acceptodds
Under review as a conference paper at ICLR 2027

Predictor-Guided Latent Space Codon Optimization for Maximizing Protein Expression

Abstract

Codon optimization, the process of selecting synonymous codons to improve mRNA translation efficiency and protein expression, is central to therapeutic protein production and mRNA vaccines, yet it remains a hard problem. The design space is discrete and combinatorially large, precluding gradient-based methods, and existing tools rely on heuristic proxies (e.g., Codon Adaptation Index or GC-content) that poorly capture true protein expression. We introduce Latent-Space Codon Optimization (LSCO), which recasts this discrete problem as a continuous one by mapping sequences into the latent space of a pretrained mRNA language model, enabling efficient gradient-based search. LSCO combines four components: a data-driven expression objective from an uncertainty-aware deep ensemble, whose epistemic uncertainty steers the search away from out-of-distribution designs; a Minimum-Free-Energy regularizer for structural stability; a naturalness prior from a protein-to-codon back-translation model; and constrained decoding for protein fidelity. On a real-world, wet-lab antibody expression dataset, LSCO attains the highest predicted expression among the compared methods under its native predictor and two evaluators held out from optimization, outperforming naturalness-driven and deep generative RNA foundation models, while retaining suitable biophysical properties. Notably, these gains do not come from reproducing conventional codon-bias heuristics. Ablations further show that search in the pretrained latent space outperforms direct input-space optimization, both on the guiding expression surrogate and under held-out evaluators.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Latent-Space Optimization, Codon Optimization, mRNA Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_sahqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors introduce Latent-Space Codon Optimization (LSCO), which performs synonymous codon design through gradient ascent in a pretrained variational masked autoencoder. They pretrain the backbone on over 15 million antibody transcripts and combine a ten-member expression-predictor ensemble with a ViennaRNA-trained minimum-free-energy predictor and a penalty for departing from a CodonBERT reference. The expression objective rewards predicted A280, absorbance at 280 nm, while penalizing ensemble uncertainty. Codon masks preserve the encoded protein, and temperature annealing converts relaxed codon distributions into discrete sequences. The authors evaluated 54 antibodies spanning nine antigens using the guiding predictor, two evaluators withheld from optimization, and sequence and folding metrics. LSCO achieved higher predicted expression than several codon-generation baselines, while MFE regularization improved folding energies at some cost to predicted expression. An ablation against direct codon-logit optimization showed an advantage under the guiding predictor but no statistically significant improvement under either held-out evaluator. However, the redesigned sequences were not experimentally tested.

Critical strengths
  • I appreciate that the authors enforce synonymous codon choices during optimization, which preserves the protein sequence by construction and keeps the search focused on the intended design problem.
  • The two held-out evaluators, including a structure-based GNN, provide useful checks on whether improvements transfer across predictors.
  • The direct input-space optimization baseline is particularly useful for assessing the need for latent search. The authors also examine MFE regularization and removal of the naturalness prior, allowing readers to assess which components contribute to the reported results.
Critical weaknesses
  • My main concern is the severely limited methodological novelty. Latent gradient optimization, ensemble uncertainty penalties, reference regularization, and constrained decoding are familiar components. I don't yet see the advance over methods such as LaMBO and NOS/LaMBO-2 beyond their adaptation to codon design. Table 2 also shows no significant advantage over direct input-space optimization on either held-out evaluator, weakening the central argument for the learned latent space.
  • The test case is too limited to establish practical usefulness. The authors optimize 54 selected antibodies with moderate starting expression, and all reported improvements for the redesigned sequences come from predictors. Even a small prospective comparison against the original constructs and direct input-space optimization would make the application substantially more convincing.
  • I also doubt whether the expression predictors have learned the effects needed for synonymous optimization. The reported test Spearman correlation of 0.65 does not establish that the model can rank synonymous variants of the same protein. Differences between antibody amino acid sequences could contribute substantially to predictive performance. Agreement between evaluator architectures and low ensemble disagreement do not resolve my concern.
Questions for authors
  1. What specific algorithmic contribution distinguishes LSCO from existing latent optimization methods? Please add a comparison with an established latent optimizer and simple synonymous hill-climbing under shared scoring objectives, and report runtime and multiple seeds across antibodies alongside DISO.

  2. How many expression measurements and unique proteins train each evaluator, and how are the splits constructed? Were all 54 starting antibodies excluded from every evaluator's training set? Please also specify which data were used to select optimization hyperparameters.

  3. Can the expression predictors rank experimentally measured synonymous variants of the same protein? Perhaps the authors can use a public benchmark reporting within-protein rank correlation and recovery of high-expression variants -- that wou;d directly test this capability and provide a reproducible evaluation beyond the proprietary antibody dataset.

