With the emergence of large collections of protein−ligand complexes complemented by binding data, as found in PDBbind or BindingMOAD, new opportunities for parametrizing and evaluating scoring functions have arisen. With huge data collections available, it becomes feasible to fit scoring functions in a QSAR style, i.e., by defining protein−ligand interaction descriptors and analyzing them with modern machine-learning methods. As in each data modeling ansatz, care has to be taken to validate the model carefully. Here, we show that there are large differences measured in R (0.77 vs 0.46) or R 2 (0.59 vs 0.21) for a relatively simple scoring function depending on whether it is validated against the PDBbind core set or validated in a leave-cluster-out cross-validation. If proteins from the same family are present in both the training and validation set, the estimated prediction quality from standard validation techniques looks too optimistic.