RNA Grammar Variational Autoencoder

• The paper demonstrates that RGVAE achieves de novo RNA design by integrating stochastic context-free grammar masking with latent-space optimization to enforce structural validity and optimize minimum free energy.
• It employs a CNN encoder and an RNN decoder with stack-based rule masking, transforming RNA sequences into parse trees of production rules for robust grammar-based generation.
• Empirical results indicate that RGVAE outperforms baseline models by effectively meeting multi-constraint objectives such as MFE, GC-content, and target secondary structures.

Searching arXiv for recent and foundational papers on RNA Grammar Variational Autoencoder and related grammar-based RNA generation. RNA Grammar Variational Autoencoder (RGVAE) is a grammar-constrained variational autoencoder for RNA design that combines continuous latent-space optimization with a stochastic context-free grammar (SCFG) intended to reflect RNA secondary-structure organization. In the formulation introduced in 2025, RNA sequences are not modeled as unconstrained nucleotide strings; instead, each sequence is mapped to a parse tree and then to a sequence of grammar production rules, which the model encodes into a low-dimensional latent representation and decodes under grammar-based masking. The central aim is multi-constraint de novo RNA design under objectives such as low minimum free energy (MFE), target GC-content, length constraints, motif inclusion or exclusion, target secondary structure, and base-position constraints (Zarnaghinaghsh et al., 21 Jul 2025).

1. Problem setting and design rationale

RGVAE is situated in the RNA design problem, where the target is to generate RNA sequences that satisfy multiple desired constraints simultaneously. The formulation explicitly includes forming a stable secondary structure, having low MFE, matching a target GC-content, satisfying length constraints, including mandatory motifs, excluding forbidden motifs, matching a target secondary structure, and satisfying base-position constraints. The motivating difficulty is combinatorial: if an RNA has length $n$, the number of possible sequences grows exponentially with $n$, making exhaustive search infeasible except for very short RNAs (Zarnaghinaghsh et al., 21 Jul 2025).

The method is motivated by a structural critique of ordinary sequence generators and unconstrained VAEs. In that view, RNA design is not merely string generation, because RNA function depends strongly on secondary structure rather than only on primary sequence. A nucleotide string that appears plausible at the sequence level may fail to fold into a useful or stable structure. Standard VAEs offer a continuous latent space for optimization, but unconstrained decoders may generate invalid or poor-quality outputs and may create “dead areas” in latent space where decoded samples are unusable. RGVAE addresses this by restricting decoding to derivations admitted by an RNA grammar, so that the decoder emits valid production-rule sequences rather than arbitrary token strings (Zarnaghinaghsh et al., 21 Jul 2025).

The resulting model is presented as a de novo generative design method rather than a sequence-editing system. Training begins from a corpus of RNAs, but the operational design workflow proceeds by latent sampling or latent optimization, followed by grammar-constrained decoding and downstream evaluation of the generated candidates. Because the reported training set consists of 101,754 tRNA sequences from Rfam, the learned design space is intentionally biased toward a family with a well-known cloverleaf consensus secondary structure; this bias is part of the model’s inductive prior rather than an incidental dataset choice (Zarnaghinaghsh et al., 21 Jul 2025).

2. Formal grammar, representation, and latent-variable formulation

RGVAE follows the general grammar-VAE paradigm in which a structured object is represented by its parse tree under a grammar and then linearized as a sequence of production rules rather than as a raw character string. The underlying grammar formalism is a CFG $G=(V,T,P,S)$, with nonterminal symbols $V$, terminal symbols $T$, production rules $P$, and start symbol $S$. A production rule has the form

$B \rightarrow \alpha$

where $B\in V$ and $\alpha \in (V\cup T)^*$. In the stochastic version used by RGVAE, production rules carry probabilities, and the probabilities of all productions leaving a given nonterminal sum to 1 (Zarnaghinaghsh et al., 21 Jul 2025).

The specific RNA grammar reported for the model is

$n$0

$n$1

$n$2

where $n$3 is the start symbol, $n$4 and $n$5 are nonterminals, $n$6 denotes any single terminal, and $n$7 denotes the base pair of $n$8. This grammar is intended to describe nested RNA secondary-structure patterns through base-pairing relations. The paper occasionally writes “SCFT” in places where the intended object is clearly the SCFG (Zarnaghinaghsh et al., 21 Jul 2025).

