Correspondence analysis has been frequently used for codon usage
studies but this method is often misused. Indeed, as amino acid
composition constraints play on codon usage, it is common to use tables
containing relative frequencies (or ratio of fraquencies) instead of
simple codon counts, this in a way to get rid of those amino acid
biases. The problem is that some important properties of correspondence
analysis, such as rows weighting, are lost in the process. Moreover,
the use of relative measures sometimes introduces other biases and
often diminish the quantity of information to analyze, this resulting
in interpretation errors. For instance, in the case of organisms such
as Mycoplasma genitalium or Borrelia burgdorferi, the use of relative
measures led to the conclusion that there was no translational
selection, while analyses based on codon counts show that there is
indeed a selection at that level in these bacteria. In this paper, we
expose these problems and we propose alternative strategies to
correspondence analysis for studying codon usage biases when it is
needed to remove amino acid composition effects.