Weak correlation between predictive power of individual sequence patterns and overall prediction accuracy in proteins

Patterns in amino acid properties (polar, hydrophobic, etc.) that characterize secondary structure motifs are derived from a database containing 75 protein structures, with the aim of circumventing the limitations due to data base size so as to increase structure prediction score. Many such sequence-structure associations with high intrinsic predictive power are found, which turn out to be correct 78% of the time when applied individually to proteins outside the learning set. Based on these associations, a prediction method is developed, which reaches the score of 62% on the 3 states alpha-helix, beta-strand, and loop, without using additional constraints. Though this score is quite good compared to that of other available prediction methods, it is much lower than could be expected from the high intrinsic predictive power of the associations used. The reasons underlying this surprising result, which indicate that prediction score and intrinsic predictive power are only weakly coupled, are discussed. It is also shown that the size of the present database still seriously limits prediction scores, even when property patterns are used, and that higher scores are expected in large databases. Clues are provided on the relative influence of neglecting spatial interactions on prediction efficiency, suggesting that, in sufficiently large databases, predicted secondary structures would correspond to those formed early in the folding process. This hypothesis is tested by confronting present predictions with available experimental data on early protein folding intermediates and on small peptides that adopt a relatively stable conformation in water. Although admittedly there are still too few such data, results suggest that the hypothesis might be well founded.