Each RNA sequence is first transformed into a parse tree under this grammar and then into a sequence of production rules extracted from that tree. Each production rule is represented as a one-hot vector over the grammar’s rule vocabulary, so the sequence has the form

$n$9

up to the paper’s slightly corrupted typesetting. The encoder is described as a deep CNN, stated to be the same as in Kusner et al.’s “Grammar Variational Autoencoder” (Kusner et al., 2017), while the decoder is an RNN that outputs a sequence of unnormalized log-probability vectors over grammar rules. The latent prior is

$G=(V,T,P,S)$0

Latent dimensions 10, 25, and 56 were compared experimentally, and dimension 10 was selected (Zarnaghinaghsh et al., 21 Jul 2025).

The key validity mechanism is stack-based grammar masking. Decoding begins with a last-in-first-out stack initialized by the start symbol $G=(V,T,P,S)$1. At each step, the model pops the top nonterminal $G=(V,T,P,S)$2, constructs a binary mask $G=(V,T,P,S)$3 that retains only the rules whose left-hand side is $G=(V,T,P,S)$4, applies that mask to the decoder logits, samples among the remaining rules, and pushes right-hand-side nonterminals onto the stack from right to left. The masked rule distribution is

$G=(V,T,P,S)$5

The denominator contains a minor indexing typo in the paper; the intended construction is a masked softmax over rules compatible with the active nonterminal (Zarnaghinaghsh et al., 21 Jul 2025).

The decoder likelihood over the production-rule sequence is written as

$G=(V,T,P,S)$6

and the VAE objective is given as

$G=(V,T,P,S)$7

The notation is somewhat inconsistent with standard VAE form, but the intended objective is the usual ELBO, with a reconstruction term over the production-rule sequence and a KL term against the Gaussian prior. The paper also states that it follows Kusner et al. in using non-centered parameterization, i.e. the reparameterization trick (Zarnaghinaghsh et al., 21 Jul 2025, Kusner et al., 2017).

3. End-to-end workflow and optimization regime

The operational workflow begins with 101,754 tRNA sequences from Rfam, chosen because tRNAs have a well-known cloverleaf consensus secondary structure. The SCFG is parameterized on the training data using the inside-outside algorithm, after which each RNA sequence is parsed into a grammar derivation and converted into a one-hot production-rule sequence. The CNN encoder maps this rule sequence into a latent Gaussian distribution, a latent vector $G=(V,T,P,S)$8 is sampled, and the model is trained by optimizing the ELBO. The paper reports three-fold cross-validation, testing of multiple hyperparameters, and evaluation across 10 different optimizations with 10 distinct random seeds per scenario, although many lower-level training details are not reported clearly (Zarnaghinaghsh et al., 21 Jul 2025).

Design is performed in latent space using Bayesian optimization. The formulation is stated generally as

$G=(V,T,P,S)$9

but the practical implementation is over latent points that are decoded into RNA sequences. For multiple constraints, the paper states that “the summation of the scores of different constraints for Bayesian optimization is implemented” and that “we summed the scores of different constraints together to make the objective function.” No explicit acquisition-function or surrogate-model formula is given, so the documented mechanism is an additive objective over decoded candidates (Zarnaghinaghsh et al., 21 Jul 2025).

The explicitly defined constraints are the MFE constraint, GC-content constraint, length constraint $V$0, secondary structure constraint $V$1, base positional constraint $V$2, mandatory motifs $V$3, and forbidden motifs $V$4. These are not described as grammar constraints directly. Instead, grammar validity is enforced during decoding, while most application-specific requirements are handled by latent-space search plus post hoc scoring and filtering. This separation between syntactic validity and design-objective satisfaction is a central feature of the method (Zarnaghinaghsh et al., 21 Jul 2025).

The reported implementation details are intentionally sparse. The paper specifies a CNN encoder, an RNN decoder, inside-outside estimation of SCFG probabilities, Bayesian optimization in latent space, and additive aggregation of constraint scores. It does not clearly report learning rate, optimizer type, batch size, number of epochs, KL annealing schedule, exact CNN or RNN layer sizes, decoder hidden sizes, training time, runtime, exact Bayesian optimization surrogate or acquisition setup, or the exact parsing algorithm used to select a single parse tree from potentially ambiguous parses. The source code and data are provided through a GitHub repository and a Zenodo DOI, which are important reproducibility details even though they do not fill all methodological gaps (Zarnaghinaghsh et al., 21 Jul 2025).

