Fig. 4

IF-SitePred “missed” predictions. Eight examples of human protein chains where IF-SitePred does not report any predicted ligand binding sites. Predictions are made on ligand-stripped chains. Ligand molecules, in orange, are superposed to illustrate how the ligandability scores recapitulate the observed binding site. These are protein representative chains and ligand molecules might not be observed in the same entry. A GDP-fucose protein O-fucosyltransferase 2, Q9Y2G5, with GDP-L-fucose (GFB) superimposed (PDB: 4AP6) [76]; B tRNA (cytosine(72)-C(5))-methyltransferase NSUN6, Q8TEA1, (PDB: 5WWT) [77] with superposed sinefungin (SFG) (PDB: 5WWR); C tubulin beta-2B chain, Q9BVA1, with phosphomethylphosphonic acid guanylate ester (G2P) (PDB: 7ZCW) [78]; D cyclic GMP-AMP phosphodiesterase SMPDL3A, Q92484, with cytidine-5'-monophosphate (C5P) (PDB: 5EBE) [79]; E tRNA (adenine(58)-N(1))-methyltransferase catalytic subunit, Q96FX7, with SAH (PDB: 5CCB) [80]; F chronophin, Q96GD0, (PDB: 5GYN) [81] with PLP (PDB: 2FCT) [82]. G Mitochondrial methylmalonic aciduria type A protein, Q8IVH4, with GDP (PDB: 8GJU) [83]; H renalase, Q5VYX0, (PDB: 3QJ4) with FAD [84]. Residues are coloured based on the ligandability score calculated by averaging the probabilities predicted by each of the 40 IF-SitePred prediction models. This is a score (Eq. 13) ranging 0–1 which is indicative of the likelihood of a given residue binding a ligand (see “Methods”). Clear pockets can be observed formed by residues with high ligandability scores (darker blue colour), which agree with the sites where ligands bind. These “missed” pockets contribute to the lower recall of IF-SitePred and strongly suggest that a more permissive ligand-binding residue selection threshold (currently all 40 models) or a different clustering scheme might be able to capture these predictions and increase the recall of this method