RNAbpFlow

RNA structure prediction has not followed protein structure prediction into solved territory. RNA molecules are conformationally flexible, the structural database is small, and the evolutionary signal that powers protein predictors is often unavailable: RNA sequence alignments are shallow and hard to build, because base-pairing interactions distort the sequence similarity that alignment tools rely on. Most deep learning RNA predictors — trRosettaRNA, RoseTTAFoldNA, RhoFold+, NuFold — inherit the multiple sequence alignment dependency along with the transformer architecture, and they return a single static model rather than the ensemble a flexible molecule actually occupies.

RNAbpFlow, from the Bhattacharya Lab at Virginia Tech, replaces evolutionary conditioning with base-pairing conditioning. It is an SE(3)-equivariant conditional flow matching model that takes a nucleotide sequence together with base-pair maps and generates all-atom 3D structural ensembles end to end, using no MSAs, no homologous templates, and no physics-based refinement stage. The base-pair map is information a user can obtain from secondary structure prediction, chemical probing, or comparative analysis — sources that remain available exactly when deep alignments are not.

Released as a preprint in January 2025 by Sumit Tarafder and Debswapna Bhattacharya, it is the first conditional flow matching method for RNA 3D structure generation, extending the unconditional RNA-FrameFlow line of work into a predictive tool.

#Key Features

• Base-pair conditioning from three annotators: Canonical and non-canonical pairs are extracted with RNAView, MC-Annotate, and DSSR, each with distinct geometric and hydrogen-bond cutoffs, and supplied as edge features so the model sees a consensus rather than one tool's opinion.
• Nucleobase center representation: Each nucleotide becomes a rigid body frame with C1′ as origin and O4′–C1′–C2′ defining orientation, from which the remaining atoms are built through ten sub-frames and nine predicted torsion angles — giving all-atom output without a separate full-atom reconstruction step.
• Evolution-free by construction: No MSA search, no template retrieval, so runtime does not depend on database coverage and performance does not collapse on orphan RNAs.
• Ensemble output rather than a point estimate: Sampling many structures per target exposes the conformational heterogeneity of flexible RNAs; roughly 25 seconds produces 10 structures for a 70-nucleotide RNA on a single GPU.
• Base-pair-centric auxiliary losses: Additional loss terms explicitly reward realizing the conditioned canonical and non-canonical pairing geometry in the generated 3D coordinates.

#Technical Details

The framework builds on FrameFlow, an SE(3) flow matching formulation for fast protein backbone generation, with rotations initialized from IGSO3(σ = 1.5) and translations from a unit Gaussian, then iteratively transported toward the data distribution by a learned denoiser. Ablations settled on 6 structure-module blocks. Training used the RNA3DB split for internal benchmarking; for CASP15 evaluation the authors curated a separate trRosettaRNA-derived set restricted to chains released before May 2022 to exclude leakage, and for CASP16 they used the full RNA3DB release augmented with a cross-distillation set of 912 sequences under 200 nucleotides drawn from bpRNA-1m.

On 52 RNA3DB test targets, conditioning on all three base-pair maps reaches an average maximum TM-score of 0.51 and lDDT of 0.71, versus 0.38 and 0.49 for the sequence-only variant — improvements of 34.2% and 44.9%. Against the MD-based sampler RNAJP, mean TM-score is 0.38 versus 0.32, and 17.3% of RNAbpFlow decoys exceed TM-score 0.45 compared with 1.73% for RNAJP. On 14 public CASP16 targets, RNAbpFlow reaches an average maximum TM-score of 0.59 and lDDT of 0.70, ahead of the best-of-five submissions from the top automated servers, trRosettaRNA2 (0.54) and AlphaFold 3 (0.49), despite those methods using MSA or template information. The margin is largest on hard targets with shallow alignments (Neff ≤ 130).

#Applications

RNAbpFlow suits structural biologists modeling RNAs where alignments are shallow or absent — synthetic constructs, designed aptamers, viral elements from sparsely sampled lineages, and newly annotated non-coding RNAs. Because base-pair maps are the conditioning signal, experimentally derived secondary structure from SHAPE or DMS probing can be fed in directly to bias the 3D ensemble toward the solution-state fold. The ensemble output is useful for downstream scoring, docking, and hypothesis generation about junction geometry and loop–loop contacts.

#Impact

The result reframes what RNA 3D prediction needs: base-pairing, not evolutionary depth, carries most of the usable signal, and a generative sampler conditioned on it can beat MSA-dependent servers on hard targets. The work has since been published in Nature Methods. Practically, accuracy depends on the quality of the input base-pair maps — with predicted rather than native pairs, performance falls — and the model produces ensembles without an intrinsic confidence ranking, so a separate scoring method such as lociPARSE is needed to select models. Code is released under GPL-3.0 and the three trained checkpoints are deposited on Zenodo under CC-BY-4.0.