Figure 5
ROC and IPX score distributions. Mirrored ROC curves for the five genome pairs in the known positive sets in KP. Inset: IPX score distributions (reproduced in Additional File 3: 'Interolog Prioritisation Index Histograms'). For each characteristic, the point at coordinate (1,1) corresponds to IPXthr = min(IPX), TPR = 100% and FPR = 100%. The point at coordinate (0,0) corresponds to IPXthr = max(IPX), TPR = 0% and FPR = 0%. Initially, IPXthr = min(IPX). Then, the score histogram is divided into 1000 segments and IPXthr is incremented until IPXthr = max(IPX) is reached. For all datasets, the decrease of TPR is slower than the decrease of FPR as IPXthr → max(IPX). This means that, as the score threshold becomes more stringent, for all datasets the number of known positive samples lost stays smaller than the number of new predictions lost. The correlation between TPR and the FPR varies depending on the dataset: in the case of the Yeast-Worm pair, 98% of known positives are retrieved when the novel prediction retrieval rate is down to about 76%. Conversely, in the Human-Yeast case, the TPR is down to about 92% for 98% FPR.