Methods for affinity prediction do not perform well in the case of peptide-
protein binding. Here we investigate the cause for this failure. We conclud
e that the affinity of a small set of complexes can be described well by si
dechain steric interactions and the number of buried polar groups, but only
if outliers are removed. In the PLS models, particularly the hydrogen bond
ing of the peptide backbone plays an anomalous role, explaining the inabili
ty of general scoring methods to predict peptide-protein binding.