The prediction of the network of protein-protein interactions (PPI) of an organism is crucial for the understanding of biological processes and for the development of new drugs. Machine learning methods have been successfully applied to the prediction of PPI in yeast by the integration of multiple direct and indirect biological data sources. However, experimental data are not available for most organisms. We propose here an ensemble machine learning approach for the prediction of PPI that depends solely on features independent from experimental data. We developed new estimators of the coevolution between proteins and combined them in an ensemble learning procedure. We applied this method to a dataset of known co-complexed proteins in Escherichia coli and compared it to previously published methods. We show that our method allows prediction of PPI with an unprecedented precision of 95.5% for the first 200 sorted pairs of proteins compared to 28.5% on the same dataset with the previous best method. A close inspection of the best predicted pairs allowed us to detect new or recently discovered interactions between chemotactic components, the flagellar apparatus and RNA polymerase complexes in E. coli.