RNA conformational ensemble generation with a diffusion model, sampling excited states and folding pathways from one structure without MSA input.
No providers recorded yet. Browse all providers
Most non-coding RNA does its job by changing shape. Riboswitches toggle between conformers to gate transcription, and regulatory elements populate short-lived excited states that recruit different partners than their ground state does. A single predicted structure — the output of RhoFold+ and other RNA structure predictors — therefore misses the functional picture. The alternative, molecular dynamics, is expensive enough that reaching transitions on biologically relevant timescales is often out of reach, and the RNA force fields available carry known pathologies: the widely used OL3 parameters trap even simple tetranucleotides in non-native intercalated states.
DynaRNA, from Shanghai Jiao Tong University, applies the generative-ensemble approach that AlphaFlow and BioEmu-1 established for proteins to RNA. A denoising diffusion probabilistic model paired with an equivariant graph neural network operates directly on 3D coordinates, sampling ensembles in minutes to hours on a single GPU where the equivalent simulation takes weeks. It uses no multiple sequence alignment, which matters for RNA more than for protein: homologous RNA families are sparse and alignments unreliable, so MSA dependence is a real constraint on the alternatives.
The distinguishing mechanism is partial noising. Rather than starting from pure noise, DynaRNA corrupts an input structure only to an intermediate step of the diffusion schedule and then denoises, which keeps generated conformers within a physically sensible neighbourhood while still allowing genuine transitions. The consequence is bidirectional sampling: seeded with a ground-state structure the model reaches the excited state, and seeded with the excited state it returns to the ground state.
Each nucleotide is coarse-grained to one particle at its C4′ atom — a minimal encoding of backbone geometry that keeps the graph small enough for long ensembles — and generated conformers are rebuilt to all-atom detail with Arena. The EGNN uses a hidden dimension of 128 throughout with LayerNorm, SiLU activations, and 0.1 dropout; node features are C4′ coordinates plus timestep embeddings, edges carry connectivity and spatial distance. Training drew 14,632 RNA structures from RNAsolo, curated to 6,820 structures of 5–200 nucleotides after removing non-RNA components, modified bases, and incomplete nucleotides, and ran about 14 days on a single NVIDIA 4090D with Adam at learning rate 1e-4 against an L2 noise-prediction loss. The partial-noising level was tuned across 200 to 1024 steps, with 800 selected.
Reconstructed all-atom structures reproduce experimental covalent geometry, with a C5′–C4′ bond length of 1.509 Å against 1.478 Å in the PDB and a mean absolute error of 0.031 Å across bond lengths. On HIV-1 TAR, clustering recovers 11% ES2 population from a ground-state seed and 16% ground state from an ES2 seed. De novo folding of tetraloops with no prior structural knowledge reaches a minimum RMSD of 0.9 Å and eRMSD of 1.01 for UUCG, and 1.3 Å with eRMSD 1.13 for GAAA, recovering all Watson-Crick base pairs; native-state clusters account for 20.6% and 18% of the respective ensembles.
The model suits structural biologists characterising RNA dynamics where NMR relaxation dispersion or chemical probing has indicated a hidden state but not resolved it, and RNA therapeutics groups screening for transient pockets that a static structure would not expose. Because inference needs only a starting PDB file, it slots in downstream of any structure predictor or experimental structure as an ensemble-expansion step, and its cost profile makes it usable as a hypothesis generator ahead of targeted simulation rather than a replacement for it.
DynaRNA extends generative conformational ensembles from proteins to RNA and shows the approach can recover excited states that both experiment and force-field simulation struggle to characterise. The evidence base is currently narrow — tetranucleotides plus two case studies, with larger structured RNAs such as riboswitches and ribozymes not yet demonstrated — and the coarse-grained C4′ representation means base-pairing and stacking detail come from the Arena reconstruction rather than the generative model. The Zenodo checkpoint is released under CC BY 4.0, but the code repository carries no license file at all, and both the journal article and the preprint are distributed under non-commercial terms.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.