RIFT-VAE

A dot-bracket string is the standard way to hand a target RNA secondary structure to a neural network, and it quietly discards the thing that matters most. The opening and closing brackets of a base pair form one structural relation, but in a flat character sequence they can sit hundreds of tokens apart, leaving the model to rediscover their pairing from position alone. A context-free grammar sees the same structure differently: one recursive production emits both halves of a pair in a single operation, and the parse tree expresses the nesting of stems and loops directly.

RIFT-VAE, from Yasubumi Sakakibara's group at Kitasato University with Keio University, builds an RNA inverse-folding model on exactly that representation. A pseudoknot-free dot-bracket target is parsed into a CFG tree, each production-rule application becomes a token, and every token carries a tree-position encoding from its depth and sibling order, so identical rules in different structural contexts stay distinguishable. A Transformer conditional variational autoencoder assigns nucleotide labels — single bases at unpaired terminals, ordered base-pair labels at paired ones — to the terminal-generating nodes, read back in 5′-to-3′ order as a sequence.

The latent-variable formulation matters because inverse folding is one-to-many. Where gRNAde conditions on a 3D backbone and yakRNA Design denoises masked tokens under Gene Ontology constraints, RIFT-VAE keeps the classical secondary-structure interface and adds an objective search-based solvers do not optimize at all: the sequence statistics of natural RNA families.

#Key Features

• Grammar-conditioned encoding: Four nested grammars of increasing granularity (Normal, Loop, Loop-Stem, Loop-Stem α) make hairpin boundaries, internal-loop asymmetry, stacked pairs and multibranch junctions explicit in the parse-tree topology.
• Conditional latent space: Sampling a latent vector from a standard normal prior under a fixed target yields many distinct candidates for one structure, not one deterministic answer.
• Self-refinement learning: A generation that folds to the wrong structure is not discarded — RNAfold's prediction for it becomes a new, valid structure–sequence training pair.
• Latent-space optimization: The cross-entropy method fits a diagonal Gaussian over the latent variable, decodes and scores candidates with RNAfold, and reweights toward the elite fraction. The decoder stays frozen — this is guided sampling, not per-target refitting.
• GC-content conditioning: A target GC fraction enters the decoder as a scalar condition alongside the structural one, useful for temperature-responsive RNAs.
• Family signature retention: Designs score above family-specific Infernal gathering cutoffs for their target covariance models, which the search-based designs tested did not.

#Technical Details

The released configuration is a six-layer Transformer encoder-decoder with model dimension 512, eight attention heads, feed-forward dimension 2,048, a structural-token vocabulary of 351, a decoder-label vocabulary of 559, a latent dimension of 128, and a maximum serialized tree length of 178 nodes; the decoder embedding and output projection share weights. Training used AdamW at learning rate 2 × 10⁻⁴, batch size 64, up to 200 epochs across six GPUs. The corpus is RNAcentral sequences of at most 500 nucleotides from families with characteristic folds — rRNA, tRNA, snRNA, snoRNA, miRNA, RNase P, ribozymes, Y RNA, RsmA/MRP — giving 2,128,591 training and 43,118 test examples, with RNAfold default MFE predictions as targets.

The ablation separates the contributions cleanly. On the RNAcentral test set, exact-match recovery (RNAfold-Correct) rises from 0.312 for the deterministic base Transformer to 0.468 with the CVAE, 0.499 with grammar expansion, 0.654 with self-refinement learning, and 0.833 with latent optimization; MCC reaches 0.994. On EteRNA100v2 the ladder runs 0.140, 0.210, 0.220, 0.250, 0.760, with MCC 0.977. The authors are explicit that the final step dominates, and call the system a hybrid generative-search framework rather than a faster replacement for search.

#Applications

The immediate use is designing structured functional RNAs — ribozymes, riboswitch scaffolds, regulatory elements — where a fold is specified up front and sequence properties also matter. A hammerhead ribozyme case study recovered the conserved catalytic-core and cleavage-site nucleotides and produced a correct-family covariance-model hit, where the DesiRNA design of the same target did neither. Under a matched 3,600-second budget on EteRNA100, RIFT-VAE candidates used as warm starts improved every downstream solver tested, most dramatically RNAInverse, whose exact-match rate went from 0.297 to 0.803.

#Impact

RIFT-VAE bridges formal-language descriptions of RNA structure, which predate deep learning by decades, and modern conditional sequence generation. Its argument is that structural correctness is not the whole objective: pretraining on natural sequences buys family-level naturalness that folding-energy optimization discards. The limits are stated plainly by the authors. Every score measures agreement with RNAfold's MFE model rather than an experimental structure; the representation covers only pseudoknot-free canonical base pairs; results report the best of 10 to 20 sampled candidates, so diversity is uncharacterized; and the family holdout excludes families by label, not by clan-level homology. DesiRNA still leads on EteRNA100 exact match, 0.907 against 0.760. Code is on GitHub without an explicit license and checkpoints are on Zenodo under CC BY 4.0; the work is a preprint, and no designed sequence has been assayed experimentally.