High-Codon

Sixty-one codons encode twenty amino acids, so any protein can be written in an astronomical number of synonymous DNA sequences — roughly 10^78 for a 100-residue peptide. Which one you pick matters: codon choice shapes translation speed, ribosome occupancy, mRNA folding, and co-translational folding, and the difference between a good and a bad choice for heterologous expression can be orders of magnitude in yield. Classical optimization tools address this statistically, swapping rare codons for host-preferred ones, which ignores context and risks depleting particular tRNA pools.

High-Codon is a deep learning codon optimizer for expressing heterologous proteins in Escherichia coli, developed at Weifang University of Science and Technology and posted as a preprint in June 2025. It frames codon selection as a sequence labelling problem borrowed from natural language processing: amino acids are the tokens, synonymous codons are the tags, and a pretrained protein language model assigns each residue its optimal codon in context.

Its specific contribution is the training signal. Earlier work typically trained only on genes with a high codon adaptation index (CAI), discarding the rest of the genome and with it the negative examples. High-Codon instead trains on the full expression range and reweights the loss by measured protein abundance, so highly expressed genes pull harder on the model without lowly expressed ones being thrown away.

#Key Features

• Sequence labelling formulation: Codon choice is cast as token classification over the 61 sense codons, so the model conditions each codon on the whole protein context rather than on the local amino acid alone.
• Expression-weighted loss: Cross-entropy is scaled by measured protein abundance through a tunable factor α, letting the strength of the expression prior be dialled rather than fixed.
• Pretrained protein backbone: The model fine-tunes ProtBert from the ProtTrans suite, which was pretrained on 217 million UniRef100 sequences, giving useful residue representations without training from scratch.
• Cross-species evaluation: Testing on proteins drawn from 34 organisms checks that the model generalises across source genomes rather than memorising a narrow set.
• Interpretable attention: Attention peaks map onto structural turns when projected onto predicted structures, regions where translation speed is known to influence folding.

#Technical Details

Expression data come from the PaxDb E. coli whole-organism integrated dataset: 3,674 relative abundance values, mapped through UniProtKB to GeneIDs and matched to coding sequences from NCBI, yielding 3,650 valid genes. Sequences longer than 1,536 nucleotides were excluded to fit ProtBert's 512-residue limit, leaving 3,128 training sequences. The last transformer layer's hidden states pass through a fully connected layer into the codon label space, with softmax over biologically permissible synonymous codons for each residue.

Sweeping the weighting factor showed a clear optimum: test-set CAI rose from 0.905 with no expression weighting to a peak of 0.934 at α = 0.1, dipped to 0.919 at α = 0.5, and recovered to 0.930 at α = 1.5 — a non-monotonic pattern indicating that both too little and too much weighting hurt. Evaluated on 100 randomly selected proteins from 34 species against the commercial optimizers GenSmart and ExpOptimizer, High-Codon produced CAI values roughly 23.45% higher than GenSmart and 13.34% higher than ExpOptimizer, with error bars 15.12% and 16.63% narrower respectively. GC content clustered between 45% and 60% with a median near 55%, comparable to both baselines.

#Applications

The immediate use case is designing synthetic genes for recombinant protein production in E. coli — the workhorse system for enzymes, vaccine antigens, and antibody fragments — where a better codon assignment translates directly into yield and cost. Because the model consumes a protein sequence and emits a coding sequence, it drops into the same slot as commercial optimization tools in a gene synthesis workflow.

#Impact

High-Codon sits alongside a growing family of learned codon optimizers, including CodonTranslator, CodonFM, and SynCodonLM, and its niche is narrow but well defined: a single host, and abundance-weighted supervision rather than CAI-filtered training data. Two limitations bound how much can be read into the results. First, the evaluation is entirely in silico and rests on CAI and GC content, proxies the authors acknowledge are imperfect — no wet-lab expression measurements were reported, and validation experiments are named as future work. Second, no code or weights accompany the preprint, so the model cannot currently be run or reproduced independently.