Although the machinery for translation initiation in Escherichia coli is very complicated, the translational efficiency has been reported to be predictable from upstream oligonucleotide sequences. Conventional models have difficulties in their generalization ability and prediction nonlinearity and in their ability to deal with a variety of input attributions. To address these issues, we employed structural learning by artificial neural networks to infer general rules for translational efficiency. The correlation between translational activities measured by biological experiments and those predicted by our method in the test data was significant (r=0.78), and our method uncovered underlying rules of translational activities and sequence patterns from the obtained skeleton structure. The significant rules for predicting translational efficiency were (1) G- and A-rich oligonucleotide sequences, resembling the Shine-Dalgarno sequence, at positions -10 to -7; (2) first base A in the initiation codon; (3) transport/binding or amino acid metabolism gene function; (4) high binding energy between mRNA and 16S rRNA at positions -15 to -5. An additional inferred novel rule was that C at position -1 increases translational efficiency. When our model was applied to the entire genomic sequence of E. coli, translational activities of genes for metabolism and translational were significantly high.