Of the 20 common amino acids， 18 are encoded by multiple synonymous codons. These synonymous codons are not redundant; in fact， all of codons contribute substantially to protein expression， structure and function. In this study， the codon usage pattern of genes in the E. coli was learned from the sequenced genomes of E. coli. A machine learning based method， Presyncodon was proposed to predict synonymous codon selection in E. coli based on the learned codon usage patterns of the residue in the context of the specific fragment. The predicting results indicate that Presycoden could be used to predict synonymous codon selection of the gene in the E. coli with the high accuracy. Two reporter genes (egfp and mApple) were designed with a combination of low- and high-frequency-usage codons by the method. The fluorescence intensity of eGFP and mApple expressed by the (egfp and mApple) designed by this method was about 2.3- or 1.7- folds greater than that from the genes with only high-frequency-usage codons in E. coli. Therefore， both low- and high-frequency-usage codons make positive contributions to the functional expression of the heterologous proteins. This method could be used to design synthetic genes for heterologous gene expression in biotechnology.