RNAGenScape: Property-Guided, Optimized Generation of mRNA Sequences with Manifold Langevin Dynamics

Danqi Liao; Chen Liu; Xingzhi Sun; Dié Tang; Haochen Wang; Scott Youlten; Srikar Krishna Gopinath; Haejeong Lee

q-bio.QMOct 14, 2025

RNAGenScape: Property-Guided, Optimized Generation of mRNA Sequences with Manifold Langevin Dynamics

Authors: Danqi Liao, Chen Liu, Xingzhi Sun, Dié Tang, Haochen Wang, Scott Youlten, Srikar Krishna Gopinath, Haejeong Lee, +3 more

Organizations: Yale University

Abstract

Generating property-optimized mRNA sequences is central to applications such as vaccine design and protein replacement therapy, but remains challenging due to limited data, complex sequence-function relationships, and the narrow space of biologically viable sequences. Generative methods that drift away from the data manifold can yield sequences that fail to fold, translate poorly, or are otherwise nonfunctional. We present RNAGenScape, a property-guided manifold Langevin dynamics framework for mRNA sequence generation that operates on a learned manifold of real data. By performing iterative local optimization constrained to this manifold, RNAGenScape preserves favorable computational viability proxies, accesses reliable guidance, and discourages excursions into regions poorly supported by the training data. The framework integrates three components: (1) an autoencoder jointly trained with a property predictor to learn a property-organized latent manifold, (2) a denoising autoencoder that projects updates back onto the manifold, and (3) a property-guided Langevin dynamics procedure that performs optimization along the manifold. Across three real-world mRNA datasets spanning two orders of magnitude in size, RNAGenScape increases median oracle-predicted property gain by up to 267% and success rate by up to 31%, while maintaining good biological viability proxy measures and achieving competitive inference efficiency compared to existing generative approaches.

Figures & tables

Figure 1: Illustration of manifold Langevin dynamics for optimization on the manifold of real data.

Figure 2: Schematic of RNAGenScape . a, We encode a given input mRNA sequence into a latent manifold organized by property, and iteratively optimize the sequence along the property landscape. The manifold projector $\Psi$ encourages the optimization to stay on or near the manifold. Notably, we can decode the intermediate optimization result after every step. b, The organized latent manifold is created by jointly optimizing reconstruction and property prediction objectives. c, The manifold projector is trained to project perturbed data points back to the manifold. d, For undersampled mRNA manifolds, we use SUGAR [ 42 ] to learn key dimensions in the manifold and fill undersampled regions. e, During optimization, the manifold projector brings off-manifold points back to the manifold.

Figure 3: Overview of the experimental design. The same procedure is repeated for every mRNA dataset. a, We first fine-tune and evaluate a pre-trained oracle model that predicts target properties from mRNA sequences. b, We then train each sequence optimization method using the training and validation sets. c, The trained methods are applied to optimize mRNA sequences starting from unseen test samples. d, Finally, the oracle model is used to evaluate property improvement and success rate.

Method	$\mathcal{M}$	Property Optimization				Biological Viability Checks
		$+$ Property		$-$ Property		Initiation Interference	Initiation Context	Structural Stability	Structural Confidence
		$\Delta$ $\uparrow$	% $\uparrow$	$\Delta$ $\downarrow$	% $\uparrow$	uORF OOF AUG % $\downarrow$	Kozak Similarity % $\uparrow$	Minimum Free Energy $\downarrow$	Mean pLDDT $\uparrow$
OpenVaccine (n $\approx$ 2k) , property to optimize: stability
Data in the test set	—	—	—	—	—	$1.75$ $\pm~0.12$	$28.70$ $\pm~0.62$	$-29.82$ $\pm~0.66$	$42.18$ $\pm~0.24$
DiffAb [ 50 ]	✗	$+0.49$ $\pm~0.09$	$72.5$ $\pm~3.4$	$-0.20$ $\pm~0.16$	$60.1$ $\pm~7.3$	$1.52$ $\pm~0.55$	$22.13$ $\pm~5.68$	$-14.41$ $\pm~3.89$	$35.57$ $\pm~3.00$
IgLM [ 66 ]	✗	$+0.59$ $\pm~0.02$	$80.6$ $\pm~0.9$	$+0.43$ $\pm~0.06$	$21.4$ $\pm~3.0$	$2.01$ $\pm~0.16$	$28.04$ $\pm~2.70$	$-15.68$ $\pm~0.51$	$38.00$ $\pm~0.77$

