Accuracy of predicting buried helices by PHDsec

Contact: Burkhard Rost (rost@EMBL-Heidelberg.de)
Date: May, 1996

• Question:
  Are buried helices predicted at lower accuracy than exposed ones?
• Method:
  Secondary structure prediction: PHDsec, Profile-based neural network system from Heidelberg (Rost & Sander, 1994; Abstract).
  Expected overall accuracy: Three-state per-residue accuracy (percentage of residue predicted correctly in either of the three states helix, strand, rest): > 72% +/- 9% (one standard deviation).
  Dataset: Cross-validation experiment on 705 sequence unique protein chains (largest sequence unique subset of PDB in May, 1996, i.e., no two proteins in that set have more than 25% pairwise identical residues).
  
• Result and conclusion:
  It has been argued in the literature that buried helices would in general be predicted less accurately than exposed helices. The figure suggest that this is wrong on average, at least for PHDsec. The following three conclusions can be drawn.
  
  1. The per-residue accuracy is highest for more buried helical residues.
  2. The per-segment accuracy is higher for buried helices than for exposed helices.
  3. For helical segments the best predictions are obtained for intermediate values of accessibility (averaged over the entire segment).

• Figure caption:
  Accuracy of predicting residues in helices (or entire helical segments) is plotted vs. the observed relative solvent accessibility (according to DSSP assignments). Values for per-residue accuracy are averaged over all residues with a given relative solvent accessibility. Values for per-segment accuracy are averaged over all helical segments with a given average accessibility.