In the genomic sequence processing domain, bias in codon usage is a well known phenomenon which expresses codon context preferences. Our method takes advantage of codon usage bias for reverse translation achievement. It provides an efficient means of finding the most probable reverse translation of an amino acid sequence. In an off-line learning step, we first scan a given genome to gather statistical information about codon context preferences. We are thus able to generate the most probable nucleotide sequence for any given protein sequence. On-line generation (i.e. reverse translation) is guided by an n-gram model that enables codon preference description. As efficiency is required, we implemented a dynamic programming algorithm. it is evaluated with the Escherichia coli genome and provided prediction results compatible with laboratory constraints.