Table 1: Our proposed RNAGenScape generates mRNA sequences that achieve strong property improvement while maintaining favorable computational viability proxies.

\mathcal{M}

denotes whether the method takes the manifold structure into consideration. In property optimization results,

\Delta

denotes the median change in property and

\%

denotes success rate (percentage of mRNAs improved). In the biological viability checks, uORF OOF AUG denote out-of-frame start codon in the upstream open reading frame, Kozak similarity is the similarity to the consensus Kozak sequence, and mean pLDDT is a structural-confidence score averaged over nucleotides.

Method	$\mathcal{M}$	Property Optimization
		$+$ Property		$-$ Property
		$\Delta$ $\uparrow$	% $\uparrow$	$\Delta$ $\downarrow$	% $\uparrow$
Zebrafish (n $\approx$ 55k) , property to optimize: translation efficiency
DiffAb [ 50 ]	✗	$+0.25$ $\pm~0.04$	$82.7$ $\pm~\phantom{0}3.0$	$+0.00$ $\pm~0.03$	$49.6$ $\pm~\phantom{0}4.4$
IgLM [ 66 ]	✗	$+0.03$ $\pm~0.00$	$56.5$ $\pm~\phantom{0}0.6$	$-0.01$ $\pm~0.00$	$52.5$ $\pm~\phantom{0}1.0$
NOS-C [ 26 ]	✗	$+0.07$ $\pm~0.02$	$77.6$ $\pm~\phantom{0}8.4$	$-0.04$ $\pm~0.04$	$67.2$ $\pm~14.8$

Table 2: Reproducing results in Table 1 using third-party checkpoints which fine-tuned UTR-LM on external data as the oracles for translation efficiency ( UTR-LM-TE ) and ribosome loading ( UTR-LM-MRL ). These oracles are not fine-tuned on our datasets. Despite potential distribution shifts that limit direct comparison of predicted property values with those in Table 1 , consistent gains suggest that our improvements are not solely due to shared predictor-oracle biases.

	DiffAb	IgLM	NOS-C	NOS-D	PCD	EM	MFM	gg-dWJS	MPGD	RNAGenScape
	[ 50 ]	[ 66 ]	[ 26 ]	[ 26 ]	[ 73 ]	[ 4 ]	[ 34 ]	[ 32 ]	[ 28 ]	(ours)
OpenVaccine (n $\approx$ 2k)	$0.424\pm 0.015$	$0.583\pm 0.005$	$0.421\pm 0.001$	$0.417\pm 0.001$	$0.546\pm 0.001$	$0.416\pm 0.002$	$0.433\pm 0.009$	$0.479\pm 0.048$	$0.412\pm 0.002$	$\textbf{0.411}\pm 0.023$
Zebrafish (n $\approx$ 55k)	$0.643\pm 0.002$	$0.642\pm 0.001$	$0.637\pm 0.004$	$0.635\pm 0.004$	$0.687\pm 0.001$	$0.653\pm 0.004$	$0.640\pm 0.003$	$0.638\pm 0.003$	$0.642\pm 0.000$	$\textbf{0.632}\pm 0.001$
Ribosome (n $\approx$ 260k)	$0.501\pm 0.001$	$0.509\pm 0.000$	$0.502\pm 0.001$	$0.497\pm 0.001$	$0.524\pm 0.000$	$0.510\pm 0.002$	$0.509\pm 0.003$	$0.504\pm 0.002$	$0.513\pm 0.001$	$\textbf{0.491}\pm 0.005$

Table 3: Mean normalized Hamming distance between generated sequences and their nearest favorable-property test-set references. Distances lie in

[0,1]

, with lower values being better.

