For three miRNAs, our predictions (condition: perfect Watson-Crick complementarity at nt 2–7) on targets experimentally validated by Miranda et al. [9].

Of 158 genes experimentally tested for regulation by miR-134, 85 occur in our database, as do 14 of 24 tested for regulation by miR-296, and 22 of 44 tested for regulation by miR-375. As Miranda et al. set different criteria for target prediction (e.g. allowing GU pairs and a mismatch in the seed regions) and selected sites for validation on that basis, not all of their validated sites could have been discovered by our approach. To correct (to the extent possible) for this difference, we excluded a further 20 target sites with GU pairs and mismatches in the seed region for miR-134, and 8 target sites for miR-296 (all target sites for miR-375 have WC matches in the seed region), and report the results of 65 target genes examined for miR-134, 6 for miR-296 (for miR-296, all six sites examined were validated), and 22 for miR-375. Miranda et al. also give details of predicted sites that failed validation (i.e. were false positive predictions); in principle this makes it possible for us to calculate the specificity of our approach. In some cases, our 40% free energy threshold is more-stringent than that of Miranda et al. (∼−16.4 kcal/mol), reducing the sensitivity calculated for our approach. As the number of false positives is small, specificity values should be interpreted as indicative only.