Poster #RP209
FORECAST (fold recognition by combining profile-profile alignment and support vector machine)
Chan-seok Jeong*, Minho Lee*, Dongsup Kim*
*Korea Advanced Institute of Science and Technology, Daejeon, 305-701, Korea
We develop an automated protein structure prediction server combining profile-profile alignment and support vector machine (http://pbil.kaist.ac.kr/forecast/). For each template its related sequence homologs at SCOP superfamily level are aligned by profile-profile alignment and the corresponding n+2 dimensional feature vectors are calculated. Those feature vectors of the template are used as the positive examples to train the template-specific support vector machine (SVM). For a given query sequence, the profile-profile alignments to the templates are conducted, the resulting feature vectors are classified by the SVMs, the SVM outputs are converted to posterior probabilities, and finally the templates are ranked by the probabilities. At fold level the similar procedure is applied to construct the SVM classifiers that are optimized to recognize the homologs related at the SCOP fold level. When comparing the performance to PSI-BLAST and profile-profile alignment resulting Z-score, our server significantly outperforms them at every SCOP level. Our server is useful for recognizing distantly related structural homologs that other methods fail to recognize due to the low sequence similarity.