ProMORNA: An End-to-End Generative Framework for Full-Length mRNA Design

The gist: ProMORNA proposes an end-to-end, protein-conditioned Reinforcement Learning framework that generates complete full-length mRNA transcripts de novo directly from a target protein sequence by optimizing for multiple biological objectives simultaneously.

How it works

ProMORNA is designed as 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. Next, to stabilize exploration over the large design space, we introduce Multi-Objective Group Relative Policy Optimization (MO-GRPO). MO-GRPO standardizes relative advantages separately for each biological metric before aggregation, avoiding directly scalarizing raw scores with different scales and variances.

The framework consists of several key components that enable protein-conditioned generation and multi-objective optimization:

1. Protein-conditioned full-length mRNA generation: This involves an encoder-decoder framework that generates complete mRNA transcripts, including the 5’ UTR, CDS, and 3’ UTR, directly from a target protein sequence. The encoder processes the input amino acid sequence of the target protein while the decoder autoregressively generates the corresponding full-length mRNA sequence step by step.

2. Multi-objective reinforcement learning for mRNA design: MO-GRPO is a GRPO-based optimization method trained over a diverse set of protein prompts. MOGRPO computes relative advantages at the metric level before aggregating them, avoiding direct scalarization of raw biological scores with different scales while enabling reward-directed optimization that generalizes to held-out protein targets.

Biological Objectives and Design Constraints

The framework defines a set of K = 5 separate biological objectives to evaluate transcript quality. The metrics used are:

** Predicted Half-life (k = 1): This objective is maximized, evaluating the predicted intracellular degradation time of the mRNA molecules. It is evaluated by a frozen ridge-regression predictor pre-trained on structural features extracted via mRNABERT. The objective is maximized.**

** Predicted Translation Efficiency (k = 2): This objective is also maximized, predicting the efficiency of translating an mRNA into protein with mRNABERT. The objective is also maximized.**

** Normalized Minimum Free Energy (MFE, k = 3): This objective serves as a robust proxy for the thermodynamic stability of the secondary structure, and lower values are preferred. It is computed with LinearFold during RL rollouts. The objective serves as a robust proxy for the thermodynamic stability of the secondary structure, and lower values are preferred.**

** U-content (k = 4): This objective is minimized, monitoring the fractional proportion of Uridine (U) nucleotides within the entire sequence because U-rich sequence elements are associated with mRNA turnover and uridine-containing RNA can activate innate immune sensing pathways. The objective is minimized.**

** UTR Length Plausibility (k = 5): This objective is maximized, encouraging biologically plausible UTR lengths with Gaussianshaped priors over the 5’ and 3’ UTR lengths. The score lies in [0, 2] and is maximized when both UTR lengths match their target values.**

RL Optimization Dynamics

MO-GRPO addresses the challenge of scalarizing raw scores by standardizing each objective within a rollout group and then aggregating the standardized scores. The aggregated baseline advantage Ai for sequence i is formulated as A i = (1/qPK) sum over k=1 to K of (w k · clip r ik − µk / σk + δk, −c, c). The weights w k denote the relative importance and optimization direction of the k-th objective, where w k > 0 for maximization and w k < 0 for minimization.

To improve training stability, MO-GRPO uses three additional mechanisms:

1. A dynamic damping factor δk is introduced for each objective; when the group variance σk drops significantly (σk ≪ δk), δk smoothly pushes the normalized advantage contribution of this metric toward zero. This naturally prevents numerical instability and shifts the model’s exploration focus toward other under-optimized metrics.

2. Extreme-value clipping with bound c = 5.0 is applied to limit the influence of outlier objective values, which is especially useful when reward components are computed from heuristic or learned proxy predictors.

3. A post-centralization step is applied to derive the final token-level advantage A hat i,t = Ai − (1/G) sum over j=1 to G of Aj. This preserves the adaptive magnitude derived from the damping factor δk, allowing the advantage to decay naturally as the policy converges.

Empirical Evaluation of Optimized mRNA Candidates

Experiments on firefly luciferase demonstrate that ProMORNA improves predicted quality of generated mRNA candidates over supervised baselines and GEMORNA, particularly on the two primary functional objectives: predicted half-life and translation efficiency. The Pareto frontier analysis further shows that MO-GRPO can guide the model toward transcript candidates with better multi-objective trade-offs. ProMORNA’s frontier shifts substantially to the upper-right, demonstrating that MO-GRPO improves both objectives simultaneously rather than trading one for the other.

