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

summary

Video file (mp4)

The gist

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

In short

ProMORNA is an end-to-end framework that generates complete mRNA transcripts directly from a target protein sequence. It uses a BART model and Multi-Objective Group Relative Policy Optimization (MO-GRPO) to simultaneously optimize five biological goals, such as predicted half-life and translation efficiency, leading to better mRNA designs than existing methods.

Key concepts

Protein-conditioned full-length mRNA generation
This involves an encoder that reads the target protein sequence and a decoder that autoregressively builds the entire mRNA. This process generates all three parts of the transcript—the 5' UTR, coding sequence (CDS), and 3' UTR—all at once, directly from the protein input.
Multi-Objective Group Relative Policy Optimization (MO-GRPO)
This is an RL method designed to optimize multiple biological objectives together. It standardizes the advantages for each objective separately before combining them. This prevents instability caused by mixing raw scores with different scales, ensuring a stable and effective search across all desired metrics.
Biological Objectives (K=5)
The framework optimizes five distinct criteria: maximizing predicted half-life (stability), maximizing translation efficiency, minimizing U-content (for safety), and maximizing UTR length plausibility. These objectives guide the generation toward biologically superior transcripts.
Dynamic Damping Factor ($\delta_k$)
This mechanism adjusts the influence of each objective during training. If an objective's variance drops too low, $\delta_k$ pushes its contribution towards zero, preventing numerical instability and allowing the model to focus its exploration on other important metrics.

Terminology used across episodes

This episode discusses

The paper

ProMORNA: An End-to-End Generative Framework for Full-Length mRNA Design · Read on arXiv

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

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.

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "ProMORNA: An End-to-End Generative Framework for Full-Length mRNA Design".

Jane: Here's the gist of the paper. 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.

Tom: 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.

Jane: Next, to stabilize exploration over the large design space, we introduce Multi-Objective Group Relative Policy Optimization (MO-GRPO).

Tom: MO-GRPO standardizes relative advantages separately for each biological metric before aggregation, avoiding directly scalarizing raw scores with different scales and variances.

Title and authors: Tom: So we're looking at ProMORNA today, which is titled "ProMORNA: An End-to-End Generative Framework for Full-Length mRNA Design". Jane, what should we expect from this paper?

Jane: Well, it’s about building a system that can design complete messenger RNA transcripts directly from the amino acid sequence of a target protein. It sounds like they are tackling the whole process, not just picking out parts of the transcript.

Lu: Exactly. The core idea is to go from a protein sequence straight to mRNA, including the five' UTR, the coding sequence, and even the three' UTR <ref:2605.01513#pg3>. It addresses that massive design space problem we talked about before where you have thousands of possible combinations just for one section <ref:2605.01513#pg3>.

Meng: I wonder if this is just another encoder-decoder setup, or if they actually managed to make it work well in practice? We need to know if this is something that can actually be deployed reliably without needing massive compute resources for every design.

Lalam: I see they start by training a BART style encoder-decoder model on over six million natural protein mRNA pairs, which gives the model a strong foundation before the reinforcement learning part kicks in <ref:2605.01513#pg1>.

The paper's summary: Tom: The summary says they are using Multi-Objective Group Relative Policy Optimization, or MO-GRPO, to handle the design problem. That’s the big mechanism they introduce to manage all those different biological goals at once.

Jane: It sounds like a way to deal with the fact that mRNA design involves balancing several different things—like how long it lasts in the cell versus how well it gets translated into protein <ref:2605.01513#pg2>.

Lu: That’s right. The summary explains that combining those metrics into one single score can be unstable because they have different units and noise levels, so MO-GRPO standardizes relative advantages separately for each biological metric before aggregating them <ref:2605.01513#pg2>.

Meng: So they’re trying to avoid the problem where you just pick one objective to maximize and ignore the others because their scales are completely different? That makes sense for real-world application.

Lalam: The paper states they define K equals five separate biological objectives used to evaluate transcript quality, which are predicted half-life, translation efficiency, Normalized Minimum Free Energy, U-content, and UTR Length Plausibility <ref:2605.01513#pg2>.

The paper's improvements: Tom: The improvements they highlight are pretty technical—they focus on how they make the reinforcement learning part more stable and effective than what’s out there now.

Jane: They introduce the idea that instead of just combining raw scores, they calculate an aggregated baseline advantage Ai using a weighted sum of standardized, clipped, and damped objective scores <ref:2605.01513#pg2>.

Lu: That aggregation formula is key; they use weights w k for each objective, where a positive weight means maximize and a negative weight means minimize the objective <ref:2605.01513#pg2>.

Meng: And they mention three specific ways they keep things stable: a dynamic damping factor delta k that pushes the advantage contribution toward zero when variance drops too low, extreme-value clipping at bound c equals five point zero, and a post-centralization step to get the final token-level advantage A hat i,t <ref:2605.01513#pg2>.

Lalam: The paper notes that this formulation guarantees that assuming the objectives are mutually orthogonal, the standard deviation of the aggregated advantages sigma(Ai) remains approximately one <ref:2605.01513#pg2>.

Conclusion: Tom: So, to wrap up this discussion on ProMORNA, it seems they've put together a framework that uses protein conditioning and multi-objective reinforcement learning to design full-length mRNA transcripts de novo.

Jane: It sounds like the main result is that this method improves the in silico Pareto frontier for predicted half-life and translation efficiency when tested on firefly luciferase, which is a pretty strong functional test <ref:2605.01513#pg1>.

Lu: They show that MO-GRPO can guide the model toward candidates with better multi-objective trade-offs, shifting the frontier substantially to the upper right where both objectives are optimized simultaneously <ref:2605.01513#pg2>.

Meng: From a deployment standpoint, it’s interesting that they built this on top of existing generative models rather than trying to learn everything from scratch for every target protein. It suggests a path forward for building these kinds of tools more efficiently <ref:2605.01513#pg2>.

Lalam: I just want to add that the paper admits a limitation, saying their evaluation is based on a single held-out reporter protein, and their main functional metrics are estimated by the same proxy models used during reinforcement learning training <ref:2605.01513#pg3>.

Tom: That’s fair. So, while this is strong evidence of in silico optimization, it doesn't yet prove improved biological activity in an experimental system. We need better predictors for things like half-life and translation efficiency before we can say this translates directly to real therapeutic design <ref:2605.01513#pg3>.

Jane: Exactly. So, ProMORNA is a really sophisticated way to explore the massive design space, but it’s just one piece of the puzzle for designing actual medicines right now. That’s where we are headed next <ref:2605.01513#pg3>.

More episodes

← Home