acceptodds
Under review as a conference paper at ICLR 2027

mRNAWM: A Latent World Model for Multi-Objective Full-Length mRNA Design

Abstract

Designing full-length messenger RNA (mRNA) sequences requires balancing translation-related properties, RNA structure, sequence composition, and manufacturability while preserving the encoded protein. These objectives interact across the coding sequence (CDS) and untranslated regions (UTRs), meaning that locally beneficial edits do not necessarily yield globally optimal designs. We formulate full-length mRNA design as a long-horizon, multi-objective decision problem and introduce mRNAWM, a protein-conditioned latent world model for controllable mRNA optimization. mRNAWM encodes a full-length mRNA into a latent state and learns action-conditioned dynamics over protein-preserving sequence edits. By rolling these dynamics forward, the model predicts the multi-step consequences of alternative edit trajectories on mRNA properties and future utility. An uncertainty-aware model predictive controller plans over these predicted trajectories, while preference-independent dynamics allow the same world model to support different objective trade-offs. mRNAWM (Full) outperforms all external methods on seven of the eight primary metrics, while the CDS-only configuration achieves the strongest Kozak score. The full-mRNA configuration reaches a codon adaptation index (CAI) of and a full-mRNA minimum free energy per nucleotide (MFE/nt) of . It further reduces homopolymer burden by 14.5% and uridine–guanosine (UG)-rich sequence burden by 19.1% relative to the strongest baseline for each metric. These results position latent world-model planning as a practical strategy for controllable biological sequence engineering under multiple interacting objectives.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
mRNA design, world models, model predictive control, multi-objective optimization, computational biology

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_GjC1OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The authors present mRNAWM, a latent world model for optimizing full-length mRNA while preserving the encoded protein. They formulate design as a sequence of synonymous CDS edits and constrained UTR substitutions, with a weighted utility over 18 sequence and structural properties. A learned transition model predicts how edits change the latent state, while preference-conditioned policy and value heads guide uncertainty-penalized beam search. The controller executes one edit and replans, allowing subsequent decisions to account for earlier changes. During search, the authors recompute 13 sequence properties exactly and predict five ViennaRNA-derived structural properties, then rescore final candidates with ViennaRNA. For me, the main computational contribution is combining learned edit dynamics with multi-step planning and adjustable objective weights. The authors compare against six methods on RefSeq-derived targets and evaluate 97 tumor-associated antigens across three UTR contexts, reporting improvements in codon adaptation and several structural and composition metrics. Planning-horizon and preference-weight studies provide additional computational support, but the study unfortunately doesn't include experimental measurements of expression or transcript stability.

Critical strengths
  • I appreciate that users can adjust the objective weights without retraining the model, which makes it straightforward to explore tradeoffs between codon usage, sequence composition, and predicted structure while preserving the encoded protein.

  • The authors recompute 13 inexpensive sequence properties exactly during search and rescore final candidates with ViennaRNA. These are sensible choices that reduce dependence on surrogate predictions and allow the final structural scores to be checked directly, so I appreciate that.

  • The improvement from one-step to two-step planning supports looking beyond the immediate edit, and the preference experiments show that changing the weights produces corresponding changes in the designs.

Critical weaknesses
  • My main concern is the missing comparison with mRNAutilus (Patel, et al. 2026), which already performs joint CDS/UTR optimization and includes experimental validation. The authors should compare both methods on shared protein targets and report the UTR constraints and search budgets. Even though the authors may have missed this, during rebuttals, I would need this comparison to judge the advance over existing approaches.

  • The biological usefulness remains unclear. All 18 properties come from sequence calculations or ViennaRNA, and I don't think improvements in these quantities alone establish that the resulting mRNAs would perform better experimentally. Even a small reporter panel measuring protein expression and transcript decay, with matched RNA chemistry and delivery conditions, would make the paper substantially stronger. I get that it may be difficult, but any published paper (conference or journal) for mRNA generation should have experimental results to validate.

