Table 3 A comparison of ligand similarity to training set with success rates. DeepPocket performs worse than average on sites that bind ligands significantly different to the training set. We calculated the USRCAT similarity for the most similar ligand in the training set to that binding each protein in the HAP-small set, and calculated the fraction difference between overall top-1 success rate (from Table 1) for each 0.2 interval of USRCAT similarity

From: Learnt representations of proteins can be used for accurate prediction of small molecule binding sites on experimentally determined and predicted protein structures