acceptodds
Under review as a conference paper at ICLR 2027

PACE-3D: A structure-aware multi-dimensional evaluation framework for codon optimization via AlphaFold2-derived metrics

Abstract

Codon optimization is the cornerstone of heterologous protein expression, but current metrics, such as the Codon Adaptation Index (CAI), only focus on translation elongation rates and ignore the complex coupling between translation dynamics and co-translational folding. Therefore, despite theoretically high expression levels, optimized sequences often suffer from misfolding and aggregation. Here, we propose PACE-3D (Protein-Aware Codon Evaluation in 3D), a structure-aware evaluation framework for quantifying the compatibility between mRNA sequences and protein 3D structures. Utilizing the AlphaFold2-derived confidence metrics, pLDDT (predicted Local Distance Difference Test) and PAE (Predicted Aligned Error), we develop six multidimensional metrics to assess local rigidity-velocity matching, domain assembly synchronization, and translational rhythm smoothness, and so on. Analysis of mainstream codon optimization models based on this framework shows that while existing models successfully optimize translation rate and static mRNA structural stability, they severely neglect the dynamic requirements of co-translational folding, leading to ultimate task failure. This work bridges the gap between sequence design and structural biology, providing a universal standard for rational mRNA design in the AlphaFold era.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Evaluation Framework, Codon Optimization, AlphaFold2

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_2qAHOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

PACE-3D is a structure-aware eval framework for codon optimization. The motivation is that conventional condon-optimization objectives such as CAI emphasize codon usage/translation efficiency without the explicit modelling of co-translational protein folding. PACE-3D uses AF2-derived pLDDT & PAE + codon-freq-derived translation-velocity profiles to define 6 metrics: STC, KDM, SVR, LSEVC, IMRC, and LRCWV. They are intended to measure local rigidity-translation coordination, kinetic synchronization, within-module translation coherence, responses to long-range structural constraints, etc.. PACE-3D analyses the eSOL E. coli dataset, observing that a substantial fraction of high-CAI proteins are insoluble and that longer proteins tend to have higher CAI but lower solubility (“CAI paradox”). PACE-3D was then applied to native E. coli seq and to seq generated by 5 optimization methods (BFC, URC, CodonTransformer, GEMORNA, LinearDesign) across 8 organisms. The conclusion is that current codon optimization methods fail to preserve structure-aware translational patterns, particularly for longer proteins.

Critical strengths
  1. The paper addresses an important limitation of conventional codon-optimization evaluation. Incorporating protein quality and co-translational folding rather than evaluating only codon usage, GC content, or mRNA properties is well motivated. The attempt to connect seq design with predicted protein structure could become useful if the structural proxies can be validated.
  2. The proposed metrics are generally interpretable. The evaluation has relatively broad scope, which includes the eval of multiple existing optimization strategies & organisms (rather than exclusively on E. coli).
Critical weaknesses
  1. PACE-3D explicitly reinterprets pLDDT (AF2 confidence metric) as a proxy for "local rigidity requirements" and PAE as a proxy for topology. Then the authors assume that high-pLDDT regions should be translated more slowly and that low-PAE residue pairs represent structural contacts requiring translational pauses. These are very strong assumptions about AF outputs. From my perspective, pLDDT is fundamentally a prediction-confidence quantity; high pLDDT doesn’t automatically imply that the corresponding region is rigid or requires slow synthesis. Similarly, PAE quantifies uncertainty in relative predicted positions. Low PAE is not equivalent to small physical distance between two residues. In other words, PACE-3D is too heavily relying on the beliefs from AlphaFold model. Especially, for LRCWV, the paper defines "significant contacts" using a PAE threshold and then weights by 1/PAE. This effectively treats prediction confidence as contact strength. Since AF coordinates themselves are available, I’m confused why actual pairwise distances/contact maps are not used. This concern affects STC/SVR through pLDDT and especially LSEVC/IMRC/LRCWV through PAE.
  2. Most PACE-3D metrics are constructed in a way that negative correlation between structural complexity and translation velocity is favorable. E.g., for STC, negative correlation is defined as successful coordination. Co-translational folding is likely more context dependent than "structurally difficult region -> slower codon at the same position." E.g., a pause can be beneficial after a structural element has been synthesized and emerged sufficiently from the ribosome, rather than directly at the residues constituting that element. The whole system is complex, from my perspective, and the context will be an important component if the authors want to further clarify the co-translational folding process. Similarly, the codon usage frequency is treated too directly as translation velocity. The velocity variable in the mansucript is a quantity derived from codon usage freq/codon adaptation weights. However, actual elongation process depends on more than marginal codon freq, including codon context, tRNA avaibility, local mRNA structure, ribosome density, etc. etc.. This is important because the paper criticizes CAI for being an inadequate translation-centric metric but subsequently derives its own kinetic conlusions from closely related codon-freq information.
  3. Some of the reported results contradict the proposed interpretation of the metrics. LSEVC is defined such that negative is favorable, yet the paper reports positive LSEVC for very short E. coli proteins and increasingly negative LSEVC for longer proteins. That trend seems opposite to the claim that short proteins exhibit stronger evolutionary coordination.
Questions for authors
  1. Can you provide direct validation of the central proxy assumptions?
  2. Why is PAE used to identify physical contacts instead of AlphaFold coordinates?
  3. Have you tested positional offsets between structural requirements and translation pauses?
  4. What exactly is the translation-velocity proxy? Any codon-weight definition, host-specific tables, smoothing window size, etc. of all six metrics to these choices?
Reviewer_Ho88OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the objectives and evaluation frameworks for codon optimization. Specifically, existing codon optimization aims to maximizing translation rate by matching the codon usage to the codon usage bias of the host species and improving the stability of mRNA quantified by MFE and GC content. In this paper, the authors argue that high translation speed doesn't lead to high protein quality due to the co-translational folding, and a good codon sequence should encodes a desiable "timetable" that guides protein folding. The authors find that a high CAI doesn't mean a high rate of folding (measured by the solubility) and longer sequences are more likely to have higher CAI but lower solubility due to their complex 3D structures. The authors proposes to use AlphaFold outputs, especially pLDDT and PAE , as proxy indicators of co-translational fold structure and proposed 6 metrics to evaluate if the translation dynamic of the optimized codon sequences matches the folding dynamic of the protein.