  4. Do the authors have any wet-lab experimental data comparing LSCO, DISO, a strong codon-design baseline, and the original constructs on a small antibody panel? You could measure secreted antibody titer under matched expression conditions and report paired improvements across antibodies.

  5. In UΨ=μΨ−λLCBσΨU_\Psi=\mu_\Psi-\lambda_{\mathrm{LCB}}\sigma_\Psi, the utility subtracts uncertainty σΨ\sigma_\Psi, weighted by λLCB\lambda_{\mathrm{LCB}}, from mean predicted expression μΨ\mu_\Psi. Does σΨ\sigma_\Psi represent disagreement between ensemble means or the total uncertainty in Eq. 8, which also includes measurement noise? Please clarify and test whether uncertainty predicts actual error on held-out measured variants.

  6. How much predicted improvement survives conversion from soft codon probabilities to the final discrete sequence? Please report this gap across optimization trajectories and clarify whether the final decoding temperature is 0.01, as in Appendix A, or 0.03, as in Table 5.

  7. Eq. 4 equates the optimized latent code zz with E(c)E(c), the encoder representation of sequence cc. Why should these remain equal after latent optimization and decoding? Please clarify which representation the naturalness penalty uses and whether the distance is squared, as stated in Appendix C.

1 comment
AuthorsOpenReview ↗

Response to Reviewer sahq

We thank the reviewer for their detailed assessment and for recognizing the value of synonymous constraints, held-out evaluation, and the direct input-space baseline. We appreciate the specific questions, which help us clarify the contribution, implementation, and limitations of the evidence.

1) Contribution and latent-space advantage

Our contribution is a framework combining pretrained mRNA representations, uncertainty-aware expression guidance, structural regularization, and protein-preserving decoding for codon optimization, rather than a new general-purpose latent optimizer. We will clarify this positioning relative to such optimizers better, i.e., LaMBO and NOS/LaMBO-2.

We acknowledge that LSCO’s statistically significant advantage over direct input-space optimization is confined to the guiding predictor; no statistically significant advantage was established under either held-out evaluator. We will temper the claims about the benefit of latent-space search accordingly. The current comparisons also do not isolate the respective contributions of the pretrained representation, optimization procedure, and learned expression predictor.

2) Training data and evaluation setup

All three evaluator ensembles were trained on the same supervised dataset, containing 2,298 expression measurements. The two external evaluators were withheld from optimization, but were not trained on independent expression datasets. We constructed six folds by first clustering mRNA sequences and then stratifying fold assignments over these clusters. Optimization hyperparameters were tuned on held-out fold. We will clarify these details and the resulting limitations in the manuscript.

3) Predictive validity and experimental validation

We acknowledge that the held-out evaluators only check whether predicted improvements extend beyond the guiding predictor. We observe evidence of transfer, although gains vary across evaluators and shared biases cannot be excluded. Importantly, the experimental measurements were used to train and evaluate the predictors; and LSCO-generated designs have not yet been wet-lab tested. We will clarify this distinction throughout the manuscript and stress that the results are improvements in predicted expression, rather than experimentally established expression gains.

4) Uncertainty penalty

We mainly follow Lakshminarayanan et al. (2017) uncertainty decomposition. Each probabilistic member of the deep ensemble outputs a mean and aleatoric variance. Epistemic uncertainty is estimated as the disagreement between the members’ mean (variance of the means), while aleatoric uncertainty is estimated as mean of the members’ variances. Our objective penalizes epistemic standard deviation, rather than total uncertainty. We will clarify this distinction and the different variance conventions used during optimization and evaluation.

5) Discretization and temperature

The temperature is linearly annealed from 1.0 toward 0.03 over 3,000 optimization steps. We will clarify this schedule and correct the conflicting values. The objective is evaluated on soft codon probabilities during optimization, whereas the reported results are obtained by rescoring the selected discrete, argmax-decoded sequences. Thus, the reported gains are evaluated on discrete designs, although the soft-to-discrete score gap has not been quantified.

6) Naturalness penalty

Thank you for pointing out the inconsistency in Equation 4. The implemented penalty uses squared Euclidean distance, averaged over positions, between the optimized latent variable and the fixed encoding of the BERT-natural reference sequence. This reference is computed once before optimization. The penalty does not involve re-encoding the decoded candidate or assume that its encoding equals the optimized latent variable. We will correct Equation 4, align the main text and Appendix C, and specify the coefficient for each variant, including when the penalty is disabled.

We thank the reviewer again for these constructive comments. They will help us describe the implementation more precisely, clarify the methodological contribution, and align the claims with the available evidence.  

