Comparison of predictive accuracy of three motif finding models using both weakly structured and stem-loop sequences in the test set.

The values show the average AUC-PR across two test sets and bold values indicate the best performing method. Rows are sorted by decreasing relative gain of RNAcontext to the best of MEMERIS & MatrixREDUCE. For all the methods, displayed values were calculated using the single best motif model for each method chosen based on the two training set performance. According to the Wilcoxon's sign rank test performed on paired AUC-PR values across 1,000 bootstrap samples from the test set results of the three methods, all differences between RNAcontext AUC-PR and that of the other algorithms are statistically significant (the largest P-value is , see Table S3 for the complete results of bootstrap analysis) except for the differences on PTB and U1A.