Fig. 3

Leveraging FunFams to better predict binding residues. The horizontal lines indicate the performance estimates for precision (Eq. 2) and recall (Eq. 3) of BindPredict-CCS and BindPredict-CC baseline predictions not using FunFams. Predictions are refined by constructing consensus predictions for the FunFams. The x-axes give different thresholds in terms of what fraction of the FunFams members need to have a binding prediction for a particular residue in order to label that residue as binding in the consensus prediction: from at least one (0.01) to all (1.0). Depending on the threshold, both precision and recall significantly increase over the standard prediction method. The two panels illustrate the improvement over two slightly different baseline prediction methods: a BindPredict-CCS using the cumulative couplings-based input features. In this case precision increases up to 61 ± 4%. Panel b shows the performance improvement for BindPredict-CC using the clustering coefficient-based input features. For low thresholds, these predictions reach recall up to 50 ± 2%