Fig. 1

Concept of using FunFam to filter binding residue predictions. For the example of protein glutathione S-transferase (identifier 1U3I [17, 18]) binding glutathione. The binding residues were shown on the structure using PyMol [19]. Correctly predicted binding residues (TP) are shown in darkblue, incorrectly predicted non-binding residues (FN) in lightblue, and incorrectly predicted binding residues (FP) in red. a Poor binding prediction: Some prediction method (here BindPredict-CCS) might correctly identify only a small fraction of all binding residues (here in red with a precision = recall = F1 = 11%). The method might even incorrectly over-predict more residues as binding (red) and might miss more observed binding residues (lightblue) than it gets right. b FunFam filter with 1% prediction agreement: Simply filtering the prediction by requiring that at least 1% of all proteins aligned at a particular residue position had the same binding residue prediction (consensus threshold = 0.01). For the example, given, this boosted recall to 67% (precision = 16%, F1 = 26%). c FunFam filter with 50% prediction agreement: Filtering the prediction by requiring consensus threshold of 0.5 (50% of the residues predicted equally) removed most predicted binding residues without removing the correctly predicted ones (correctly predicted residues shown in darkblue identical in a and c; precision = 20%, recall = 11%, F1 = 14%)