Table 7

	OpenVaccine (n $\approx$ 2k)				Zebrafish (n $\approx$ 55k)				Ribosome-loading (n $\approx$ 260k)
	$+$ property		$-$ property		$+$ property		$-$ property		$+$ property		$-$ property
SUGAR ratio	$\Delta$ $\uparrow$	% $\uparrow$	$\Delta$ $\downarrow$	% $\uparrow$	$\Delta$ $\uparrow$	% $\uparrow$	$\Delta$ $\downarrow$	% $\uparrow$	$\Delta$ $\uparrow$	% $\uparrow$	$\Delta$ $\downarrow$	% $\uparrow$
0	$+0.44$	$73.2$	$-0.97$	86.2	$+0.14$	$54.7$	$-0.98$	$73.0$	$+0.83$	$83.7$	$-0.74$	$67.9$
0.01	$+0.08$	$56.9$	$-0.62$	$67.5$	$+0.69$	$70.7$	$-1.20$	80.0	$+0.53$	$78.3$	$-0.20$	$51.3$
0.1	$+0.20$	$58.1$	$-0.51$	$74.0$	$+0.65$	$70.8$	$-0.93$	$75.3$	$+0.49$	$69.6$	$-1.18$	$71.4$
1.0	$+\textbf{0.92}$	83.6	$-0.77$	$84.2$	$+\textbf{0.79}$	70.9	$-\textbf{1.25}$	$77.3$	$+\textbf{0.93}$	88.8	$-\textbf{1.93}$	92.8

Table 6: Ablation study on the data augmentation ratio of SUGAR.

Panel of Figure 4 — Figure 4: RNAGenScape optimization is step-efficient and remains stable over a range of optimization steps.

Figure 6: A sample latent space trajectory of RNAGenScape over 10 optimization steps. The trajectory follows a smooth path with steady improvement in the target property. a, The trajectory in the PHATE space and the corresponding 3D structures. b, Intermediate results of RNAGenScape and competing methods midway through property optimization (at the 5 th step out of 10 steps).

Figure 7: Conceptual comparison of sampling and guidance mechanisms relevant to our property-guided optimization problem. a, Unguided reverse diffusion. b, Diffusion posterior sampling [ 17 ] adds guidance to the reverse diffusion. c, Manifold Preserving Guided Diffusion (MPGD) [ 28 ] constrains guidance through a local manifold projection before reverse diffusion. d, Guided Langevin sampling in energy-based models [ 4 ] combine guidance with energy descent; contours depict the learned energy landscape. e, RNAGenScape alternates property-guided updates and a SUGAR-assisted denoising projection in the OAE latent space. The distinct projection mechanisms used by MPGD and RNAGenScape are represented in the same color since they share the same high-level concept, whereas implementation details and stochastic perturbations are omitted from the illustration.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	$\lambda_{\text{Recon}}$ / $\lambda_{\text{Pred}}$	Noise levels	Langevin (step size, temperature)
OpenVaccine	5.0 / 1.0	{0.5, 0.2, 0.1}	$1\times 10^{-2},\,1\times 10^{-3}$
Zebrafish	5.0 / 1.0	{0.5, 0.2, 0.1}	$1\times 10^{-3},\,1\times 10^{-3}$
Ribosome	5.0 / 1.0	{0.5, 0.2, 0.1}	$1\times 10^{-3},\,1\times 10^{-3}$

Appendix

Table S1: Hyperparameters used for different datasets. All runs use latent dimension

128

, denoising autoencoder manifold projector with

\mathrm{sugar}_{w}{=}1

, and

100

Langevin steps.

