Figure 2

ROC curves comparing in silico methods. (a) A PSWM constructed from 20 medium and strong PPREs that contain multiple variations, and a PSAM constructed using the single nucleotide data and ten initializations of PPRE classifier created based on Table 1 and random sampling of Figure 1 and Additional data file 1 were compared for their ability to detect binding. True positive rates (TPRs) and false positive rates (FPRs) were calculated, with false positives given when no binding was detected despite prediction, and false negatives given when binding was detected but not predicted (correlation of matrix scores to predicted binding was done based on lines fitted to correlation plots shown in Additional data file 3). A line of no discrimination is a diagonal line and optimum performance approaches the value (0, 1). For clarity, only one representative instance of a PPRE classifier is shown in (a). (b) To assess how good the predicted experimental binding estimates were, the performance of the method used was tested with a 15% tolerance interval for a match to experimental binding (5% when prediction was 15% or less) using a single cut-off (the optimal cut-off was 3% for the classifier, 25% or a score of 0.0000015 for PSAM, and 20% or a score of 4.7 for PSWM) and calculating again the FPR and TPR for each method. False positives in this case represented predictions that were too high and false negatives predictions that were too low.