A novel statistical analysis of non-bonded contacts in a set of known protein structures shows that the natural residue types fall into five or six groups distinguishable by nearest neighbor preference. The observed pattern of contact specificities clearly reflects residue hydrophobicity and charge. Its most striking feature is that residues in the hydrophobic group make about twice as many contacts with one another as would be expected on a random basis. A similar increase in hydrophobic contact frequency can be observed at the level of individual proteins. Native proteins make, on average, about twice as many hydrophobic contacts as corresponding misfolded proteins, generated by computer. On the basis of these observations increased hydrophobic contact frequency is proposed as a simple model of the hydrophobic effect.