RNADiffFold

A folded RNA is often not one structure. Riboswitches, ribozymes and many regulatory elements interconvert between conformations that differ by a handful of base pairs, and both states are correct. Yet nearly every secondary-structure predictor — thermodynamic folders such as Mfold and RNAfold, and the deep learning methods that followed — returns exactly one answer per sequence, because each is built to score candidate pairings and take the maximum. The ensemble is discarded by construction.

RNADiffFold, from the Hangzhou Institute of Medicine, Chinese Academy of Sciences with collaborators at Zhejiang University of Technology and the Hangzhou Institute for Advanced Study of the University of Chinese Academy of Sciences, replaces that argmax with sampling. It represents a secondary structure as a binary L×L contact map — a black-and-white image in which a lit pixel is a base pair — and recasts prediction as pixel-wise segmentation. A multinomial (discrete-state) diffusion process corrupts the true contact map toward a uniform categorical distribution over 20 steps, and a U-Net learns to reverse it. At inference the model starts from noise, denoises under conditioning derived from the sequence, and emits a structure; a different random seed can yield a different one.

Nothing in the generative path imposes nesting or a thermodynamic prior, so the model is free to place pairings that hard-constrained methods rule out in advance. It appeared as a preprint in June 2024 and was published in Briefings in Bioinformatics that November.

#Key Features

• Contact maps as images: Each pixel of the L×L map is embedded into an eight-dimensional categorical vector, giving an L×L×8 tensor that the diffusion process corrupts with Gumbel-Softmax sampling. Because the denoiser is a fully convolutional U-Net, variable sequence lengths need no padding scheme.
• Four-stream conditioning: The condition construction unit supplies one-hot sequence encoding, a probability map from a UFold scoring network, and both sequence embeddings and multi-head attention maps from RNA-FM, each reduced by an MLP and fused into the denoiser at every timestep.
• Sampling and voting: Because prediction is generative, several structures can be drawn per sequence, clustered by similarity, and reduced to the most frequent one. Going from one sample to ten improved accuracy across all four test sets.
• Cross-family generalization: One trained checkpoint is applied to RNA families absent from training, the regime where constraint-heavy predictors degrade most sharply.
• Conformational coverage: On 14 CoDNaS-RNA clusters each holding two conformations of the same sequence, the model recovered the structural profiles of 11 and distinguished the alternative states in part.

#Technical Details

The scoring branch takes the Kronecker product of the one-hot sequence with itself (16 × L × L), concatenates a CDPfold pairing-probability matrix to reach 17 channels, and passes the result through a U-Net initialized from pretrained UFold weights and then fine-tuned. RNA-FM contributes a 640 × L embedding and a 240 × L × L attention tensor assembled from 20 heads across its 12 encoder blocks. Training runs in two stages on a single NVIDIA A40 with Adam at a learning rate of 1e-3 for up to 400 epochs: first on RNAStrAlign (30,451 sequences, 8 families), bpRNA TR0 and 2,717 Contrafold-pseudo-labeled mutants, then fine-tuning on PDB-derived structures. Reported F1 is 0.880 on ArchiveII and 0.711 on bpRNA TS0 — an 8.7% gain over UFold and 2.7% over RNA-FM on TS0 — 0.659 on the cross-family bpRNA-new set, and 0.736 on the PDB test sets. On the 993 TS0 sequences with long-range pairs (span greater than L/2) it reaches 0.739 F1 against UFold's 0.687. UFold remains ahead on ArchiveII (0.905) and on PDB (0.774), and accuracy falls for sequences beyond roughly 500 nucleotides.

#Applications

The practical draw is any setting where a single maximum-likelihood fold is the wrong abstraction: mapping riboswitch and ribozyme conformational switching, screening designed constructs whose function depends on an alternative fold, and annotating structure in newly sequenced families with no homologs to draw on. Repeated sampling also gives a rough confidence signal — sequences whose draws cluster tightly are ones the model is consistent about — useful triage before committing to chemical probing or crystallography.

#Impact

RNADiffFold is an early demonstration that discrete diffusion, developed for categorical image and text data, transfers cleanly to RNA base-pairing matrices, and that a generative formulation buys conformational diversity that argmax decoders cannot express. Its gains are clearest where the field is weakest — unseen families and long-range pairs — while its within-family numbers sit close to the supervised methods it conditions on. The authors are candid that the approach inherits whatever its frozen conditioners encode, that no conditioning network was retrained for the diffusion objective, and that the segmentation loss is poorly matched to contact maps in which most pixels are trivially unpaired.