Codon biases that predict expression in PLS models of the combined datasets.

A genetic algorithm was used to identify codon biases that best explained expression levels for the combined datasets. The algorithm evolved 888 unique codon subsets with root mean square error in cross-validation within 5% of that of the best predictive subset. These evolved subsets contained an average of 14 codons each. The codon biases for 10 amino acids were represented by at least one codon in greater than 99% the majority of the subsets. All other amino acids were represented at less than 30% in the subsets. Codon frequency data for the 10 highly represented amino acids is shown. Column 3 (��GA incl.��), frequency of inclusion of the specified codon in the 888 selected subsets. Fc, codon usage frequency per cognate amino acid, shown for a subset of naturally highly expressed E. coli genes (��HE_coli��) [5], for the entire combined dataset (��Dataset��), and for 10 highly expressed genes among the dataset (��Best variants��; see text). %AA, percent usage of the indicated amino acid in the scFv and polymerase (��Pol��) genes as well as that measured for the E. coli proteome (see Materials and Methods). ��tRNA senstitivity�� is an estimate of the sensitivity of the charged cognate tRNA supply to amino acid limitation [36].