Kind regards,
The Authors

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

The paper introduces Latent-Space Codon Optimization (LSCO), an mRNA design framework that performs gradient-based optimization in the latent space of a pretrained mRNA language model. Optimization is guided by learned predictors of protein expression and RNA structural stability, with uncertainty-aware guidance and a naturalness regularizer that encourages the optimized latent representations to remain close to a reference generated by a pretrained protein-to-codon back translation model. LSCO achieves the highest predicted expression under its internal evaluator and competitive performance under two external evaluators.

Critical strengths
  1. Consistently outperforms the compared methods on the metric evaluated by the internal, external predictors.

  2. Using the protein-to-codon back translation model to encourage generated samples for naturalness prior is interesting.

  3. Well organized, generally easy to read and follow. The background on mRNA and codon optimization is good.

Critical weaknesses
  1. Novelty and missing related work. The paper overlooks prior work RNAGenScape (Liao et al., 2025, arXiv.2510.24736), which also performs property-guided mRNA optimization in a learned latent space while constraining updates to remain near the learned latent space. The conceptual overlap is a lot. The authors should discuss this work and clarify the methodological distinctions and novelty of LSCO.

  2. Limited empirical evaluation and results: Only one dataset in the evaluation, and only 54 mRNAs. The filtering logic for choosing median-expressing mAbs seems unnecessary. Despite the large improvement on the internal evaluators, the improvement is relatively much smaller on independent (external) evaluators, especially on the 1D CNN predictor. LSCO may be exploiting the shared model architecture prior between the internal and external transformer evaluators. And on the bio heuristics metrics, LSCO does not outperform the other methods.

  3. The limited advantage of LSCO compared to DISCO (direct search in input space): LSCO and DISCO do not seem to differ much in performance, especially on the external evaluators.

Questions for authors
  1. How independent are the internal and external 1D Transformer predictors? They appear to share the same architecture. Do they share training data or data splits?
  2. How similar are the generated sequences to the original sequences and to examples in the training dataset? Does the uncertainty-aware objective and back translation encourage homogeneous designs rather than diverse samples?
  3. What primarily drives the performance improvements: the pretrained latent representation, the optimization procedure, or the learned expression predictor?

I think I am willing to change my recommendation if the authors can 1) test on more different datasets, 2) use independent evaluators (other pretrained mRNA models, different architectures, etc.) and show that the LSCO maintains the same improvement, and 3) discuss missing prior works such as RNAGenScape.

1 comment
AuthorsOpenReview ↗

Response to Reviewer HXDh

We thank the reviewer for their constructive assessment and for recognizing the presentation of the work and the naturalness prior. We appreciate the suggestions for strengthening positioning and evaluation. Below, we clarify the contribution and acknowledge the limitations of the current evidence.

1) Relationship to RNAGenScape and methodological contribution

Thank you for bringing RNAGenScape to our attention. We agree that it is relevant prior work and will add an explicit discussion. Both approaches use property-guided optimization in a learned latent space, and we do not claim that this general principle is new. The methodological distinction concerns how the representation is learned and how optimization is constrained. RNAGenScape jointly trains a property-organized autoencoder and uses a learned manifold projector with property-guided Langevin dynamics. LSCO instead searches within a pretrained mRNA representation, combining expression guidance, an uncertainty term, an MFE regularizer, a protein-conditioned back-translation reference, and synonymous decoding constraints that preserve the target protein, with coefficients varying across variants.

We will position LSCO more precisely as a framework for constrained, predictor-guided codon optimization, rather than as introducing property-guided latent optimization itself.

2) Evaluation scope and selection of starting antibodies

The 54 starting antibodies constitute a focused optimization panel, rather than evidence of applicability across datasets and expression systems. The selection was intended to cover multiple antigens while leaving possible room for expression improvement.

Panel selection involved two stages. We first ranked antigens by their highest measured antibody A280 and selected the nine antibodies immediately below the median. Within each antigen, we then randomly sampled six eligible mAbs as starting sequences, filtering by measured purity above 95%.

For the other experiment featured in Appendix B.3 and Figure 6, we instead grouped antibodies by ranked A280 and randomly selected five starting sequences from the lowest-A280 group.

We acknowledge that these two selection strategies do not necessarily establish performance across the full expression range and broad generalizability. We will clarify this scope.

3) Evaluator independence and smaller held-out gains

We agree that evaluators withheld from optimization are not necessarily statistically independent. Shared training data or related architectures lead to correlated errors. To clarify, we use three deep-ensemble predictors of A280 from V-region mRNA, each including 10 probabilistic ensemble members that output a mean and an aleatoric variance estimate:

  1. A 1D transformer regressor with a deep VMAE backbone (6 layers, 648 hidden dimensions), used as the guiding predictor.
  2. A shallow 1D transformer regressor with an autoencoder backbone (2+2 layers, 324 hidden dimensions, 4 heads).
  3. A GNN using Chebyshev graph convolutions on estimated RNA secondary structure, without a pretrained backbone.