  • I am having a hard time reconciling the CDS-only Kozak results with the stated definition. For transcript xx, the score K(x)=0.51[x−3∈{A,G}]+0.51[x+4=G]K(x)=0.5\mathbf{1}[x_{-3}\in\{A,G\}]+0.5\mathbf{1}[x_{+4}=G] awards half a point for A/G at position −3-3 and half for G at +4+4 relative to the start codon. Fixed UTRs preserve the first term, and synonymous CDS edits preserve whether +4+4 is G. The increase from 0.625 for CodonTransformer to 0.770 for mRNAWM (CDS) therefore needs an explanation. Please check the scoring and confirm that the methods were evaluated on identical protein targets and UTRs.

Questions for authors
  1. Could the authors compare with mRNAutilus on the same proteins, with compatible UTR lengths and shared evaluation metrics? Please report the sampling or search budget for each method.

  2. The authors should test a small panel of designed reporter mRNAs for protein expression and transcript decay. Comparing against the starting constructs and strong baselines under matched RNA chemistry and delivery conditions would make the biological contribution much more convincing to me.

  3. Since 13 of 18 properties are recomputed exactly, what improvement comes from the learned latent transitions? Please compare against beam search that directly encodes each candidate sequence, retaining the same structural predictor, initialization, and candidate budget, and report runtime.

  4. The value target Gt=∑j=tT−1γj−trjG_t=\sum_{j=t}^{T-1}\gamma^{j-t}r_j sums utility gains rjr_j from step tt to trajectory end TT, with discount γ\gamma. These returns come from random/greedy training edits. How well do they predict returns under MPC? Please include an ablation removing the future-value contribution, βV=0\beta_V=0.

  5. In Appendix A.4.6, the uncertainty penalty includes ensemble standard deviations σR,k,i\sigma_{R,k,i} for property ii at rollout step kk. Are these set to zero for the 13 properties computed exactly? Please also report structural prediction error over one to five imagined edits and whether disagreement tracks that error.

  6. How does the CDS-only Kozak score increase from 0.625 to 0.770? Under Table 4's definition, fixed UTRs preserve position −3-3, and synonymous substitutions preserve whether position +4+4 is G. Please clarify the scoring and confirm that the comparisons use identical proteins and UTRs.

7 comments

Dear Reviewer GjC1,

Thank you for your thoughtful and constructive feedback. We appreciate your recognition of the practical relevance of controllable multi-objective mRNA design, the use of exact sequence-property evaluation, and the benefits of multi-step planning.

To address your concerns, we have conducted additional comparisons with mRNAutilus (Patel et al., 2026) on 10,000 sampled test targets (Table R4), together with new evaluations against conventional search strategies, including direct re-encoding beam search (Table R2), a paired runtime comparison, and component ablations examining the contributions of future-value estimation, policy guidance, and uncertainty regularization (Table R3). We also provide an additional analysis of planning-horizon performance and decision consistency (Table R1) and further clarify the evaluation protocol and CDS-only Kozak scoring.

The additional results show that mRNAWM outperforms mRNAutilus on six of eight reported computational metrics, achieves the best mean performance on seven of eight metrics against conventional search strategies, and offers a substantial computational advantage, with direct re-encoding beam search requiring 2.65× the mean inference time of mRNAWM. Component ablations further demonstrate that learned latent dynamics and future-value estimation make substantial contributions to overall design quality.

We hope these new results and clarifications help address your concerns. Our detailed, point-by-point responses are organized into three parts in the accompanying comments.

AuthorsOpenReview ↗

Response to Reviewer GjC1 — Part 1/3: W1–W2 & Q1–Q2 — mRNAutilus Comparison and Biological Validation

[W1 & Q1] Comparison with mRNAutilus

We thank the reviewer for highlighting mRNAutilus (Patel et al., 2026), an important related work on joint CDS/UTR optimization. To address this concern, we conducted an additional comparison on 10,000 sampled targets from our held-out test set, evaluating both methods using the same eight sequence- and structure-based metrics.

