Evaluating Representations for the Shine-Dalgarno Site in Escherichia coli

Several methods for identifying individual motif instance by exhaustive evaluation of k-mers (k ≤ 10) are applied to the pooled Upstream Regions (USR) of all 4289 Escherichia coli ORFs. Instances of the Shine-Dalgarno (SD) site are readily identified using these methods. Using these motif instances as starting points, various motif representations and training methods, including several new algorithms, are applied to characterize the complete SD motif. Motif representation languages of increasing power give increasingly better characterizations of the SD motif, permitting more SD sites to be reliably identified. In particular, matrix representation is better than IUPAC which is better than k-mer prototype. However, overly powerful representation also results in suboptimal characterization. A variety of matrix techniques using different representations, objective functions and learning methods yield approximately the same motif, providing evidence for the robustness of the result and the effectiveness of the methods. By these measures, about 1/4 of the ORFs have no better than random SD sites. More biologically realistic motif representation languages might further reduce that fraction.