shape — RNA inverse folding — is the generative counterpart of structure prediction and the route to purpose-built aptamers, ribozymes, and riboswitches. It is harder than the protein equivalent for one dominant reason: experimentally determined RNA tertiary structures are scarce. A model restricted to experimental data works from a corpus of roughly two thousand chains, so learning efficiency, rather than raw scale, is the binding constraint.
RIdiffusion, developed at the School of Pharmacy of Fudan University with Byterna Therapeutics and East China Normal University, addresses that constraint by changing the geometry the model reasons in. Instead of embedding the RNA backbone graph in Euclidean space, it maps the graph into a Poincaré ball, a model of hyperbolic space in which distances between nodes grow the way they do in a tree. RNA structure is hierarchically organized — base pairs assemble into helices and loops, which assemble into junctions and pseudoknots, which assemble into a global fold — and a hyperbolic embedding captures that hierarchy in fewer dimensions while separating similar backbones more sharply than a Euclidean one.
Generation itself is framed as discrete diffusion. Uniform noise is progressively added to the nucleotide labels of a real RNA until the sequence distribution is pure noise; generation reverses the process, starting from a uniformly sampled sequence and iteratively denoising it conditioned on the fixed backbone. RIdiffusion is a distinct lineage from gRNAde, which approaches the same task with geometric message passing rather than hyperbolic diffusion.
seq_generator.py entry point that redesigns sequences for a directory of new PDB backbones without retraining.Training used experimentally derived structures from the PDB and the RNAsolo database, split 8:1:1 into 1,773 training, 221 validation, and 223 test structures following the protocol established by RDesign, then reclustered by PSI-CD-HIT into three similarity-capped subsets. Across those subsets RIdiffusion reached sequence recovery rates of 43.37%, 43.97%, and 45.80% with Macro-F1 scores of 43.24, 43.33, and 45.71, the top figure in every column; baselines were gRNAde, GVP-GNN, GraphTrans, PiFold, and StructGNN. At a strict 80% sequence-similarity cutoff the best of those, StructGNN, reached 36.51% recovery. Splitting the same subset by length, RIdiffusion recovered 45.43% on medium sequences of 50–100 nucleotides while training on only 225 examples, against 43.37% on the full 761-sample subset, indicating that performance does not collapse as training data shrinks. Ablating the hyperbolic embedding and the hyperbolic graph transform together dropped recovery from 45.80% to 43.90%. Folding the designed sequences back with the RhoFold+ structure predictor and aligning with US-align produced samples within 2 Å RMSD of the input for several targets.
RIdiffusion is intended for structure-guided RNA engineering: given a backbone from an experimental structure, a predicted model, or a design pipeline, it proposes sequences expected to adopt that fold. That covers aptamer and ribozyme redesign, engineering of structured regulatory elements for synthetic biology, and scaffold-constrained diversification where a functional motif is preserved while the surrounding sequence varies. Because the released checkpoints run on arbitrary new backbones through a single inference script, the model also serves as a generative baseline for groups benchmarking new RNA design methods.
RIdiffusion was peer reviewed and published in the Journal of Chemical Information and Modeling in June 2025, though its evaluation is entirely computational — recovery rates, novelty scores, and in-silico refolding, with no reported wet-lab validation of designed sequences. Its contribution is methodological: it demonstrates that the choice of embedding geometry, not just model capacity, materially affects generative performance when structural training data are limited, a finding that generalizes beyond RNA to any structural domain where hierarchical topology and small datasets coincide. The code and pretrained weights are released under an MIT license, making it one of the more directly reusable entries in the small field of RNA 3D inverse folding; the article itself is distributed under standard ACS terms rather than an open licence, so the benchmark tables cannot be redistributed freely.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.