Table R4. Comparison with mRNAutilus on 10,000 sampled test targets. Best results are highlighted in bold. Ensemble diversity is evaluated by proximity to the target of 0.10.

MetricmRNAWM (Full)mRNAutilus
CAI ↑0.94250.7627
Kozak ↑0.66570.7327
Full-MFE/nt ↓−0.3960−0.3558
Ens. diversity (target 0.10)0.09860.1458
Homopolymer ↓0.64500.6686
Window GC deviation ↓0.30400.2812
uAUG density ↓0.004500.00998
UG-rich density ↓0.010560.01190

mRNAWM outperforms mRNAutilus on six of the eight evaluated metrics, including substantially higher CAI (0.9425 vs. 0.7627), more favorable full-transcript MFE/nt (−0.3960 vs. −0.3558), and ensemble diversity closer to the target of 0.10. It also achieves lower homopolymer burden, uAUG density, and UG-rich density.

These differences are consistent with the methods' distinct optimization objectives. mRNAutilus employs masked discrete diffusion with Monte Carlo Tree Guidance to optimize predicted mRNA half-life, translation efficiency, and protein abundance, whereas mRNAWM evaluates 18 sequence- and structure-derived properties and optimizes their preference-weighted combination through model predictive control. In particular, mRNAWM explicitly rewards codon adaptation and RNA folding-related properties, helping explain its advantages in CAI, MFE/nt, and ensemble diversity. Its direct consideration of sequence-composition and manufacturability properties also provides optimization pressure on homopolymer and UG-rich burdens.

Conversely, mRNAutilus achieves a higher Kozak score (0.7327 vs. 0.6657) and slightly lower window GC deviation (0.2812 vs. 0.3040). The former is consistent with its translation-efficiency-oriented guidance, which may favor initiation-related sequence contexts. The latter may reflect sequence-composition preferences learned through its generative model and functional guidance, rather than direct optimization of the window GC metric.

Overall, these results demonstrate that mRNAWM achieves stronger computational performance on most of the evaluated sequence and structural objectives, while the two methods emphasize complementary aspects of mRNA design.

[W2 & Q2] Biological relevance and experimental validation

We thank the reviewer for emphasizing the importance of experimental validation. Our work focuses on establishing a generalizable computational framework for protein-preserving, multi-objective full-length mRNA design, with particular emphasis on learned edit dynamics, long-horizon planning, and controllable optimization.

To further demonstrate the practical relevance of our approach, we evaluate zero-shot transfer on 97 tumor-associated antigens across three UTR contexts (291 design tasks), without antigen-specific fine-tuning (Section 4.6, Table 3, and Appendix B.10). These CEDAR-derived antigen targets provide a therapeutically relevant evaluation setting. As shown in Table 3, mRNAWM achieves the highest CAI (0.9458), the lowest homopolymer burden (0.5457), and improved start-region unpaired fraction relative to CodonTransformer (0.4657 vs. 0.4446), while preserving the encoded antigen sequences.

These results highlight the generalizability and potential therapeutic applicability of mRNAWM, complementing our large-scale computational benchmarks. Experimental characterization of protein expression and transcript stability represents a valuable next step toward translating these computational designs into practical therapeutic applications.

AuthorsOpenReview ↗

Response to Reviewer GjC1 — Part 2/3: Q3–Q4 — Learned Latent Dynamics, Search Efficiency, and Future-Value Ablation

[Q3] Contribution of learned latent transitions versus direct re-encoding beam search

We thank the reviewer for this important question. Since 13 of the 18 properties can be computed directly from nucleotide sequences, we agree that the contribution of learned latent dynamics should be distinguished from conventional search.

To address this concern, we conducted additional comparisons against three conventional search strategies: Greedy Search, Randomized Greedy Search, and Direct Re-encoding Beam Search.

