Research map / Probabilistic and scientific modeling / Molecules and proteins

$\texttt{RNAGenScape}$: property-guided, optimized generation of mRNA sequences with manifold Langevin dynamics

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative protein structure design

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RNAGenScape, a property-guided manifold Langevin dynamics framework that optimizes mRNA sequences in a property-organized latent manifold (autoencoder + property predictor, denoising-autoencoder projector, SUGAR augmentation), improving property gain and success rate over diffusion/flow baselines on three mRNA datasets.

Problem

Generating property-optimized mRNA sequences is hard due to scarce, imbalanced data, nonlinear sequence-function relationships, and the narrow space of biologically viable sequences; generative methods drifting off the data manifold yield nonfunctional sequences.

Approach

Performs iterative Langevin-style property-gradient updates in a latent space organized by property and learned by a jointly trained autoencoder/predictor, with each step projected back onto the data manifold by a denoising-autoencoder projector, augmented with SUGAR diffusion-geometry samples for sparse data.

Key results

• Up to 148% increase in median property gain (Zebrafish, +0.77 vs best baseline) and up to 30.4% higher success rate over baselines.
• On OpenVaccine, median stability gain +0.54 with 77.5% success rate vs MFM +0.27 / 62.5%.
• Throughput of 1.754 sample/ms, more than 1.6x faster than the most efficient biological sequence optimization baseline.

Surprising findings

• Removing the manifold projector makes the '+property' median gain negative on OpenVaccine (-0.44) and Zebrafish gains drop sharply, showing off-manifold drift hurts optimization.
• On the Ribosome dataset, the gain over the best baseline is essentially zero for the '+property' direction (+0.0%), with IgLM achieving a higher median gain than RNAGenScape-level values in the table before the proposed method's row is compared.

Limitations

• Models only mRNA sequence without explicit structure information.
• Relies on an external oracle (UTR-LM) with moderate correlation on OpenVaccine (Pearson 0.64).
• Evaluation limited to three datasets and 5' UTR/short sequences; mentions GitHub release only in impact statement.

Experimental setup

Datasets: OpenVaccine Zebrafish 5' UTR Ribosome (mean ribosome load) 5' UTR UTR-LM pretraining corpus (480k 5' UTR sequences)

Benchmarks: OpenVaccine Zebrafish Ribosome

Metrics: median property gain (Delta) success rate (%) uORF OOF AUG % Kozak similarity % minimum free energy normalized edit distance to held-out near-optimal sequences nucleotide accuracy Pearson correlation Spearman correlation throughput (sample/ms)

Baselines: DiffAb IgLM NOS-C NOS-D MFM gg-dWJS VAE DDPM LDM flow matching gradient ascent MCMC

Builds on: SUGAR Langevin dynamics Riemannian SGD denoising autoencoder MFM

Closest papers by meaning

• HELM: Hierarchical Encoding for mRNA Language Modeling (ICLR 2025 · similarity 0.86)
• Beyond Sequence: Impact of Geometric Context for RNA Property Prediction (ICLR 2025 · similarity 0.85)
• BEACON: Benchmark for Comprehensive RNA Tasks and Language Models (NeurIPS 2024 · similarity 0.85)
• gRNAde: Geometric Deep Learning for 3D RNA inverse design (ICLR 2025 · similarity 0.85)
• Training Compute-Optimal Protein Language Models (NeurIPS 2024 · similarity 0.84)
• BAnG: Bidirectional Anchored Generation for Conditional RNA Design (ICML 2025 · similarity 0.84)