	Property Prediction
	Pearson Corr. $\uparrow$	Spearman Corr. $\uparrow$
OpenVaccine (n $\approx$ 2k)	0.62	0.73
Zebrafish (n $\approx$ 55k)	0.82	0.81
Ribosome (n $\approx$ 260k)	0.93	0.91

Appendix

Table S2: The UTR-LM oracle for property evaluation.

region	$N$	density	nearest-neighbor distance	grad norm	prediction variance
on-manifold, dense	500	$39.1\pm 8.6$	$6.273\pm 0.428$	$0.398\pm 0.146$	$0.127\pm 0.103$
on-manifold, sparse	500	$\phantom{0}9.1\pm 3.2$	$8.254\pm 0.669$	$0.345\pm 0.119$	$0.182\pm 0.142$
off-manifold	500	$\phantom{0}0.0\pm 0.0$	$399.3\pm 25.7$	$0.192\pm 0.130$	$0.583\pm 0.430$

Appendix

Table S3: Property-gradient analysis in the OAE latent space on the Ribosome dataset. Reported values are the within-region mean and standard deviation averaged over five random seeds. Gradient norms are the raw

\ell_{2}

norms

\|\nabla_{z}f(z)\|_{2}

of the OAE property head.

	OpenVaccine (n $\approx$ 2k)				Zebrafish (n $\approx$ 55k)				Ribosome-loading (n $\approx$ 260k)
	$+$ property		$-$ property		$+$ property		$-$ property		$+$ property		$-$ property
$\Psi$ design	$\Delta$ $\uparrow$	% $\uparrow$	$\Delta$ $\downarrow$	% $\uparrow$	$\Delta$ $\uparrow$	% $\uparrow$	$\Delta$ $\downarrow$	% $\uparrow$	$\Delta$ $\uparrow$	% $\uparrow$	$\Delta$ $\downarrow$	% $\uparrow$
$k$ NN ( $k{=}1$ )	$+0.01$	$53.2$	$+0.00$	$50.3$	$-0.22$	$45.8$	$-0.42$	$61.5$	$+0.05$	$55.5$	$-0.09$	$56.7$
$k$ NN ( $k{=}5$ )	$+0.06$	$56.2$	$+0.05$	$45.4$	$-0.22$	$46.0$	$-0.37$	$59.7$	$-1.05$	$\phantom{0}9.7$	$-1.17$	$92.4$
$k$ NN ( $k{=}10$ )	$+0.10$	$58.6$	$+0.07$	$43.1$	$-0.22$	$45.7$	$-0.29$	$58.2$	$-1.09$	$\phantom{0}8.9$	$-1.21$	$92.5$
DAE, single-step	$+\textbf{0.93}$	88.3	$-\textbf{0.79}$	87.4	$+0.24$	$58.1$	$-0.83$	$72.1$	$+0.75$	$84.4$	$-0.99$	$74.4$

Appendix

Table S4: Ablation study on the design of the manifold projector

\Psi

. A simple

k

NN retractor is competitive in some settings but does not consistently support optimization in both directions. With a denoising autoencoder (DAE) as

\Psi

, multi-step denoising outperforms single-step denoising on Zebrafish and Ribosome-loading , and stays reasonably close on OpenVaccine .

	OpenVaccine (n $\approx$ 2k)				Zebrafish (n $\approx$ 55k)				Ribosome-loading (n $\approx$ 260k)
	$+$ property		$-$ property		$+$ property		$-$ property		$+$ property		$-$ property
$\lambda_{\text{Recon}}$	$\Delta$ $\uparrow$	% $\uparrow$	$\Delta$ $\downarrow$	% $\uparrow$	$\Delta$ $\uparrow$	% $\uparrow$	$\Delta$ $\downarrow$	% $\uparrow$	$\Delta$ $\uparrow$	% $\uparrow$	$\Delta$ $\downarrow$	% $\uparrow$
0.0	$-1.29$	$16.2$	$-2.40$	$91.0$	$-0.79$	$33.9$	$-0.34$	$54.9$	$+0.05$	$54.0$	$-0.21$	$57.3$
0.1	$+1.00$	$66.8$	$-1.60$	$88.5$	$-0.11$	$51.3$	$-1.08$	$76.1$	$+0.39$	$77.5$	$-1.20$	$90.6$
0.5	$+0.93$	$87.2$	$-\textbf{2.92}$	99.4	$+0.51$	$63.7$	$-1.07$	$76.7$	$+0.85$	$88.7$	$-0.98$	$69.5$
1.0	$+\textbf{1.11}$	91.4	$-1.31$	$87.8$	$+\textbf{1.13}$	81.2	$-1.09$	$77.2$	$+0.71$	$85.4$	$-0.88$	$69.6$