  • Greedy Search selects the edit with the highest predicted immediate utility among policy-guided candidate proposals, without multi-step lookahead or learned future-value estimation.
  • Randomized Greedy Search uniformly samples up to KK legal edit candidates at each step and selects the one with the highest predicted immediate utility, testing randomized exploration combined with local optimization.
  • Direct Re-encoding Beam Search retains multi-step beam planning but explicitly constructs and re-encodes each edited candidate sequence, using the observation head to estimate structural properties instead of recursively propagating states through learned latent dynamics.

All methods operate under the same protein-preserving editing constraints and property definitions. The 13 inexpensive sequence-level properties are computed exactly, while structural properties are estimated using learned predictors during candidate search.

Table R2. Comparison with conventional search strategies. Best results are highlighted in bold. Ensemble diversity is evaluated by proximity to the target of 0.10.

MetricmRNAWMGreedyRandom-GreedyRe-encoding Beam
CAI ↑0.9450.93210.92530.9268
Kozak ↑0.6770.5900.5900.560
Full-MFE/nt ↓−0.401−0.3738−0.3804−0.3881
Ens. diversity (target 0.10)0.0980.10990.10590.1201
Homopolymer ↓0.6500.7380.7120.790
Window GC deviation ↓0.3140.31530.32070.3207
uAUG density ↓0.00540.00410.00330.005394
UG-rich density ↓0.01570.02160.02150.0214

mRNAWM achieves the best mean performance on seven of eight evaluated metrics, including CAI, full-transcript MFE, ensemble diversity, and several sequence-composition measures.

Importantly, mRNAWM also outperforms Direct Re-encoding Beam Search on seven of eight metrics, supporting the effectiveness of learned latent imagination for multi-step editing decisions.

Computational efficiency. We additionally conducted a paired runtime evaluation between mRNAWM and Direct Re-encoding Beam Search.

MetricmRNAWMRe-encoding BeamRatio
Mean inference time (s)74.678197.9982.65×
Median inference time (s)39.001129.734—

Direct Re-encoding Beam Search requires approximately 2.65× the mean inference time of mRNAWM. This reflects the additional cost of repeatedly encoding edited candidate sequences, whereas mRNAWM propagates imagined successor states through learned latent transitions.

Together, these results support a favorable design quality–computational efficiency trade-off from learned latent dynamics, beyond simply increasing search depth or optimizing additional properties.

Reviewer_7bBhOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes mRNAWM, a multi-objective optimization model based on latent world models and model predictive control. mRNAWM has three major components: 1) a protein-conditioned mRNA state encoder that embeds the current mRNA sequence and target protein, 2) a latent world model that predicts the next mRNA state conditioned on both the current state and the action taken, and 3) a series of value functions that predicts the properties of the next latent state. The paper shows mRNAWM generally outperforming other deep learning based mRNA design methods on various mRNA design metrics.

Critical strengths

The main strength of mRNAWM seems to be its strong empirical results, showing consistent improvement over other deep learning based methods across different mRNA design metrics.

I appreciate the authors’ background on mRNA design including UTR, CDS, etc. It well justifies the multi-objective optimization instead of just focusing on one or two mRNA properties.

Critical weaknesses

Major: One of the biggest weaknesses of this paper is the lack of justification for using a latent world model, and the overall choice to perform value/property prediction and beam search in the trained latent space, especially considering all the properties could be efficiently computed in the raw mRNA sequence space.

Most of its ablation studies test hyperparameter sensitivity: number of horizons/beams used in planning, edit budget vs. performance, and inference preference trade-off by adjusting the property weight vector, which is pretty standard in multi-objective optimization.

Although its empirical results are strong, the compared baselines are all deep learning based methods. It would be more compelling to compare against greedy search, random search with multiple trials, and beam search in the sequence space. Also, it’s not discussed in the paper why/what makes mRNAWM perform much better than these DL based methods.

