Skip to content
rnaprog

Codon optimisation without the hype: what CAI does and does not buy you

Codon optimisation raises a number. Whether that number predicts protein yield depends on the host, the transcript and the assay — here is what the literature actually supports, and how to use a codon-optimised library without over-trusting it.

rnaprog editorial 8 min read

Every design tool will hand you a codon-optimised sequence and a CAI near 0.95. That number feels like a grade. It is not a prediction of anything you can measure at the bench.

What CAI actually measures

The codon adaptation index compares the codons in your sequence to the codon usage of a reference set of highly expressed genes in the same organism. It is a geometric mean over codons, so a single rare codon drags it down disproportionately. Values run from 0 to 1.

What it tells you: how closely your sequence resembles the codon bias of genes that are, in that host, expressed well. What it does not tell you: whether your specific transcript will express well. CAI is a proxy built from a population of genes, not a model of translation kinetics.

Where the correlation holds and where it breaks

For bacterial hosts with strong codon bias — E. coli being the classic case — the correlation between codon adaptation and yield is real but modest, and it is strongest for the first ~15 codons, where the ribosome is still committing. Past that, mRNA secondary structure near the start site usually matters more than codon choice.

In mammalian systems the picture is messier. Human codon usage is comparatively flat, so the spread between “best” and “worst” codon is small. Optimising to human bias can raise CAI substantially while changing very little about yield. What tends to matter more in mRNA therapeutics: the 5′ UTR and its Kozak context, the presence of structured elements near the start codon, uridine content (for IVT and innate sensing), and the 3′ UTR and poly(A) design.

So a construct with CAI 0.78 and a clean 5′ UTR can outperform one with CAI 0.96 and a hairpin over the AUG. We have seen this argued in both directions in the literature, which is itself the point: CAI is a weak instrument for a question it is often used to answer.

The failure modes optimisation introduces

Optimisation is not free. Three things go wrong repeatedly:

Hidden regulatory elements. Rare codons are not always noise. Some are functional — they slow translation at domain boundaries so a protein folds co-translationally. Replacing every rare codon with the most frequent one can flatten that rhythm. If your protein has domains that need to fold in order, consider keeping rare codons at domain junctions.

Cryptic splice sites and poly(A) signals. Changing codons changes the nucleotide sequence, and sometimes creates a sequence that a splicing machinery recognises or that resembles a polyadenylation signal. Good tools scan for these; many do not.

Restriction sites and repeat structure. An optimiser choosing the same optimal codon repeatedly produces identical triplets in a row. That creates homopolymer runs and repeats that synthesis vendors reject or that recombine in the vector. Any optimiser worth using breaks these up — at some cost to CAI.

What a candidate library is for

Here is the practical reason to generate six or twelve candidates rather than one: you are not trying to find the optimal sequence, you are trying to find one that survives contact with reality. Different codon choices produce different GC profiles, different local structures, different repeat patterns. All can have CAI above 0.9 and behave differently.

A ranked pool lets you look at the distribution — if candidate #1 has a 6 nt homopolymer and #2 has a GC spike at position 300, maybe #3 is the one you order. The ranking is a sorting convenience, not a verdict.

How to use the number honestly

  • Treat CAI as a filter, not a score to maximise. Anything below ~0.75 for a human construct is worth a second look; anything above 0.85 is not meaningfully better than 0.92.
  • Read it alongside GC, homopolymer runs, U content and structure near the 5′ end.
  • If you are making mRNA for a therapeutic programme, the CAI of your CDS is one line in a document that will also need 5′ UTR rationale, poly(A) strategy and modified-nucleotide decisions.
  • If you are expressing a protein in E. coli, spend your attention on the first 15 codons and on the ribosome binding site, not on pushing CAI from 0.88 to 0.97.

The honest summary: codon optimisation reliably improves synthesisability and removes obvious problems. It improves expression sometimes, in some hosts, by amounts that are hard to predict. Design a pool, order the one that looks clean, and measure.

Your first design run starts with 300 free credits

Enough for a 6-variant mRNA construct set or a 12-candidate siRNA pool, with QC and export included. No card required.

300 credits · No card required · You keep the rights to your sequences