Fig. 6

A comparison of the benefits of combining predictions for multiple structures. Left: Predictive power for binding residue annotation improves when predictions for 10 structures were combined for all three methods (blue, orange, green) compared to when predictions for single structures were used (grey). IF-SitePred had a slightly higher original F1 score, and also saw the greatest improvement of the three methods, increasing to 0.59. This trend is also reflected when the Matthews correlation coefficient is calculated (Additional file 1:Figure 3). Right: Precision-recall curves for all three methods reveal that IF-SitePred has a higher average precision (AP) (0.58) than P2Rank (0.50) and DeepPocket (0.50). Iso-F1 curves are shown in grey, demonstrating that IF-SitePred achieves higher F1 scores across all probability thresholds