Critical strengths

The paper studies an important problem in drug discovery which is how to evaluate an optimized codon sequences before wet-lab validation. The paper proposes evaluations beyond the commonly used CAI and MFE and studies the fitness of the translation dynamic and the folding dynamic which is novel and can be potentially impactful to the whole codon optimization community. The paper is also well written.

Critical weaknesses

Although the paper studies a novel perspective of codon optimization, I'm not convinced by the important assumptions that "regions with well defined structures (rigid) need to have slow translation" implicitly used in this paper (see questions). I would be happy to increase my score if the authors address this concern during rebuttal.

Questions for authors
  1. Could the authors justify why the proposed metrics are correlated with the quality of proteins? Do those 6 metrics predict experimentally measured protein solubility (ideally within-protein synonymous codon sequences) better than other metrics, such as CAI?

  2. Could the authors justify why codon frequency is a good indicator of the translation speed and why pLDDT and PAE from AlphaFold are good proxy indicators of the folding dynamic? AlphaFold confidence metric measures the uncertainty in model prediction and a high prediction confidence doesn't directly imply that it requires slower translation.

Minor:

  1. What does the result of the first-order derivative mean in Figure 1? How is it calculated? Is it a proxy of the translation and folding dynamic?
Reviewer_92byOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces PACE-3D, which overcomes the limitations of standard codon optimization which can guarantee high yield but not necessarily the foldability of expressed protein with a structure aware evaluation framework. The framework consists of 6 quantitative metrics incorporating pLDDT and pAE values from AlphaFold-2.

Critical strengths
  • Introduces a diverse set of 6 quantitative metrics based on pLDDT and pAE for translation and foldability of a codon that is not dependent on the length of the protein.
  • The “CAI paradox” framing (high-CAI long proteins, lower eSOL solubility) is a clean, concrete motivation.
  • All six metrics given in closed form, so re-implementable in principle.
  • Solid in silico evaluation: was applied to a diverse set of 8 proteomes and 5 different generation methods.
Critical weaknesses
  • Missing mandatory “AI Use Statement” as separate from the Ethics statement
  • Framework never validated against experimental outcomes, eSOL is used only for motivation, never to test whether the 6 metrics predict solubility.
  • It is unclear whether model confidence metrics like pLDDT/pAE correlate to biophysical properties like codon optimality without experimental validation.
Questions for authors
  • Is there any experimentally correlated metric or wet-lab validation that can be performed to test whether these metrics are actually predictive of codon optimality and translation speed?
  • What are STC/SVR/LRCWV values under a synonymous-shuffle null, computed separately per length group?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.