Appendix

Table S5: Ablation study on the reconstruction weight

\lambda_{\text{Recon}}

. Although

\lambda_{\text{Recon}}=1.0

outperforms

\lambda_{\text{Recon}}=5.0

on OpenVaccine in both optimization directions, we decide to use

\lambda_{\text{Recon}}=5.0

across all datasets to maintain a consistent hyperparameter setting.

Method	Property Optimization
	$+$ Property		$-$ Property
	$\Delta$ $\uparrow$	% $\uparrow$	$\Delta$ $\downarrow$	% $\uparrow$
OpenVaccine (n $\approx$ 2k)
VAE [ 36 ]	$+0.23$	$56.5$	$+0.23$	$43.5$
DDPM [ 30 ]	$-0.33$	$35.1$	$-0.33$	$64.9$
LDM [ 61 ]	$-0.41$	$32.4$	$-0.41$	$67.6$

Appendix

Table S6: RNAGenScape achieves better property optimization than de novo generative methods and classic optimization algorithms that operate on our learned data manifold.

Figure S1: Latent space interpolation trajectories from 5 sources to 4 targets. Each trajectory is shown as a line fading from bright to dark in a consistent color. RNAGenScape produces smooth and coherent paths on the manifold between arbitrary input-target mRNA pairs.

Figure S2: Latent space $\ell_{2}$ distances during interpolation show smooth and monotonic transition from the source to the target. Results are averaged over all data samples.

Figure S3: More examples of latent manifold trajectories. Trajectories in PHATE space and corresponding 3D structures are shown.

Figure S4: More examples of latent manifold trajectories. Trajectories in PHATE space and corresponding 3D structures are shown.

Figure S5: More examples of latent manifold trajectories. Trajectories in PHATE space and corresponding 3D structures are shown.

Explore similar work

May 29, 2026

mRNAutilus: Multi-Objective-Guided Discrete Generation of mRNA with Optimized Therapeutic Properties

Therapeutic mRNA design requires coordinating multiple interacting sequence features across the full transcript, where codon usage, untranslated regions (UTRs), and their coupling jointly determine stability, translation efficiency, and protein expression. Here, we present mRNA generation via unrolled trajectories and informed latent updates (mRNAutilus), a framework for simultaneous codon optimization and de novo UTR design directly from sequence. mRNAutilus combines a masked discrete diffusion model trained on millions of full-length mRNAs with Monte Carlo Tree Guidance to generate Pareto-efficient sequences under multiple functional objectives, using lightweight regressors over model embeddings to predict half-life, translation efficiency, and protein abundance. Unlike recent methods that design coding sequences and UTRs separately or rely on post hoc assembly and screening, mRNAutilus generates complete transcripts in a single process optimized across properties. Across diverse targets, zero-shot mRNAs encoding P. pyralis luciferase achieve over 400-fold higher expression than wild-type and outperform commercial and machine learning-designed baselines, including zero-shot generative approaches. Zero-shot SARS-CoV-2 Spike mRNAs exceed clinically used and commercial constructs and match or surpass lab-optimized designs with improved durability. We further demonstrate generality in therapeutic settings, including prime editing (PEMax) and programmable proteome modulation, where mRNAutilus-designed constructs enhance expression of peptide-guided E3 ligases (uAbs) for beta-catenin degradation. These results establish a sequence-based, multi-objective framework for generating functional mRNAs tailored to diverse biological applications.

