Abstract
BACKGROUND: For over 30 years potentials of mean force have been used to evaluate the relative energy of protein structures. The most commonly used potentials define the energy of residue-residue interactions and are derived from the empirical analysis of the known protein structures. However, single-body residue 'environment' potentials, although widely used in protein structure analysis, have not been rigorously compared to these classical two-body residue-residue interaction potentials. Here we do not try to combine the two different types of residue interaction potential, but rather to assess their independent contribution to scoring protein structures. RESULTS: A data set of nearly three thousand monomers was used to compare pairwise residue-residue 'contact-type' propensities to single-body residue 'contact-count' propensities. Using a large and standard set of protein decoys we performed an in-depth comparison of these two types of residue interaction propensities. The scores derived from the contact-type and contact-count propensities were assessed using two different performance metrics and were compared using 90 different definitions of residue-residue contact. Our findings show that both types of score perform equally well on the task of discriminating between near-native protein decoys. However, in a statistical sense, the contact-count based scores were found to carry more information than the contact-type based scores. CONCLUSION: Our analysis has shown that the performance of either type of score is very similar on a range of different decoys. This similarity suggests a common underlying biophysical principle for both types of residue interaction propensity. However, several features of the contact-count based propensity suggests that it should be used in preference to the contact-type based propensity. Specifically, it has been shown that contact-counts can be predicted from sequence information alone. In addition, the use of a single-body term allows for efficient alignment strategies using dynamic programming, which is useful for fold recognition, for example. These facts, combined with the relative simplicity of the contact-count propensity, suggests that contact-counts should be studied in more detail in the future.