← Research

Product

EvoCodon: a genome-learned codon model for Komagataella phaffii

June 28, 2026

A self-developed, genome-learned codon model trained on public K. phaffii genomes — reproducing host-specific codon rhythm beyond frequency-based methods, and competing on the same scorecard as every other candidate.

The default approach to codon optimization is frequency: count how often K. phaffii uses each synonymous codon, then pick the most frequent one at every position. This is the logic behind max-CAI optimization, and it is a known failure mode. A sequence built from maximum-frequency codons can be translationally poor — it ignores the rhythm of codon choice that real host genomes exhibit.

EvoCodon is our self-developed, genome-learned codon model for Komagataella phaffii. Trained on public K. phaffii genomes, it learns host-specific codon rhythm — the way codon choice varies along a transcript in patterns that frequency tables alone cannot capture.

Beyond frequency

Frequency-based methods collapse codon choice to a single lookup table: one preferred codon per amino acid. EvoCodon models the sequence-level structure of how a host actually uses its codons — the %MinMax rhythm, the codon-pair bias, the regional variation. The result is a candidate that resembles a host-written transcript, not a maximally-frequent one.

One candidate among equals

EvoCodon's candidate enters the same panel as every other: uniform, host-weighted, GC-balanced designs. They all compete on the same four-axis scorecard — manufacturability, host-likeness, 5′ initiation, protein developability. The learned candidate does not auto-win. Its rank reflects the data, visible alongside its competitors.

This is deliberate. A model that always crowned its own output would be persuasion, not provenance. EvoCodon earns its place by being inspectable on the same terms.

Genome-derived and interpretable

EvoCodon is trained on public K. phaffii genomes. Its host-likeness contribution is measured using codon usage and %MinMax rhythm — genome-derived signals that reject the degeneracy of naive max-CAI over-optimization. The model proposes a candidate; the scorecard judges it.

References

EvoCodon is trained on public K. phaffii genomes and positions itself against the frequency-based methods below — each a real, citable source.

  1. Sharp PM, Li WH. The codon Adaptation Index — a measure of directional synonymous codon usage bias, and its potential applications. Nucleic Acids Res, 1987. PMID:3547335
  2. Gustafsson C, et al. Codon bias and heterologous protein expression. Trends Biotechnol, 2004. PMID:15245907
  3. Zha J, et al. Advances in Metabolic Engineering of Pichia pastoris Strains as Powerful Cell Factories. J Fungi (Basel), 2023. PMID:37888283
Scope. EvoCodon produces a host-aware codon candidate. It is part of Kairos's dry-lab design ranking, not an expression or titer prediction. No absolute model accuracy figures are published — the candidate is judged by its scorecard, not its pedigree.