To better motivate the latent world model choice, the authors need to better demonstrate the advantage of operating in the latent space rather than in raw space, and conduct ablation studies that explain where the gains are coming from.

Minor: Additionally, this paper largely uses RL, control models, and world models but never explains or provides background or preliminaries for these methods, and is thus quite hard to understand and read for a general ML audience unfamiliar with RL jargon/methodologies.

Also, it would be much more readable to include a preliminary or background section on RL/world models, and largely trim Figure 2 to make it less dense and focus on the overall architecture/framework instead of trying to illustrate each modeling detail.

Questions for authors

Why/what makes mRNAWM perform much better than these DL based methods. Is it simply optimizing more objectives? What are the advantages of running in latent space?

8 comments

Dear Reviewer 7bBh,

Thank you for your thoughtful and constructive feedback. We appreciate your recognition of the strong empirical results and the motivation for multi-objective mRNA optimization.

To address your concerns, we have conducted additional experiments comparing mRNAWM against conventional search strategies, including greedy search, random-greedy search, and re-encoding beam search (Table R2), together with a paired runtime evaluation. These results show that mRNAWM achieves the best performance on seven of eight reported metrics and that explicit re-encoding beam search incurs 2.65× the mean inference time of mRNAWM on the paired evaluation.

We also clarify the motivation and contributions of learned latent dynamics, distinguish multi-step planning from conventional multi-objective optimization, and address the need for clearer methodological explanations.

Our detailed responses are organized into three parts:

  • Part 1/3: Motivation, novelty, and the role of latent world models.
  • Part 2/3: Comparisons with conventional search strategies and computational efficiency.
  • Part 3/3: Component attribution, ablation studies, and presentation improvements.

We hope these additional results and clarifications help address your concerns.

AuthorsOpenReview ↗

Response to Reviewer 7bBh — Part 1/3: Motivation, Novelty, and Latent World Models

We thank the reviewer for raising the important question of what distinguishes our latent world-model framework from conventional mRNA optimization approaches.

Motivation and methodological novelty. The central contribution of mRNAWM is a protein-conditioned predictive world model that enables controllable, long-horizon optimization of full-length mRNA sequences. Its advantages arise from three complementary capabilities:

  1. Controllable multi-objective optimization. mRNAWM separates preference-independent sequence dynamics from preference-conditioned planning. The learned world model captures how protein-preserving edits affect mRNA states and biological properties, while the controller adapts its decisions to different objective preferences. This allows a single trained model to accommodate diverse downstream design requirements without retraining. The preference-calibration experiments (Fig. 3D) demonstrate this capability.

  2. Action-conditioned prediction and multi-step planning. Rather than evaluating isolated edits, mRNAWM learns latent transition dynamics that simulate how alternative editing actions influence future sequence states and properties. Recursive latent rollouts enable efficient comparison of candidate editing trajectories, while receding-horizon control continuously updates the plan after each executed edit. Figure 4 illustrates this process, including coordinated CDS/UTR editing, coupled property changes, and predicted planning signals.

  3. Long-term utility-aware decision-making. mRNAWM combines predicted multi-step property improvements with a learned future-value function that estimates additional gains beyond the planning horizon. This enables the controller to account for the downstream consequences of current edits rather than selecting actions solely based on immediate utility improvements.

Advantages of learned latent dynamics. The action-conditioned transition model provides a reusable predictive mechanism for evaluating alternative editing trajectories without repeatedly encoding every imagined intermediate state. Combined with exact evaluation of sequence-derived properties and learned structural-property predictions, this enables efficient exploration of coupled, multi-objective editing decisions.

Importantly, our additional experiments directly validate these methodological advantages. Replacing learned latent rollouts with re-encoding beam search or raw-space search reduces the Final Score from 3.5347 to 3.4390 and 3.4307, respectively. Furthermore, removing the future-value contribution decreases the score to 3.4314, the largest degradation among the individual component ablations. These results highlight the contributions of both learned latent dynamics and future-value-guided planning to the overall optimization performance.

