RNAGenScape: Latent RNA Design Framework

• RNAGenScape is a latent-space neural framework that jointly represents RNA sequences and structures to generate and optimize biologically valid molecules.
• It spans two formulations: the 2021 model emphasizes RNA secondary structure generation via variational autoencoding, while the 2025 model focuses on mRNA property optimization using an organized autoencoder and Langevin dynamics.
• Key operations include latent-space navigation with manifold projection to ensure that decoded sequences remain close to biologically feasible configurations.

RNAGenScape denotes latent-space neural frameworks for RNA design in which sequence representations are organized so that generation, optimization, and interpolation remain aligned with biochemical or functional constraints. In the 2021 formulation, the system is presented as a model for neural representation and generation of RNA secondary structures, jointly embedding primary sequence, dot-bracket structure, planar molecular graphs, and junction-tree decompositions within a variational architecture (Yan et al., 2021). In the 2025 formulation, RNAGenScape is presented as a property-guided optimization and interpolation framework for mRNA sequences, combining an organized autoencoder, a denoising manifold projector, and manifold Langevin dynamics to steer sequences toward desired property regimes while remaining close to a learned viable manifold (Liao et al., 14 Oct 2025).

1. Conceptual scope and lineage

A recurring source of confusion is that RNAGenScape does not refer to a single frozen architecture. The term spans two related but distinct formulations. The earlier system focuses on RNA secondary-structure-aware representation learning and variational generation; the later system focuses on mRNA property optimization under scarce and imbalanced data through iterative latent updates with projection back to the manifold. Both formulations center latent-space geometry, but they do so with different inductive biases and optimization operators.

Aspect	2021 formulation	2025 formulation
Molecular object	RNA sequence + secondary structure	mRNA sequence
Core machinery	VAE + CNF prior + decoders	Organized autoencoder + projector + Langevin dynamics
Main operations	generation, targeted optimization	property-guided optimization, interpolation

In the 2021 system, the latent variable is used to jointly represent sequence and structural modalities and to support targeted optimization with respect to RNA-binding protein interaction. In the 2025 system, the latent variable is explicitly organized by a target property such as translation efficiency, stability, or ribosome load, and updates are performed by a discrete Langevin rule followed by a learned projection. This suggests a methodological shift from direct generative modeling of RNA structure toward controlled navigation on a learned mRNA manifold.

2. Multi-modal RNA representation in the 2021 formulation

The 2021 formulation represents an RNA molecule in four coupled forms: the primary sequence $x=(x_1,\dots,x_L)$ with $x_i\in\{A,C,G,U\}$; the dot-bracket string $\mathcal S=(\dot x_1,\dots,\dot x_L)$ with $\dot x_i\in\{.,(,)\}$; the planar molecular graph $\mathcal G=(V,E)$ whose nodes are nucleotides and whose edges include backbone edges $(i,i+1)$ and base-pair edges; and a junction-tree hypergraph $\mathcal T=(\{\hat G_1,\dots,\hat G_D\},F)$ in which each hypernode $\hat G_i$ is labeled as one of four substructure types $\mathcal L(\hat G_i)\in\{S,H,I,M\}$, corresponding to stem, hairpin, internal loop, and multiloop (Yan et al., 2021).

Three encoders are defined over these representations. The sequence-plus-dot-bracket encoder, denoted LSTMVAE, uses a vocabulary of size $4\times 3=12$ consisting of all $x_i\in\{A,C,G,U\}$0 pairs, followed by one-hot embedding, a stacked Bi-LSTM, multi-head self-attention, and global max-pooling to obtain $x_i\in\{A,C,G,U\}$1, from which linear heads produce $x_i\in\{A,C,G,U\}$2 and $x_i\in\{A,C,G,U\}$3. The graph-based encoder, GraphVAE, builds an edge-feature MPNN with gating. For each directed edge $x_i\in\{A,C,G,U\}$4, it initializes

$x_i\in\{A,C,G,U\}$5

where $x_i\in\{A,C,G,U\}$6 is a one-hot nucleotide embedding and $x_i\in\{A,C,G,U\}$7 encodes backbone versus base-pair status. After $x_i\in\{A,C,G,U\}$8 message-passing steps, node embeddings are aggregated, processed by a Bi-LSTM over $x_i\in\{A,C,G,U\}$9, and max-pooled to obtain $\mathcal S=(\dot x_1,\dots,\dot x_L)$0.

