Fig. 5

Case study illustrating the effect of changing the PLEC distance cutoff on model attributions. a The original synthetic protein-ligand complex generating using the Polar generative process, labelled as active as a result of the circled interaction. b Perturbed examples. We generated 50 synthetic protein-ligand complexes, which were all identical apart from the HBA synthetic residue contained in the circle, whose position was perturbed in relation to the ligand HBD with which it interacts. The 5 spheres within the circle illustrate 5 of the 50 positions occupied by the interacting residue. c Illustration of how the relative importance of the key ligand atom changes as the distance between it and the perturbed synthetic residue changes. Despite the synthetic residue and ligand atom no longer interacting when the distance between them is greater than 4 Å, the RF_PLEC_4.5 and RF_PLEC_5 models continued to rank the ligand atom highly until the synthetic residue-ligand atom distance was greater than the respective PLEC distance cutoff