Together, these capabilities distinguish mRNAWM from conventional sequence generation and search approaches by enabling reusable, preference-controllable prediction of editing consequences and planning toward long-term multi-objective gains.

AuthorsOpenReview ↗

Response to Reviewer 7bBh — Part 2/3: Conventional Search Comparisons and Efficiency (Q1)

Comparison with conventional search. To directly investigate the contribution of learned latent dynamics, we additionally compare mRNAWM against greedy search, random-greedy search, and re-encoding beam search (Table R2).

These comparisons are particularly informative because they examine alternatives to learned latent rollouts. In particular, re-encoding beam search explicitly updates candidate sequences and re-encodes successor states instead of recursively predicting them using the learned transition model.

Table R2. Comparison with conventional search strategies. Best results are highlighted in bold. Ensemble diversity is evaluated by proximity to the target of 0.10.

MetricmRNAWMGreedyRandom-GreedyRe-encoding Beam
CAI ↑0.9450.93210.92530.9268
Kozak ↑0.6770.5900.5900.560
Full-MFE/nt ↓−0.401−0.3738−0.3804−0.3881
Ens. diversity (0.10)0.0980.10990.10590.1201
Homopolymer ↓0.6500.7380.7120.790
Window GC deviation ↓0.3140.31530.32070.3207
uAUG density ↓0.00540.00410.00330.005394
UG-rich density ↓0.01570.02160.02150.0214

mRNAWM achieves the best result on seven of the eight evaluated metrics, including CAI, RNA folding energy, and several sequence-composition measures. Random-greedy search achieves the lowest uAUG density, reflecting the trade-offs inherent in multi-objective optimization.

Computational efficiency. We additionally compare inference runtime between mRNAWM and re-encoding beam search. The mean inference times are 74.678 s and 197.998 s, respectively, meaning that re-encoding beam search requires approximately 2.65× more time.

These results support the advantage of using learned latent transitions to evaluate candidate editing trajectories without repeatedly performing full state encoding during imagined rollouts.

Why does mRNAWM outperform existing generation baselines? We do not attribute the improvement solely to optimizing more properties. Existing methods differ in their optimization objectives, design spaces, and generation mechanisms. Our external baseline comparisons establish overall system performance, while the additional search comparisons more directly examine the contribution of the planning mechanism. In particular, mRNAWM combines protein-preserving joint CDS/UTR editing, preference-conditioned multi-objective evaluation, and learned dynamics for multi-step planning within a unified framework.

Reviewer_JBauOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents mRNAWM, a latent world model that optimizes mRNA sequences ( coding sequences and untranslated regions) while preserving encoded protein sequences. The model is able to predicts effects of sequences edits on design properties and plans multiple edits ahead to balance competing objectives. The authors report improvements over baselines on several computational measures of sequence quality and RNA structure. The problem is well motivated and has strong practical application and impact in biotech industry.

Critical strengths
  1. Problem formulation is elegant and it jointly optimizes coding sequences and UTR while preserving the encoded protein.

  2. Computational evaluation is broad and compares against several existing baselines and including relevant ablations

Critical weaknesses

The contribution of the world model is not really clear from the results and it's really difficult to attribute the improvement to the learned latent dynamics and the validation of the model is not sufficient.

Questions for authors
  1. Can you show the precision accuracy changes w.r.t planing depth?
  2. Can you show the performance of the method w.r.t simpler search approaches under the same initialization, configuration?
  3. Why the performance of full model is worse than CDS model on four metrics?
6 comments

Dear Reviewer JBau,

Thank you for your thoughtful and constructive feedback and for recognizing the motivation, practical relevance, and empirical strengths of our work.

