Motivated by known preferences for certain amino acids in positions around a-helices, we developed neural network-based predictors of both N and C a-helix ends, which achieved about 88% accuracy. We applied a similar approach for predicting the ends of three types of secondary structure segments. The predictors for the ends of H, E and C segments were then used to create input for protein secondary-structure prediction. By incorporating this new type of input, we significantly improved the basic one-stage predictor of protein secondary structure in terms of both per-residue (Q3) accuracy (+0.8%) and segment overlap (SOV3) measure (+1.4).
展开▼
