NatCodon: protein-structure-informed de novo codon sequence design for efficient protein expression

Technical Summary: NatCodon – Protein-Structure-Informed De Novo Codon Sequence Design

1. Problem Statement

The translation of genetic information into functional proteins is a co-translational process where nascent polypeptide folding is inextricably coupled to ribosomal elongation kinetics. Current codon optimization strategies predominantly rely on frequency-based heuristics (e.g., maximizing Codon Adaptation Index [CAI] or tRNA Adaptation Index [tAI]) to accelerate peptide elongation. These approaches often decouple translation rates from the specific folding requirements of the protein, leading to misfolding, inclusion body formation, and low soluble yields in heterologous hosts. This limitation is exacerbated in the era of AI-driven de novo protein design, where novel protein sequences lack evolutionary ancestors or native coding templates, rendering traditional "look-up table" optimization insufficient. Existing deep learning models for codon design largely treat the problem as a sequence-to-sequence mapping task, discarding critical structural information and overlooking the regulatory role of translational rhythm in coordinating co-translational folding (CTF).

2. Methodology

The authors introduce NatCodon, a generative model that reframes codon design as a structure-informed kinetic problem. The methodology integrates multi-omics data (genomic sequences, proteomic context, and protein structures) to learn the "structure-codon grammar" embedded in natural translation systems.

• Data Curation: The model was trained on a non-redundant dataset of 45,126 bacterial genome-proteome-structure triplets. Protein secondary structures were discretized into 9 structural tokens (e.g., $\alpha$-helix, $\beta$-bridge, flexible loops) using the DSSP algorithm.
• Architecture: NatCodon employs a decoder-only Transformer architecture inspired by the LLaMA framework. It jointly encodes amino acid sequences and discrete structural tokens into a unified representation, utilizing Rotary Positional Embeddings (RoPE) to capture long-range dependencies.
• Biologically Constrained Decoding:
  • Codon Gate: A hard mask module ensures strict translational fidelity by filtering out non-synonymous codons at each position, guaranteeing the generated sequence translates to the correct amino acid sequence.
  • Focus-Controlled Sampling: A tunable parameter ($\phi$) modulates the sharpness of the probability distribution, allowing a balance between deterministic optimization and sequence diversity.
• Training Objective: The model minimizes a composite loss function combining negative log-likelihood (NLL), cross-entropy with frequency weighting to correct codon imbalance, and an entropy-inspired smoothing loss to improve generalization.
• Evaluation Metric (NCI): The authors introduce the NatCodon Index (NCI), a structure-aware metric that quantifies the concordance between codon usage and structural context. It integrates structure-specific adaptiveness coefficients and Jensen-Shannon divergence to measure alignment with native structure-codon patterns.

3. Key Contributions

• Structure-Informed Generative Model: NatCodon is the first model to explicitly integrate local protein structural context into the generative process of codon sequence design, moving beyond frequency-based heuristics.
• Deciphering Structure-Codon Grammar: The study provides empirical evidence and mechanistic insights into how synonymous codon usage is tuned to structural demands, specifically identifying that less frequent codons are strategically deployed at structurally sensitive sites (e.g., domain boundaries, flexible loops) to induce ribosomal pausing for proper folding.
• Novel Metric (NCI): The development of NCI offers a predictive tool for functional protein yield that outperforms traditional metrics (CAI, tAI) and supervised AI predictors (Encodon) in generalizable settings.
• Experimental Validation: The framework was validated across a panel of heterologous proteins, including difficult-to-express targets, demonstrating significant improvements in soluble protein production.

4. Results

• Cross-Species Benchmarking: On datasets spanning E. coli, human, plant, and viral proteins, NatCodon outperformed competing models (CodonTransformer, ICOR). It reduced translational kinetics divergence by 96.5–97.5% and achieved an order-of-magnitude lower structure-conditioned codon divergence (SCCD), indicating a superior preservation of native coding architecture.
• Experimental Performance: In E. coli expression experiments, NatCodon-designed sequences yielded 2- to 33-fold improvements in soluble protein compared to sequences optimized by GeneOptimizer (OP).
  • SMURF1HECT: A previously intractable protein that produced no detectable soluble protein with OP yielded 0.57–0.61 mg·ml⁻¹ with NatCodon.
  • Other Targets: Significant improvements were observed for DHOase, Luciferase, GFP, and HPPD.
• Mechanistic Insights:
  • Explainable AI (XAI): SHAP and Integrated Gradients analyses revealed that NatCodon assigns nearly equal weight to structural features and amino acid identity. It selectively uses high-tAI codons for rigid structural cores and low-tAI codons for flexible loops and domain boundaries to create necessary translational pauses.
  • Mutagenesis: Reverse mutagenesis of NatCodon-designed DHOase confirmed that increasing translation speed in loop regions reduced yield, while increasing speed in $\alpha$-helices enhanced it, validating the model's kinetic choreography.
  • Translational Ramp: NatCodon successfully recapitulated the native 5'-terminal "translational ramp" (slower initial codons), a feature often lost in other optimization methods.
• Predictive Power of NCI: NCI showed the highest correlation with experimental soluble yield and functional output (luminescence/fluorescence) across multiple protein libraries, outperforming CAI, tAI, and Encodon in generalizable scenarios.
• Compatibility with AI Structures: The framework remains robust when using AlphaFold3-predicted structures, maintaining high sequence similarity and metric concordance with designs based on experimental structures.

5. Significance and Claims

The paper claims that NatCodon establishes a principled foundation for programmable protein expression by transcending frequency-based heuristics. By embedding structural constraints directly into sequence generation, the model orchestrates translation dynamics to align with co-translational folding requirements. This approach unlocks the production of therapeutically and industrially critical targets, including de novo AI-designed sequences previously deemed inexpressible. The authors posit that the NatCodon–NCI framework shifts the paradigm from maximizing elongation speed to "kinetic choreography," offering a generalizable solution for the reverse flow of information from protein sequence to nucleic acid sequence in the absence of evolutionary templates. The tools are made freely accessible to the community to facilitate the production of complex proteins.