To address your concerns, we provide additional analyses of planning-horizon performance and multi-step structural-property prediction accuracy (Table R1), together with new comparisons against three conventional search strategies (Table R2). We also further clarify the distinctive capabilities of our world-model planning framework and the differences between CDS-only and full-length mRNA optimization.

The additional results demonstrate that mRNAWM maintains relatively stable structural-property prediction errors across five imagined editing steps, achieves improved design quality through multi-step planning, and attains the best mean performance on seven of the eight evaluated metrics compared with conventional search strategies. These findings provide further empirical support for the effectiveness of learned latent dynamics and long-term utility-aware planning.

We hope these additional results and clarifications help address your concerns. Our detailed, point-by-point responses are organized into three parts in the accompanying comments.

AuthorsOpenReview ↗

Response to Reviewer JBau — Part 1/3: W1 & Q1 — World-Model Contribution and Planning-Depth Accuracy

[W1] Contribution and validation of the learned world model

We thank the reviewer for highlighting the importance of understanding the contribution of the world model. We would like to elaborate on three key capabilities that distinguish mRNAWM from conventional sequence generation and optimization approaches.

  1. Controllable multi-objective design. mRNAWM learns preference-independent latent dynamics and vector-valued property changes, enabling the same trained world model to accommodate diverse downstream mRNA design requirements through inference-time preference adjustment, without retraining. Figure 3D demonstrates this controllability across competing design objectives.

  2. Model-based reasoning through latent imagination. Rather than generating complete sequences directly, mRNAWM performs edit-level reasoning by imagining alternative multi-step editing trajectories and predicting their consequences before committing an edit. This allows the planner to evaluate how individual protein-preserving edits influence subsequent sequence states and coupled biological properties. Figure 4 visualizes this reasoning process through accepted edit trajectories, property changes, predicted returns, and uncertainty estimates.

  3. Long-term utility-aware planning. Unlike greedy optimization that prioritizes immediate improvements, mRNAWM evaluates cumulative utility over imagined future edits and incorporates learned continuation values to guide action selection. Its receding-horizon controller executes one selected edit and replans, enabling decisions to account for the evolving sequence state and future optimization opportunities.

These capabilities are supported by complementary experimental evidence. The planning-horizon results in Figure 3A and Table R1 show the benefit of multi-step lookahead, while the structural-property prediction results in Table R1 evaluate the reliability of imagined future states. Furthermore, the direct comparison in Table R2 shows that mRNAWM outperforms conventional search approaches, including multi-step re-encoding beam search, on seven of eight evaluated metrics.

[Q1] How does prediction accuracy change with planning depth?

Since mRNAWM performs generative sequence optimization rather than sequence classification, we evaluate planning-depth accuracy through two complementary aspects: final design quality across planning horizons and structural-property prediction accuracy across imagined rollout depths.

Table R1(a). Planning performance across different horizons. Agreement denotes the percentage of final designed sequences identical to those produced with H=5.

HorizonFinal Score ↑Agreement with H=5
H=13.51021138%
H=23.52916570%
H=33.53039178%
H=43.53079886%
H=53.531323100%
H=63.53138992%
H=103.53124388%

As shown in Figure 3A and Appendix B.6, multi-step planning improves final design quality, with the largest gains occurring between H=1 and H=2. Performance subsequently stabilizes around H=5, indicating that short-horizon lookahead captures most of the planning benefit.

To assess the predictive reliability of the learned world model, we additionally compare imagined structural-property predictions against ViennaRNA-evaluated properties of the corresponding edited sequences at rollout depths H1–H5.

Table R1(b). One- to five-step structural-property prediction accuracy, measured by mean absolute error (MAE) against ViennaRNA-evaluated properties. Lower is better.

