Codon Usage Bias in Human vs Mouse: What Changes When You Switch Expression Host

Human and mouse codon usage tables look almost identical at first glance. Both organisms are placental mammals, both use the same 64 codons, and both have similar GC content in coding regions. So when a construct that expresses well in HEK293 cells underperforms in CHO or NIH/3T3, the codon table is rarely the first thing people check. Often, that’s exactly where the problem is.

This post compares human and mouse codon usage in the specific positions where they diverge enough to matter, walks through when re-optimization is worth doing, and offers a practical decision tree for switching hosts.

How similar are human and mouse codon tables?

The published codon usage frequencies from the Kazusa Codon Usage Database show that for the vast majority of amino acids, human and mouse preferences are within a few percentage points. Both species heavily favor GC-rich codons (consistent with mammalian genome composition), both penalize AT-rich codons, and both suppress CpG dinucleotides moderately.

The codon adaptation index (CAI) for a typical human-optimized gene runs in mouse cells anywhere from 0.85 to 0.95 of its CAI in human cells. That’s usually high enough that the gene expresses, but the difference is real and shows up in protein yield for high-expression workhorse constructs.

Where human and mouse codon usage actually diverges

The differences cluster at specific amino acids. These are the codon positions where switching hosts can change expression measurably:

Across most amino acids, the differences are 1–2 percentage points, which is well within the noise of any individual gene expression measurement. The codons where the differences accumulate matter when you have proteins with biased amino acid composition (e.g., Arg-rich nuclear localization signals, Leu-rich coiled coils, Pro-rich linkers).

When re-optimization actually matters

Most genes that work in human cells will work in mouse cells without re-optimization. The exceptions:

• Genes with extreme amino acid bias. Histones, ribosomal proteins, transcription factors with long disordered regions, and antibody constant regions all have skewed amino acid composition. Small per-codon differences amplify.
• Genes near the expression ceiling of the host. If you’re trying to express a recombinant antibody at 5 g/L in CHO, even a 5% codon penalty becomes economically significant. For research-scale expression at 1–100 mg/L, the penalty is invisible.
• Genes with mouse-incompatible regulatory elements embedded in the ORF. Some codon-optimized human sequences acquire cryptic mouse-active splice sites or microRNA target sites by accident. Switching hosts can suddenly activate (or disable) a hidden regulatory effect.
• Genes whose protein product has a different fold or activity in mouse cells. Codon usage affects translation speed, which affects co-translational folding. A protein optimized for human translational rhythm may misfold in mouse cells with different translational dynamics.

Beyond per-codon frequency: the structural rules that matter more

Per-codon frequency is the easiest thing to compare, but most expression problems caused by codon issues come from structural features of the ORF, not raw frequency:

• 5′ UTR and start-of-ORF mRNA structure. Stable secondary structure within the first 30–40 codons after ATG suppresses translation initiation. Both human and mouse ribosomes are sensitive to this. A codon-optimization tool that doesn’t check 5′ mRNA folding will produce sequences that look optimal but translate poorly.
• Codon context. Adjacent codons interact — certain dinucleotide pairs slow the ribosome regardless of codon frequency. The dinucleotides UpA and CpG in particular are suppressed in mammalian ORFs.
• Codon ramp. Many highly-expressed mammalian genes have a deliberate stretch of slower codons in the first 30–50 codons (the “ramp”), thought to space ribosomes evenly along the mRNA. Aggressive codon optimization tools sometimes flatten this ramp and reduce expression in the process.
• Splice-site avoidance. Mammalian cells will splice cryptic donor/acceptor sites in any ORF if the score is high enough. Splice-site avoidance during codon optimization is more important than perfect codon frequency matching.

These rules apply to both human and mouse hosts. A codon-optimized sequence that respects these rules will work in both species. Many of the same considerations apply when moving in the other direction — for example, when porting a mammalian gene to E. coli, where the rules of codon optimization for E. coli expression overlap with mammalian rules but include additional bacterial-specific concerns like ribosome binding site spacing and rare-codon clusters.

Decision tree: do I need to re-optimize?

Before re-ordering a synthetic gene, work through this decision tree:

1. Was the original sequence optimized for the source host or the target host? If it was optimized for human cells and you’re moving to mouse, the difference is small. If it was optimized for E. coli or yeast and you’re moving to mammalian cells, you need to re-optimize.
2. How is the protein expressing now? If it’s expressing within 2–3x of your target, codon usage is unlikely to be the bottleneck. Look at promoter strength, copy number, and 5′ UTR before re-ordering.
3. How critical is the expression level? Research-scale: don’t bother re-optimizing for human-to-mouse. Production-scale (CHO/NS0 antibody manufacturing): re-optimize for the specific cell line you’ll produce in.
4. Have you ruled out structural issues? Run the existing sequence through 5′ mRNA folding (RNAfold, mfold, ViennaRNA) and a splice-site predictor (NetGene2, Human Splicing Finder). If the sequence has stable 5′ structure or strong cryptic splice sites, re-optimization with structural awareness is more valuable than just changing the codon table.

Practical workflow for switching hosts

If you’re moving a project from human cells to mouse cells (or the reverse), here’s the order of operations that catches the most problems for the least effort:

1. Test the existing construct in the new host first. If protein levels are within 2–3x of the original, accept the loss and move on.
2. If expression drops more than 3x, check 5′ mRNA structure and cryptic splice sites in the existing sequence before re-ordering.
3. If the existing sequence has structural problems, re-design with structure-aware optimization, not just frequency-table optimization. Tools like the PlasmidStudio codon optimizer handle structure-aware optimization in addition to per-codon frequency matching.
4. If structural features are clean and expression is still low, swap the promoter or 5′ UTR before re-optimizing the codon table. Promoter and UTR swaps usually produce larger expression changes than codon swaps.
5. As a last step, re-optimize the codon table for the specific target host. The difference between human and mouse codon tables is small enough that this is rarely the highest-leverage change you can make.

The honest answer to “does codon usage differ between human and mouse?” is yes, but rarely enough to drive a re-order on its own. The bigger wins almost always come from structural sequence features, regulatory elements, and host-cell-specific 5′ UTR effects — all of which a thoughtful codon optimization workflow handles in the same pass.