genewiz produced cds sequences  (Azenta)

Journal: bioRxiv

Article Title: HalluCodon enables species-specific codon optimization using multimodal language models

doi: 10.64898/2026.03.31.715573

Figure Lengend Snippet: A . Feature fusion process. Amino acid (AA) vectors and codon vectors are extracted from protein and CDS sequences using the protein language model ESM2 (650M) and the RNA language model mRNA-FM, respectively. The two sets of vectors are combined by weighted summation to generate merged vectors representing integrated protein and CDS features. B . Training of the CodonNAT module. CDSs ranked in the top 10% by the codon similarity index (CSI) within the host genome are first selected. A masked language modeling task is then constructed by labeling 15% of codons to learn host-specific codon context features. A Naturalness score is defined as the geometric mean of the predicted probabilities for the labeled codons and is used to quantify the compatibility between the input CDS and the codon context of the host genome. C . Training of the CodonEXP module. Each CDS is assigned a binary label indicating High or Low protein expression. CodonEXP predicts the probability of high protein expression given a CDS. A Fitness score is defined as the product of the Naturalness score and the predicted probability of high protein expression and is used to select the optimal CDS. D . Two-step generation of optimized CDSs. In the CDS initialization step, the CDS corresponding to the input protein sequence is fully masked at the codon level, after which CodonNAT predicts the most probable synonymous codon for each AA. This step, termed CodonIni, produces an initial CDS for further optimization. In the second step, the CDS is optimized using either a genetic algorithm (CodonGa) or hallucination design (CodonHa). Optimization proceeds through multiple iterations until the Fitness score reaches a plateau, resulting in a final CDS that balances compatibility with host codon usage and a high probability of protein expression.

Article Snippet: Genewiz produced CDS sequences with an average CAI of 0.98, indicating that its algorithm also relies largely on codon usage frequency.

Techniques: Construct, Labeling, Expressing, Sequencing

Journal: bioRxiv

Article Title: HalluCodon enables species-specific codon optimization using multimodal language models

doi: 10.64898/2026.03.31.715573

Figure Lengend Snippet: a, b . Fitness scores increase and reach plateaus during iterative optimization of the DsRed2 CDS using the CodonGa (a) and CodonHa (b) algorithms. c, d . GC3 content increases and reaches plateaus during iterative optimization of the DsRed2 CDS using the CodonGa (c) and CodonHa (d) algorithms. Green dots represent the Fitness score or GC3 content of each optimized CDS throughout the iterative process. The red line shows the overall trend of Fitness score or GC3 content, fitted using locally weighted scatterplot smoothing (LOESS). e, f . Clustered patterns of CAI (e) and GC3 content (f) calculated from the optimized CDSs of the fifteen benchmark proteins generated by the six methods. CodonTrans denotes the CodonTransformer method. g, h . Mean Jaccard index (g) and mean sequence similarity (h) among optimized CDSs of the fifteen benchmark proteins generated by different methods.

Article Snippet: Genewiz produced CDS sequences with an average CAI of 0.98, indicating that its algorithm also relies largely on codon usage frequency.

Techniques: Generated, Sequencing