The hierarchical encoder, HierVAE, adds an explicit substructure decomposition. It first computes nucleotide embeddings $\mathcal S=(\dot x_1,\dots,\dot x_L)$1 with the G-MPNN, then forms hypernode inputs

$\mathcal S=(\dot x_1,\dots,\dot x_L)$2

where $\mathcal S=(\dot x_1,\dots,\dot x_L)$3 is the assignment map from hypernodes to nucleotides. A tree-GRU message-passing procedure on the junction tree generates hypernode states, and a depth-first traversal of $\mathcal S=(\dot x_1,\dots,\dot x_L)$4 through another Bi-LSTM followed by max-pooling yields $\mathcal S=(\dot x_1,\dots,\dot x_L)$5. Linear heads then produce $\mathcal S=(\dot x_1,\dots,\dot x_L)$6 and $\mathcal S=(\dot x_1,\dots,\dot x_L)$7.

The significance of this representational stack is the explicit integration of RNA folding regularity into the encoder. The molecular graph captures local connectivity and base pairing, while the junction-tree hierarchy encodes higher-order structural motifs. A plausible implication is that the model’s latent geometry is shaped not only by sequence similarity but also by topological regularities of nested secondary structure.

3. Variational generation and targeted design of RNA secondary structures

The 2021 system is trained as a variational autoencoder with a learnable prior $\mathcal S=(\dot x_1,\dots,\dot x_L)$8 and decoder $\mathcal S=(\dot x_1,\dots,\dot x_L)$9, using an approximate posterior $\dot x_i\in\{.,(,)\}$0. Its objective is the ELBO

$\dot x_i\in\{.,(,)\}$1

with $\dot x_i\in\{.,(,)\}$2 annealed during training (Yan et al., 2021).

The prior is implemented as a continuous normalizing flow. The model transforms $\dot x_i\in\{.,(,)\}$3 to $\dot x_i\in\{.,(,)\}$4, with density

$\dot x_i\in\{.,(,)\}$5

At training time, $\dot x_i\in\{.,(,)\}$6 is evaluated exactly by sampling $\dot x_i\in\{.,(,)\}$7 and using the CNF density; at test time, one samples $\dot x_i\in\{.,(,)\}$8 and inverts the flow to obtain $\dot x_i\in\{.,(,)\}$9.

Two decoders are used. The linear string decoder is a single-layer unidirectional LSTM whose initial hidden state is a linear projection of $\mathcal G=(V,E)$0 and which emits one of the 12 $\mathcal G=(V,E)$1 tokens plus an end symbol under teacher forcing and cross-entropy loss. The hierarchical tree-plus-segment decoder interleaves topology prediction, node-label prediction, and segment decoding during a depth-first traversal, with the segment LSTM initialized by

$\mathcal G=(V,E)$2

Structural regularization is enforced at inference time by hard masks on decoder logits. The rules are: base-pair complementarity restricted to A–U, G–C, or G–U; hairpin loops with at least 3 unpaired nucleotides, so that if an open “(” occurs at position $\mathcal G=(V,E)$3 then its matching “)” at $\mathcal G=(V,E)$4 requires $\mathcal G=(V,E)$5; and pairing at most once per nucleotide, with no crossing or pseudoknots. These constraints are designed to guarantee valid, nested, biochemically feasible secondary structures.

The same formulation also introduces a semi-supervised VAE for RBP-binding properties. A small MLP classifier $\mathcal G=(V,E)$6 is placed on top of $\mathcal G=(V,E)$7, and the joint loss is

$\mathcal G=(V,E)$8

Targeted design is then performed by activation maximization:

$\mathcal G=(V,E)$9

optionally with small Gaussian noise, followed by decoding and evaluation by an external oracle.

4. Organized autoencoder for mRNA property supervision

The 2025 formulation defines RNAGenScape as a property-guided manifold Langevin dynamics framework for mRNA sequences. Its front end is an organized autoencoder (OAE). Inputs are one-hot encoded mRNAs of length $(i,i+1)$0 over a 7-token alphabet: $(i,i+1)$1, A, U, T, G, C, N. The encoder $(i,i+1)$2 consists of three 1D-convolutional blocks with GroupNorm, GELU, and SE-block, followed by adaptive average pooling to length 8 and a linear head that produces $(i,i+1)$3 with $(i,i+1)$4. A property predictor $(i,i+1)$5 is a 3-layer MLP with GELU activations and dropout $(i,i+1)$6, mapping $(i,i+1)$7. The decoder $(i,i+1)$8 is a progressive 1D decoder that linearly projects $(i,i+1)$9 to a seed map of shape $\mathcal T=(\{\hat G_1,\dots,\hat G_D\},F)$0, repeatedly upsamples by factor 2 with two residual convolutional blocks until reaching $\mathcal T=(\{\hat G_1,\dots,\hat G_D\},F)$1 positions, and refines with two final convolutional blocks to produce logits $\mathcal T=(\{\hat G_1,\dots,\hat G_D\},F)$2 (Liao et al., 14 Oct 2025).

