acceptodds
Under review as a conference paper at ICLR 2027

DynCodon: Multi-objective Dynamic Learning Strategies for Robust Codon Optimization

Abstract

Existing codon optimization methods often blindly pursue theoretical extreme values for evaluation metrics. However, these extreme sequences at the upper boundary of the Top 10% distribution are like unreplicable "geniuses", lacking universal biological adaptability; in contrast, the robust regions of the Top 10% distribution are more like high-performing "high achievers", containing generalizable evolutionary patterns. This paper proposes a codon optimization model guided by a multi-objective dynamic learning strategy, called DynCodon, aiming to achieve robustness comparable to the expression levels of "high achievers". This model is based on a dual BERT encoder-decoder architecture and employs a two-stage learning paradigm: pre-training through large-scale mixed species to capture universal amino acid-codon correspondences; fine-tuning using this multi-objective dynamic learning strategy for target species to achieve synergistic optimization across multiple key metrics. Experimental results show that DynCodon effectively avoids the pitfall of over-optimization towards extreme values, achieving high-precision reconstruction of natural sequence features within the Top 10% robust regions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Codon Optimization, Bioinformatics, Transformer

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_pK6AOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DynCodon tackles codon optimization by matching nature’s Top-10% “robust region” instead of maximizing CAI/GC. A dual-BERT encoder–decoder is pre-trained on 164 organisms, then fine-tuned per species with a metric-based adaptive learning rate.On CodonTransformer’s 8-species benchmark, it reportedly lands in that region far more often than all baselines.

Critical strengths
  1. Broad experimental coverage The method is evaluated on eight organisms with different codon preferences, providing evidence of its applicability across species.
  2. Comparisons with diverse baselines The experiments include heuristic methods and recent deep learning approaches, such as CodonTransformer, LinearDesign, and GEMORNA.
Critical weaknesses
  1. Unclear optimization mechanism The paper computes metric deviations from generated discrete sequences but does not explain how gradients flow through these metrics. This makes the proposed training objective difficult to verify or reproduce.
  2. Insufficient justification for the Top 10% target The paper assumes that the top 10% of natural sequences represents a biologically robust region, but provides no direct evidence or sensitivity analysis to support this threshold.
  3. Weak ablation design The multi-objective ablation changes the fine-tuning dataset rather than directly removing the metric-based objective. Thus, it does not clearly isolate the contribution of multi-objective optimization.
Questions for authors
  1. The paper argues that existing methods suffer from over-optimization. However, what evidence shows that their extreme metric values actually lead to worse biological outcomes? Without such evidence, is the proposed problem sufficiently justified?
  2. How are gradients computed for the metric-based loss? Since CAI, GC content, and CFD are calculated from discrete generated sequences, how is the gradient of the metric deviation obtained in Equation 5?
  3. Why choose the top 10% as the threshold? Why is the Top 10% region selected? Have the authors compared different thresholds, such as Top 5%, 15%, or 20%, to evaluate the sensitivity of the method?
Reviewer_DJQQOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors claim that existing codon optimization tools sacrifice performance by pushing metrics like CAI to theoretical extremes; they hypothesize that this over-optimization leads to non-robust codon sequence recovery. Instead, the authors propose a method to recover more natural/robust codon sequences using a multi-objective dynamic learning strategy. Their model is called DynCodon. DynCodon employs an encoder-decoder architecture and two-stage training: initial pretraining to capture natural residue <-> codon correspondences across species, and then finetuning to optimize multiple objectives toward mean metric values of the natural top 10% distribution. Finally, the authors run experiments to show that DynCodon sequences fall inside the natural top 10% more than baselines.

Critical strengths

The paper addresses an important problem, codon optimization, from a promising angle: the target should be the robust region of the natural codon usage distribution rather than a metric's theoretical extreme.

Critical weaknesses
  1. I'm not convinced this method is a significant contribution as a codon optimization method. The dynamic learning phase is just modulating the learning rates, and the relative weighting between objectives is fixed by γ\gamma and λk\lambda_k in Δt\Delta_t. So it doesn't seem like DynCodon actually "dynamically balances gradients from multiple objectives".
  2. The case for why the top 10% band is the right target is made via: the "genius" and "high-achiever" analogy and citation to Tuller et al. (2010, PNAS) which from an initial read doesn't strictly support it. The case would be far more convincing if it were grounded in something measurable, such as expression or synthesis outcomes, or improved recovery of native codon sequences. Separately, I feel as though methods that over-optimize metrics like CAI etc. can be regularized via smaller loss weights on these metrics (such as tuning LinearDesign's λ\lambda). Necessary controls would therefore include a tuned LinearDesign and e.g. DynCodon with static Ψ\Psi, to isolate the contribution of the dynamic learning mechanism.
Questions for authors

See weaknesses.

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

The paper presents DynCodon, a dual-BERT encoder-decoder model for generating coding sequences from protein inputs. It combines pretraining on 164 organisms with species-specific fine-tuning, using deviations from reference statistics derived from natural Top 10% CSI sequences to define a metric penalty and adjust the learning rate. The aim is to produce sequences with properties close to the selected natural reference region. Computational experiments across eight host datasets compare DynCodon with five baselines on codon-usage statistics, sequence similarity, and predicted RNA folding energy, alongside component ablations.

Critical strengths
  1. DynCodon achieves the highest reported CSI-region coverage among the evaluated generators across all eight host datasets, which shows the approach can adapt generated sequences toward the chosen reference statistics.
  2. The choice of hosts with contrasting GC preferences gives the evaluation a useful test of species-specific adaptation, since matching the natural references for mouse and tobacco calls for different compositional behavior.
Critical weaknesses
  1. The current experiments leave open whether the observed over-optimization comes from the learned model or its decoding strategy. Even a model that accurately captures natural codon frequencies can produce extreme CSI/CFD values when greedy decoding consistently selects common codons. A comparison of decoding strategies using the same checkpoints would help clarify which part of this behavior the proposed training method addresses.
  2. Matching CAI/CSI, GC, and CFD does not establish that the model learns meaningful sequence context. Permuting synonymous codons among positions encoding the same amino acid preserves these metrics while altering local codon patterns and possible RNA structures. DTW and MFE also depend on composition, so the reported result fail to isolate an advantage from learned sequence organization.
  3. There is a tension between the mean-based penalty in Equation (2) and the goal of recovering a natural sequence distribution. The penalty encourages outputs to concentrate around reference means, including by reducing variation that already exists within the selected robust region, but in the paper this is interpreted as recovery of natural sequence distributions.
  4. The shared species-level targets raise a feasibility concern for the learning-rate schedule. Amino-acid composition constrains the metric values attainable for each protein, so some inputs may remain far from the reference means even at their best achievable synonymous encoding. In those cases, a large deviation would trigger a higher learning rate without necessarily indicating an error that further optimization can resolve.
Questions for authors
  1. How were sequences longer than the 128-token context window handled in the reported experiments? If sliding windows were used, how were their outputs joined?
  2. Does decoding explicitly restrict each position to synonymous codons and enforce the expected sequence length and termination? What fraction of generated sequences translate exactly to the input protein, and how were outputs that failed this check handled in the evaluation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.