Sawan Patel, Sophia Tang, Yesol Kim +8

1Atom Bioworks, Inc. · Department of Computer and Information Science, University of Pennsylvania · Department of Bioengineering, University of Pennsylvania +2

Multi-Objective Optimization Diffusion Models

Ask PDF

May 2, 2026

Protein-Conditioned Multi-Objective Reinforcement Learning for Full-Length mRNA Design

Designing therapeutic messenger RNA (mRNA) requires creating full-length transcripts that carefully balance stability, translation efficiency, and immune safety. To address this challenge, we propose ProMORNA, a multi-objective generation framework that produces complete mRNA transcripts de novo directly from a target protein sequence. Our approach begins by training a BART-style encoder-decoder model on over 6 million natural protein-mRNA pairs. We then introduce Multi-Objective Group Relative Policy Optimization (MO-GRPO) to simultaneously optimize for various biological objectives in a unified way. As a case study, we evaluated ProMORNA on the widely used firefly luciferase target, excluding it from both our supervised training data and the prompt pool. The results indicate that ProMORNA improves the in silico Pareto frontier for predicted half-life and translation efficiency relative to standard supervised baselines. Additionally, it achieves higher predicted functional scores than a state-of-the-art baseline under the same evaluation pipeline. These computational findings demonstrate the feasibility of using multi-objective reinforcement learning for full-length mRNA design on unseen targets.

Zixi Shao, Tao Wang, Yibei Xiao +1

Nanjing Foreign Language School, Nanjing, China · School of Science, China Pharmaceutical University, Nanjing, China · School of Pharmacy, China Pharmaceutical University, Nanjing, China

Multi-Objective Reinforcement Learning Molecular Generation RL for Generative Models

Ask PDF

May 27, 2026

An Evolutionary Approach for Designing Stable and Highly Expressible Low-Immunogenicity Therapeutic mRNA Sequences

Messenger RNA (mRNA) sequences as therapeutics require optimized design to ensure efficient translation, structural stability, and minimal immunogenicity. This study presents a two-stage in-silico framework that integrates deep learning and evolutionary computation for rational mRNA optimization instead of existing state-of-the-art models. In the first stage, a pretrained CodonTransformer (BERT-like Large Language Model) generates biologically coherent mRNA sequences encoding the target antigen. In the second stage, a genetic algorithm (GA) evolves these candidate sequences through codon-aware crossover and synonymous mutation guided by human codon usage preferences. Fitness functions for evaluation combined translation-related metrics (CAI, tAI, codon-pair bias), mRNA structural stability (local and global MFE via RNAfold, GC content), and reduced immunogenicity (CpG/UpA motif frequency). Over successive generations (38th, 40th, and 42nd), the GA improved (achieved CAI values of 0.73 to 0.74 and tAI values of 0.63 to 0.64) CAI and tAI by over 6% and codon-pair bias is high and consistent (0.97 ) and improved ribosomal accessibility at the 5' end, with an unpaired_30 fraction reaching 0.87; Global Minimum Free Energy (MFE) converged to a balanced range of -346 to -356 kcal/mol, achieving approximately 84% base-paired structural stability, and reduced immune-stimulatory motifs - lowering the average immune penalty to 27.3 in the final generation. Linear Design produces hyper-stable transcripts (MFE < - 2000 kcal/mol) that risk translation inefficiency due to extreme rigidity, and BiLSTM-CRF focuses solely on high CAI (0.96 to 0.98) without structural constraints, our framework achieves an optimal translation-stability equilibrium, highlighting the proposed BERT-GA framework as an effective, data-driven approach for the design and optimization of in-silico mRNA sequences.

Dhawa Sang Dong, Mausam Gurung, Suraj Kandel

Department of Artificial Intelligence School of Engineering, Kathmandu University Dhulikhel, Nepal · Department of Electronics and Computer Engineering Kathmandu Engineering College, Trivubhan University Kathmandu, Nepal · Hetauda School of Management and Social Science Trivubhan University Hetauda, Nepal

Molecular Optimization Large Language Model-Guided Optimization Molecular Generation

Ask PDF