The VMAE and autoencoder backbones were independently pretrained on the large-scale OAS antibody-mRNA dataset. All three ensembles were trained on the same supervised expression dataset using six-fold splits; the two external evaluators were withheld from optimization, not trained on independent expression datasets.

We acknowledge that improvements are larger under the guiding predictor and smaller under the held-out evaluators. These results provide evidence of transfer beyond the guiding predictor, but cannot rule out shared biases or establish generalizability to other datasets or expression systems. We will make these limitations explicit.

4) Sequence similarity and diversity

Thank you for raising this question. Across the 54-antibody panel, LSCO designs change approx 56–60% of the starting codons on average.

MethodCodon Hamming distk-mer cosine dist
LSCO-LCB0.5970.748
LSCO-MFE-LCB0.5590.714
ICOR0.6250.812
CodonTransformer0.4470.609
GEMORNA0.4800.622
freq0.4390.610
rand0.6560.809
VMAE-BERT0.1920.304

These results suggest that LSCO makes substantial synonymous changes to the starting sequence. However, distance from the starting sequence itself does not establish diversity across repeated designs.

5) Contribution of latent search versus direct input-space optimization

We agree that the advantage over DISO is limited under the held-out evaluators. We will temper the claim accordingly. The current comparisons do not isolate contributions of pretrained representation, optimization procedure, and learned expression predictor. We will avoid attributing the observed gains primarily to any one of these components.

We thank the reviewer again for these constructive suggestions. We will revise related material accordingly.

Kind regards,
The Authors

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

The paper studies the codon optimization problem for RNAs, i.e. selecting synonymous codons corresponding to a protein sequence, aiming to improve mRNA translation efficiency and protein expression. Toward this end they propose Latent-Space Codon Optimization (LSCO), which converts the problem from discrete to continuous by mapping sequences into the latent space of a pretrained mRNA language model, allowing efficient gradient-based search. Their approach combines an expression surrogate model trained on mRNA-protein data, an MFE regularizer, a naturalness prior, and constrained decoding that guarantees fidelity to the reference protein. They perform a real world wet lab antibody expression experiment, finding that LSCO attains the highest predicted expression among the compared methods.

Critical strengths

Empirical results show that the method improves predicted expression, including on two evaluators held out from the optimization process.

Method seems overall well designed in a way that balances between generating natural sequences, increasing predicted expression levels, MFE, and using constrained decoding, to avoid overoptimizing any particular metric.

Critical weaknesses

The increased expression is not yet validated by real world wet lab experiments. While this is understandably more challenging, I think considering that the paper's claims include that the popularly used proxy measures are not a good enough measure of true protein expression, it is hard for such claims to be convincingly established without such results. It is good that the predicted expression is done using two held out evaluators, but still not fully convincing as there could be correlations between the evaluators. Additionally, the proposed method is clearly better on the trained evaluator, but not as universally or strongly better on the other two.

Less important point but I find the claiming in the abstract possibly slightly misleading (as some readers may interpret it as having actual wet lab validation of the proposed method). Even if it is technically correct, I would prefer to avoid phrasings that can lead to such misunderstandings.

Questions for authors

See weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer EJGa

We thank the reviewer for their thoughtful assessment and for recognizing the value of the multi-objective formulation and evaluation under predictors held out from optimization. We appreciate the comments on experimental validation and evaluator dependence, which will help us clarify the scope of the evidence.

1) Experimental validation and the scope of our claims

We agree that the abstract’s wording about experimental data can give the misleading impression that LSCO-generated designs were experimentally validated. The experimental measurements were used to train and evaluate the expression predictors; the generated designs have not yet been tested in the wet lab. We will revise the abstract and relevant statements throughout the manuscript to distinguish improvements in predicted expression from experimentally measured A280 gains.

We also agree that these computational results alone cannot establish that conventional proxy measures are inadequate for improving actual protein expression. We will qualify the corresponding claims accordingly.

2) Dependence between evaluators and weaker held-out gains

The evaluators withheld from optimization provide useful checks on whether improvements extend beyond the guiding predictor, but they share supervised training data and may therefore share biases. Their agreement cannot replace experimental validation.

We also acknowledge that improvements are substantially larger under the guiding predictor and smaller under the held-out evaluators. We will make this difference explicit rather than imply uniformly strong gains across models. Our conclusion is therefore limited to evidence of transfer across predictors, with gains that vary by evaluator, rather than experimentally demonstrated gains in protein expression.

We thank the reviewer again for these constructive comments. They will help us present the computational findings more precisely and make the need for prospective experimental validation clearer.

Kind Regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.