The latent space is explicitly supervised by the target property. In addition to reconstructing $\mathcal T=(\{\hat G_1,\dots,\hat G_D\},F)$3, the model requires $\mathcal T=(\{\hat G_1,\dots,\hat G_D\},F)$4 to be predictive of $\mathcal T=(\{\hat G_1,\dots,\hat G_D\},F)$5, where the target may be translation efficiency, stability, or ribosome load. The loss is

$\mathcal T=(\{\hat G_1,\dots,\hat G_D\},F)$6

with $\mathcal T=(\{\hat G_1,\dots,\hat G_D\},F)$7 for Zebrafish and $\mathcal T=(\{\hat G_1,\dots,\hat G_D\},F)$8 for the smaller datasets.

This organization by property is central to the later RNAGenScape formulation. Rather than treating optimization as an external search over a latent space learned only for reconstruction, the OAE makes the latent representation itself predictive of the target variable. This reduces the mismatch between decoding fidelity and optimization direction.

5. Manifold projector, Langevin dynamics, and interpolation

After training the OAE, RNAGenScape performs guided random walks in latent space. The discrete update is

$\mathcal T=(\{\hat G_1,\dots,\hat G_D\},F)$9

where $\hat G_i$0 is the step size, $\hat G_i$1 is a temperature controlling exploration versus exploitation, and $\hat G_i$2 is the utility function: to maximize the property, $\hat G_i$3; to minimize it, $\hat G_i$4 (Liao et al., 14 Oct 2025). The noise term is isotropic Gaussian noise, and the factor $\hat G_i$5 matches overdamped Langevin discretization.

The operator $\hat G_i$6 is a denoising manifold projector trained to retract off-manifold points back near the manifold. Training uses clean latents $\hat G_i$7 formed from original embeddings together with geometry-aware SUGAR samples, followed by a short Gaussian corruption chain

$\hat G_i$8

and the projector objective

$\hat G_i$9

In practice $\mathcal L(\hat G_i)\in\{S,H,I,M\}$0, and for Zebrafish the corruption scales are $\mathcal L(\hat G_i)\in\{S,H,I,M\}$1.

Optimization proceeds by encoding $\mathcal L(\hat G_i)\in\{S,H,I,M\}$2, iterating drift computation, noisy update, and projection for $\mathcal L(\hat G_i)\in\{S,H,I,M\}$3, and decoding the final latent to $\mathcal L(\hat G_i)\in\{S,H,I,M\}$4. Interpolation between source and target latents replaces the property gradient by a normalized pull toward $\mathcal L(\hat G_i)\in\{S,H,I,M\}$5:

$\mathcal L(\hat G_i)\in\{S,H,I,M\}$6

The decoded intermediates are intended to remain close to the viable mRNA manifold throughout the trajectory.

A notable feature of this formulation is that optimization and interpolation are not separated. Both are instances of latent-space dynamics with projection, differing only in the choice of drift term.

6. Training regimes, empirical results, and technical significance

The 2021 formulation trains on unlabeled human-transcriptome snippets of lengths 32–512 nts, with 1.15M train and 20K test sequences, and on labeled RNAcompete-S data of length 40 nts across seven RBP datasets, each containing 500K positives versus 500K negatives with an 80/20 train/test split (Yan et al., 2021). Its reported hyperparameters include latent dimension $\mathcal L(\hat G_i)\in\{S,H,I,M\}$7, hidden units $\mathcal L(\hat G_i)\in\{S,H,I,M\}$8 in all encoders and decoders, G-MPNN message-passing $\mathcal L(\hat G_i)\in\{S,H,I,M\}$9, T-GRU tree passes $4\times 3=12$0, learning rate $4\times 3=12$1 with AMSGrad, and batch size $4\times 3=12$2. KL annealing uses 5 warm-up epochs with $4\times 3=12$3, followed by a linear increase over the next 6–10 epochs up to $4\times 3=12$4, or $4\times 3=12$5 for HierVAE.

