Abstract
Protein secondary structure prediction has been extensively discussed for almost 50 years and the machine learning is one of feasible methods for it with more than 70% accuracy. PSIPRED, PHD and PROF are well-known machine learning approaches and based on the three-state prediction, helix, strand, and coil. Various prediction tools based on the machine learning have been proposed. However, these tools may make a lot of effort to develop and their accuracy was close to or even lower than PSIPRED. Under the concern, making use of or combining outputs from existing methods is an alternative to make improvements. RAP is a post-processing method using linear transformation and normalization to refine scores of three-state prediction. Hence, RAP can be easily applied to any protein secondary structure prediction tool if it uses three-state prediction. RAP was tested on the CASP data set with 181 targets and a large-scale data set with 69534 chains separated from 31402 proteins in PDB. In the experiment, PHD, PROF and PSIPRED were used to give scores of three-state prediction for each target protein; then, RAP predicted secondary structures by refining the scores from them. More secondary structural segments were detected by RAP than by PHD, PROF and PSIPRED. Moreover, prediction results of combining methods with RAP can achieve higher accuracy than without RAP. RAP is freely available via http://ensembl.cs.nthu.edu.tw/RAP/.