Identifying structure fragments by sequence alignments to predict protein secondary structures

Protein secondary structure prediction is an important intermediate step for many biological procedures. A better result is achieved with an enlarged database and refined searching approaches. Based on local sequence segments information, Rosetta is quite successful in predicting the three-dimensional structure of a folded protein from its linear sequence of amino acids. We thus use the segments obtained by BLAST to predict secondary structures of sequences from CB513. Two weight-assigning methods are applied to the matches. One uses the reciprocal of similarity or identity scores and their powers (1/sim[superscript d] or 1/id[superscript d]); the other uses similarity or identity scores and their powers (sim[superscript c] or id[superscript c]) directly (d and c are positive real numbers). Similarity/identity score levels and zones, solvent accessibility and occurrences of matching residues, sequence direction and length, etc are considered to be influential factors on prediction accuracy. Normalized scores for each position of the query sequence are defined and calculated. Some artificial approaches are used to select the most proper secondary structure element according to normalized scores of residues of query sequences. The present study shows comparable prediction accuracy with currently popular methods. Also, we learned about some factors that may influence secondary structure prediction, such as hydrophobic residues might be more important in the formation and maintaining protein structures; the more occurrences of the matching residues, the more reliable the prediction; and matching segments of different similarity levels play different roles in prediction. We find that our method works almost equally well for sequences of different sizes and structure classes in CB513. We used some artificial intelligence approaches in the last step of our prediction to determine the final secondary structure element according to normalized scores and achieve 3-4% improvement in accuracy