Recombinant protein production in microbial systems is well-established, yet half of these experiments have failed in the expression phase. Failures are expected for ‘difficult-to-express’ proteins, but for others, codon bias, mRNA folding, avoidance, and G+C content have been suggested to explain observed levels of protein expression. However, determining which of these is the strongest predictor is still an active area of research. We used an ensemble average of energy model for RNA to show that the accessibility of translation initiation sites outperforms other features in predicting the outcomes of 11,430 experiments of recombinant protein production in Escherichia coli . We developed TIsigner and showed that synonymous codon changes within the first nine codons are sufficient to improve the accessibility of translation initiation sites. Our software produces scores for both input and optimised sequences, so that success/failure can be predicted and prevented by PCR cloning of optimised sequences.