RNAbpFlow Rivals AlphaFold 3 in RNA Structure Prediction Without Evolutionary Data

AlphaFold 2’s breakthrough in protein folding sparked intense interest in applying artificial intelligence to RNA 3D structure prediction, but most existing RNA predictors rely heavily on multiple‑sequence alignments (MSA) or homologous templates, struggle to capture base‑pair constraints, and output only a single static conformation.

To address these three bottlenecks, Debswapna Bhattacharya and Sumit Tarafder at Virginia Tech developed RNAbpFlow, a SE(3)-equivariant flow‑matching model that conditions on the RNA nucleotide sequence and explicit base‑pair matrices. The model inherits the FrameFlow architecture, replaces the protein backbone with NuFold’s nucleotide representation, and incorporates three independent L×L binary matrices (derived from RNAView, MC‑Annotate, and DSSR) as a three‑channel bias input.

Training data were drawn from the RNA3DB repository (version released 2024‑04‑26). After rigorous quality control—removing structures with missing atoms, protein contaminants, or incomplete base‑pairing—the authors obtained a non‑redundant training set of 560 sequences (30‑200 nt) and a test set of 48 sequences. For CASP15 and CASP16 blind‑test evaluations, separate non‑overlapping training pools were constructed: 731 RNAs for CASP15 and 994 RNAs plus 2,170 cross‑distilled high‑confidence structures for CASP16. All splits obey strict sequence‑structure non‑overlap to prevent data leakage.

Model training used PyTorch‑Lightning on eight NVIDIA H100 GPUs, Adam optimizer with a learning rate of 0.0001, and 1,500 epochs. The flow‑matching component learns a time‑dependent vector field Uₜ that transports Gaussian noise to the target distribution of atomic coordinates, following the design of FrameFlow and leveraging AlphaFold 2’s structural modules as the backbone.

Benchmarking against RNAJP—a coarse‑grained molecular dynamics sampler that also uses base‑pair information—showed that RNAbpFlow achieved an average lDDT of 0.66 versus 0.59 for RNAJP, and a TM‑score of 0.38 versus 0.32. In terms of successful fold detection, RNAbpFlow recovered correct global topology for 66.7 % of targets (TM‑score ≥ 0.5) and 25 % based on lDDT, compared with 41.7 % and 0 % for RNAJP. Moreover, 13.4 % of RNAbpFlow’s 12 000 generated conformations exceeded a TM‑score of 0.45, while only 1.73 % of RNAJP’s conformations reached that threshold.

On the CASP15 blind‑test set, providing experimentally derived base‑pair matrices boosted RNAbpFlow’s average TM‑score to 0.48 (RMSD 7.77 Å, INF‑NWC 0.62). When only predicted base‑pair matrices from IPKnot, SPOT‑RNA, and RibonanzaNet were used, performance dropped to TM‑score 0.40, RMSD 10.70 Å, INF‑NWC 0.48, highlighting the importance of accurate pairing information.

In the CASP16 competition (14 targets ≤200 nt), RNAbpFlow produced at least one correctly folded conformation for 12 targets (85.7 %), whereas AlphaFold 3 succeeded on only 8 targets (57.1 %). For longer RNAs (>200 nt), RNAbpFlow still outperformed NuFold, trRosettaRNA2, and DRfold2 but fell behind AlphaFold 3, a shortfall attributed to reduced base‑pair fidelity on extended sequences.

Overall, RNAbpFlow demonstrates that a deep‑generative, SE(3)-equivariant framework conditioned solely on sequence and base‑pair constraints can surpass AlphaFold 3 on difficult RNA targets without any evolutionary information, offering a promising route for large‑scale RNA conformational sampling and dynamics studies.