On unsupervised generation, posterior decoding uses 20K test RNAs with 5 draws each, and prior decoding uses 10K latent samples with 10 draws each. With structural constraints, HierVAE attains Validity $4\times 3=12$6, FE DEV $4\times 3=12$7, Normed FE DEV $4\times 3=12$8, and Diversity $4\times 3=12$9; even unconstrained decoding gives Validity $x_i\in\{A,C,G,U\}$00 and FE DEV $x_i\in\{A,C,G,U\}$01. By contrast, LSTMVAE and GraphVAE both yield lower validity, around $x_i\in\{A,C,G,U\}$02, if unconstrained. In the semi-supervised setting, embedding AUROC is approximately $x_i\in\{A,C,G,U\}$03–$x_i\in\{A,C,G,U\}$04, posterior-constrained sampling is $x_i\in\{A,C,G,U\}$05 valid, FE DEV is approximately $x_i\in\{A,C,G,U\}$06, and reconstruction is approximately $x_i\in\{A,C,G,U\}$07–$x_i\in\{A,C,G,U\}$08. For targeted design, starting from 10K negative RNAs and optimizing for 15–30 steps yields success rates of approximately $x_i\in\{A,C,G,U\}$09–$x_i\in\{A,C,G,U\}$10 and mean $x_i\in\{A,C,G,U\}$11 of approximately $x_i\in\{A,C,G,U\}$12–$x_i\in\{A,C,G,U\}$13.

The 2025 formulation trains on three real mRNA datasets: Zebrafish 5′ UTR with approximately 55k sequences of length $x_i\in\{A,C,G,U\}$14 and target translation efficiency; OpenVaccine with approximately 2.4k sequences of length $x_i\in\{A,C,G,U\}$15 and target mRNA stability; and a ribosome-loading dataset with approximately 260k sequences of length $x_i\in\{A,C,G,U\}$16 and target mean ribosome load (Liao et al., 14 Oct 2025). OAE training uses AdamW with base learning rate $x_i\in\{A,C,G,U\}$17, warm-up from $x_i\in\{A,C,G,U\}$18, cosine decay, batch size $x_i\in\{A,C,G,U\}$19, up to 200 epochs, and early stopping after 20 epochs; projector training uses AdamW with learning rate $x_i\in\{A,C,G,U\}$20, batch size $x_i\in\{A,C,G,U\}$21, up to 200 epochs, and early stopping. The Langevin procedure typically uses $x_i\in\{A,C,G,U\}$22 steps, $x_i\in\{A,C,G,U\}$23, and $x_i\in\{A,C,G,U\}$24. SUGAR upsampling ratios are Zebrafish $x_i\in\{A,C,G,U\}$25, OpenVaccine $x_i\in\{A,C,G,U\}$26, and Ribosome-loading $x_i\in\{A,C,G,U\}$27.

Across 5 seeds, the optimization results are reported as follows: for Zebrafish, $x_i\in\{A,C,G,U\}$28 with $x_i\in\{A,C,G,U\}$29 success and $x_i\in\{A,C,G,U\}$30 with $x_i\in\{A,C,G,U\}$31 success; for OpenVaccine, $x_i\in\{A,C,G,U\}$32 with $x_i\in\{A,C,G,U\}$33 success and $x_i\in\{A,C,G,U\}$34 with $x_i\in\{A,C,G,U\}$35 success; for Ribosome-load, $x_i\in\{A,C,G,U\}$36 with $x_i\in\{A,C,G,U\}$37 success and $x_i\in\{A,C,G,U\}$38 with $x_i\in\{A,C,G,U\}$39 success. Inference cost is approximately $x_i\in\{A,C,G,U\}$40 ms/sample on A100, compared with $x_i\in\{A,C,G,U\}$41 ms for classic gradient ascent. Manifold fidelity, measured as average $x_i\in\{A,C,G,U\}$42 to the test set, is approximately $x_i\in\{A,C,G,U\}$43, compared with about $x_i\in\{A,C,G,U\}$44 for VAE and about $x_i\in\{A,C,G,U\}$45 for MCMC. Ablations show that removing $x_i\in\{A,C,G,U\}$46 causes property changes to collapse to approximately $x_i\in\{A,C,G,U\}$47, success to about $x_i\in\{A,C,G,U\}$48, and manifold distance to triple to about $x_i\in\{A,C,G,U\}$49. Trajectories in PHATE space are smooth and monotonic in predicted property, decoded intermediates fold into valid 2D and 3D structures by ViennaRNA and RhoFold, and interpolation produces a nearly linear decrease in $x_i\in\{A,C,G,U\}$50 with a corresponding increase in $x_i\in\{A,C,G,U\}$51.

Taken together, these results define RNAGenScape as a latent-manifold design paradigm rather than a single architecture. The 2021 system emphasizes structural validity, stability, and diversity in RNA secondary-structure generation, while the 2025 system emphasizes controllable mRNA optimization, interpolation, and manifold fidelity under scarce or undersampled data. A plausible implication is that the common scientific contribution lies in coupling biologically structured latent representations with constrained or projected latent traversal so that decoded sequences remain meaningful throughout generation or optimization.