4. Reported empirical behavior

The experiments emphasize that grammar choice materially affects performance. In addition to the primary grammar $V$5, the study tests an alternative grammar

$V$6

$V$7

$V$8

A major reported finding is that optimized sequences generated using $V$9 achieve better MFE than the training data, whereas using $T$0 yields performance worse than the training data. Within the paper’s framing, the SCFG is therefore not a neutral representation layer but a strong inductive bias (Zarnaghinaghsh et al., 21 Jul 2025).

Several use cases illustrate the reported design behavior:

Scenario	Best RGVAE outcome	Baseline note
MFE-only design	$T$1	Training: $T$2; RNAGEN: $T$3
GC target 50%	$T$4	Training: $T$5; RNAGEN: $T$6; Random: $T$7
GC target 50% + length 100–150	$T$8, GC 51%, length 149	Training: $T$9; Random grammar-based: $P$0
Target structure ......(((....))) + first two bases “UC” + GC 50%	8 sequences found; one exact match UCACUUGGCUGUAGCU	Training, random, random-with-grammar, RNAGEN: none
105-nt SAM riboswitch, two target structures	Best summed distance 11.47	Histograms lower than training data and RNAGEN

The most direct quantitative claim for single-objective optimization is the MFE-only experiment, where the minimum MFE in the training data is $P$1, the minimum MFE generated by RGVAE with grammar $P$2 is $P$3, and the minimum MFE generated by RNAGEN is $P$4. For mixed MFE and GC-content objectives, the model is evaluated at target GC-contents of 20%, 30%, 50%, and 70%, selecting the best sequence within $P$5 of the target. At 50% GC, the reported values are particularly strong: training $P$6, RGVAE $P$7, RNAGEN $P$8, and random $P$9. At 20% GC, random generation slightly beats RGVAE in best MFE, but RGVAE still outperforms training data and RNAGEN; RNAGEN could not generate valid sequences at 20% or 30% under that criterion (Zarnaghinaghsh et al., 21 Jul 2025).

The multi-constraint examples are especially important for understanding the intended use of the model. Under GC-content around 50% and length between 100 and 150, the best RGVAE candidate has MFE $S$0, GC 51%, and length 149, compared with training MFE $S$1 and random grammar-based generation MFE $S$2. In motif-constrained experiments, one setup imposes mandatory motif $S$3, forbidden motifs $S$4, and target GC-content 50%; the best RGVAE example satisfies the motif constraints, has GC-content 48.43%, and MFE $S$5. A second setup with mandatory motif $S$6, forbidden motifs $S$7, and target GC-content 50% yields a best RGVAE example with GC-content 51.12%, MFE $S$8, and satisfied motif constraints (Zarnaghinaghsh et al., 21 Jul 2025).

The clearest structural success case uses target secondary structure

$S$9

with the additional constraints that the first two nucleotides must be “UC” and the target GC-content is 50%. After filtering, the training data contains no sequence satisfying all requirements; random generation without grammar finds none; random generation with grammar finds none; RNAGEN finds none; and RGVAE finds 8 sequences, among which one sequence, UCACUUGGCUGUAGCU, has GC-content 50% and the exact target secondary structure. The SAM riboswitch experiment extends the setting to a 105-nt riboswitch with two alternative secondary structures; for sequences of length 80–100, histograms show lower summed alignment scores for RGVAE than for training data and RNAGEN, and the best RGVAE sequence attains a summed distance of 11.47. This suggests that the model can be applied to multi-structure behavior rather than only to a single target fold (Zarnaghinaghsh et al., 21 Jul 2025).

5. Relation to grammar VAEs, RNA grammars, and non-grammar generative models

RGVAE directly inherits its architectural logic from the general Grammar VAE framework of Kusner et al., where valid structured objects are represented as parse trees under a CFG, encoded as production-rule sequences, and decoded by a recurrent model whose output is masked according to the active nonterminal. In that earlier formulation, grammar-based decoding improved validity and latent-space coherence relative to character-level VAEs, and the RGVAE paper explicitly states that its encoder is the same as in Kusner et al.’s GVAE (Kusner et al., 2017, Zarnaghinaghsh et al., 21 Jul 2025).