The comparison results and ablation study are presented in Fig. 5. ProMORNA (denoted as RL) achieves the highest scores on the two primary functional objectives: predicted half-life and translation efficiency. While GEMORNA also outperforms the purely supervised BASE and SFT models, ProMORNA builds on this by directly optimizing the full-length transcript using multi-objective reinforcement learning, rather than recombining independently generated UTR and CDS elements. Furthermore, ProMORNA significantly reduces accessible TLR sensing motifs compared to BASE and SFT, suggesting that MO-GRPO captures this safety-relevant nucleotide pattern.

Overall, these results suggest that on-policy reinforcement learning can effectively expand the search capability of full-length mRNA generative models beyond supervised sequence modeling. The current study also has important limitations, as the evaluation is based on a single held-out reporter protein and the main functional metrics are estimated by the same proxy models used during RL training. These results therefore provide evidence of in silico optimization, but they do not yet establish improved biological activity in experimental systems. Future work should focus on improving the accuracy and experimental validity of halflife and translation efficiency predictors. Extending the framework to longer transcripts and evaluating it across a broader set of held-out protein targets may further clarify its practical applicability. Finally, integrating experimental feedback from in vitro or in vivo assays would enable closed-loop optimization and provide a stronger foundation for applying ProMORNA to therapeutic mRNA design.

--- Page 26 ---

References [1] Fernando P. Polack, Stephen J. Thomas, Nicholas Kitchin, Judith Absalon, Alejandra Gurtman, Stephen Lockhart, John L. Perez, Gonzalo P´erez Marc, Edson D. Moreira, Cristiano Zerbini, et al. Safety and efficacy of the BNT162b2 mRNA Covid-19 vaccine. [2] Lindsey R. Baden, Hana M. El Sahly, Brandon Essink, Karen Kotloff, Sharon Frey, Rick Novak, David Diemert, Stephen A. Spector, Nadine Rouphael, C. Buddy Creech, et al. Efficacy and safety of the mRNA-1273 SARS-CoV-2 vaccine.

1. Improved system can perform de novo full-length mRNA design directly from a target protein sequence without requiring a wild-type mRNA template at inference time. This is achieved by using an encoder-decoder framework conditioned on the complete target protein sequence, generating the 5’ UTR, CDS, and 3’ UTR simultaneously.

2. Improved system can optimize for multiple biological objectives in a unified way using Multi-Objective Group Relative Policy Optimization (MO-GRPO). This method standardizes relative advantages separately for each biological metric before aggregation, which avoids the instability of directly scalarizing raw scores with different scales and variances.

3. Improved system can handle high-dimensional reward optimization by calculating the aggregated baseline advantage Ai using a weighted sum of standardized, clipped, and damped objective scores. The formulation guarantees that, assuming the objectives are mutually orthogonal, the standard deviation of the aggregated advantages std(A1,..., AG) remains approximately 1.

4. Improved system can incorporate biological constraints during generation by restricting decoding to synonymous codons for CDS and allowing nucleotide tokens for UTRs. This guarantees prefix-level validity, which is then validated post-hoc to ensure region boundaries, frame consistency, start and stop codons, and exact translation into the target protein.

5. Improved system can learn sequence composition regimes that move away from supervised models during on-policy optimization. The UMAP visualization suggests that MO-GRPO does not only resample near the supervised model, but also explores new sequence-composition regimes during optimization.

Designing therapeutic messenger RNA (mRNA) requires generating full-length transcripts that optimally balance intracellular stability and translation efficiency. While generative models have advanced mRNA design, existing approaches rely on region-wise generation, expensive inference-time search, or target-specific model updates. To overcome these limitations, we introduce ProMORNA, a zero-shot, end-to-end generative framework for full-length mRNA design. ProMORNA aligns the design process with modern Large Language Model training pipelines: it utilizes a BART-style architecture pre-trained on over 6 million natural protein-mRNA pairs, followed by a two-stage post-training process comprising cold-start supervised fine-tuning and multi-objective reinforcement learning. By standardizing and aggregating relative advantages at the individual metric level with domain-specific stabilization, our post-training stage directly aligns the generative policy with conflicting biological constraints while avoiding direct scalarization of raw objective scores. In silico evaluations demonstrate that ProMORNA pushes the Pareto frontier of predicted stability and translation efficiency beyond standard baselines and GEMORNA, a leading state-of-the-art method. In three independent in vitro Fluc2P experiments, Candidate 2 achieved a 12-72h reporter-activity AUC 3.33 times that of the conventional Benchmark (Holm-adjusted P=0.01285), while its estimated 1.29-fold advantage over the reported best GEMORNA candidate remained exploratory. These results support ProMORNA as a promising framework for generating functionally active full-length mRNA candidates.