Structural property (MAE ↓)H1H2H3H4H5
Full-transcript minimum free energy per nucleotide (Full MFE/nt)0.011330.011230.011010.010900.01082
Translation-initiation-region minimum free energy per nucleotide (Start MFE/nt)0.057590.057980.058460.058040.05846
RNA ensemble diversity per nucleotide (Ensemble diversity/nt)0.084420.077160.078930.078360.07564
Translation-initiation-region unpaired nucleotide fraction (Start unpaired fraction)0.054810.053590.054690.056780.05697

Across all four properties, prediction errors remain broadly stable from H1 to H5, without a systematic increase with rollout depth. These results indicate that recursive latent predictions maintain relatively stable property-level accuracy over the evaluated short offline editing trajectories.

Together, the results show that short multi-step planning improves final sequence quality while maintaining stable structural-property prediction errors across the evaluated rollout depths. This supports the use of learned latent imagination within a receding-horizon controller that updates its predictions after each committed edit.

AuthorsOpenReview ↗

Response to Reviewer JBau — Part 2/3: Q2 — Comparison with Conventional Search Approaches

[Q2] Comparison with simpler search approaches

We thank the reviewer for this suggestion. We conducted additional experiments comparing mRNAWM against three conventional search strategies: Greedy Search, Randomized Greedy Search, and Direct Re-encoding Beam Search, complementing the existing comparisons against six mRNA design methods in Table 1.

These baselines examine the contributions of immediate optimization, randomized exploration, and multi-step planning:

  • Greedy Search selects the edit with the highest predicted immediate utility among policy-guided candidate proposals, without multi-step lookahead or learned future-value estimation.
  • Randomized Greedy Search uniformly samples up to KK legal edit candidates at each step and selects the one with the highest predicted immediate utility. This tests randomized candidate exploration combined with local optimization, without multi-step planning.
  • Direct Re-encoding Beam Search retains multi-step beam planning but explicitly re-encodes each edited candidate sequence and predicts its structural properties using the observation head, instead of recursively propagating states through learned latent dynamics. This provides a direct comparison with the latent-imagination planning mechanism.

All methods operate under the same protein-preserving editing constraints and property definitions. The 13 inexpensive sequence-level properties are computed exactly, while structural properties are estimated using learned predictors during candidate search.

Table R2. Comparison with conventional search strategies. Best results are highlighted in bold. Ensemble diversity is evaluated by proximity to the target of 0.10.

MetricmRNAWMGreedyRandom-GreedyRe-encoding Beam
CAI ↑0.9450.93210.92530.9268
Kozak ↑0.6770.5900.5900.560
Full MFE/nt ↓−0.401−0.3738−0.3804−0.3881
Ens. diversity (0.10)0.0980.10990.10590.1201
Homopolymer ↓0.6500.7380.7120.790
Window GC deviation ↓0.3140.31530.32070.3207
uAUG density ↓0.00540.00410.00330.005394
UG-rich density ↓0.01570.02160.02150.0214

mRNAWM achieves the best mean performance on seven of the eight evaluated metrics. Compared with greedy and randomized greedy search, it achieves higher CAI, more favorable full-transcript MFE, reduced homopolymer and UG-rich burdens, and ensemble diversity closer to the target of 0.10.

Importantly, mRNAWM also outperforms Direct Re-encoding Beam Search on seven of eight metrics, including CAI (0.945 vs. 0.9268), full MFE/nt (−0.401 vs. −0.3881), ensemble diversity (0.098 vs. 0.1201), and homopolymer burden (0.650 vs. 0.790). The uAUG densities are nearly identical (0.005400 vs. 0.005394).

These comparisons highlight two complementary advantages. First, the improvements over greedy strategies support the benefit of reasoning about future edit consequences rather than optimizing immediate gains alone. Second, the improvements over direct re-encoding beam search suggest that learned latent imagination offers a more effective mechanism for evaluating multi-step editing trajectories than repeatedly re-encoding candidate sequences for direct property prediction.

Together with the planning-horizon and prediction-accuracy analyses in Part 1, these results provide complementary evidence for the effectiveness of combining learned edit dynamics with multi-step reasoning and long-term utility-aware planning.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.