Within RNA-specific grammar research, RGVAE occupies only one point in a larger design space. “A Graph Grammar for Modelling RNA Folding” formalizes RNA secondary structure as a graph $B \rightarrow \alpha$0 with nucleotide-labeled vertices, backbone edges, and base-pair edges, and models folding as double-pushout graph rewriting over motif classes such as hairpins, helices, bulges, internal loops, and multi-branched loops. That framework separates local structural admissibility from global free-energy guidance and can be read as a symbolic substrate for a validity-preserving RNA decoder, but it is not itself a VAE and does not define a latent-variable model (Mamuye et al., 2016).

At a different level of generality, “Topological language for RNA” introduces RNAFeatures$B \rightarrow \alpha$1, a labeled context-free grammar over $B \rightarrow \alpha$2-structures that is capable of generating any RNA structure including pseudoknot structures by encoding crossing information into finite labels attached to otherwise noncrossing derivations. That formalism is relevant because it shows how a context-free derivation can be retained even when pseudoknotted topology is present, but it is substantially more elaborate than the nested SCFG used in RGVAE and is not the grammar adopted in the 2025 design model (Huang et al., 2016).

A contrasting non-grammar line is represented by “Uncovering the Folding Landscape of RNA Secondary Structure with Deep Graph Embeddings,” which introduces a geometric scattering autoencoder for RNA secondary-structure graphs. That model learns a latent organization of RNA graphs by structure and energy and supports approximate graph generation through a separate inversion network, but it does not define a formal RNA grammar, does not decode derivation rules, and does not guarantee syntactic validity through a grammar-constrained decoder. The contrast clarifies what RGVAE adds: grammar validity at decode time rather than approximate validity learned indirectly in graph space (Castro et al., 2020).

6. Guarantees, limitations, and interpretive boundaries

The strongest guarantee provided by RGVAE is grammar validity during decoding. Because rule selection is constrained by a stack-tracked nonterminal state and by rule masks derived from the SCFG, every completed derivation is valid with respect to the chosen RNA grammar. This improves structural plausibility in the narrow syntactic sense of the grammar and increases the proportion of validly structured outputs relative to unconstrained decoding (Zarnaghinaghsh et al., 21 Jul 2025).

That guarantee should not be conflated with a guarantee of thermodynamic correctness. The RGVAE paper states that using the SCFG can ensure that the generated RNA sequence can form a thermodynamically stable secondary structure, but the reported methodology evaluates stability mainly through downstream quantities such as MFE. The paper does not prove that every decoded sequence folds into a unique or intended minimum-energy structure, nor does it derive thermodynamic optimality from the grammar alone. The most justified interpretation is therefore that grammar constraints provide a structural prior and validity mechanism, while thermodynamic stability is optimized statistically in latent space and assessed downstream through folding and energy evaluation (Zarnaghinaghsh et al., 21 Jul 2025).

Several limitations follow directly from the reported experiments and design assumptions. Performance depends strongly on grammar design, as shown by the comparison between $B \rightarrow \alpha$3 and $B \rightarrow \alpha$4. The grammar is aimed at nested RNA secondary-structure patterns, and the paper does not discuss pseudoknots, so there is no evidence that the model handles them. Constraint handling is mostly score-based rather than hard-wired into decoding: MFE, GC-content, motifs, length, target structure, and positional requirements are combined through latent-space optimization and filtering rather than enforced uniformly at every generation step. Training-data bias is substantial because the model is trained on tRNAs, and the conclusion explicitly suggests using different RNA families when other structural classes are desired. Reproducibility is also limited by omitted implementation details such as optimizer, batch size, epoch count, exact CNN and RNN specifications, and the parsing strategy used for ambiguous SCFG derivations (Zarnaghinaghsh et al., 21 Jul 2025).

A plausible implication is that future RNA grammar VAEs could be hybridized with adjacent grammar and graph formalisms rather than remaining tied to a simple nested SCFG. The graph-grammar approach suggests a decoder action space over explicit motif-preserving structural moves (Mamuye et al., 2016), the topological-language approach suggests a route toward pseudoknot-capable labeled derivations (Huang et al., 2016), and the graph-scattering approach suggests multiscale encoders and latent organization criteria that are not themselves grammar-based (Castro et al., 2020). Such combinations are not part of the reported RGVAE model, but they define a natural continuation of the broader grammar-constrained RNA generative modeling program.