On the performance of combined dichotomic predictors of natively unfolded proteins
arXiv:0806.4838
Abstract
The performance of single folding predictors and combination scores is critically evaluated. We test mean packing, mean pairwise energy and the new index gVSL2 on a dataset of 743 folded proteins and 81 natively unfolded proteins. These predictors have an individual performance comparable or even better than other proposed methods. We introduce here a strictly unanimous score S_{SU} that combines them but leaves undecided those sequences differently classified by two single predictors. The performance of the single predictors on a dataset purged from the proteins left unclassified by S_{SU}, significantly increases, indicating that unclassified proteins are mainly false predictions. Amino acid composition is the main determinant considered by these predictors, therefore unclassified proteins have a composition compatible with both folded and unfolded status. This is why purging a dataset from these ambiguous proteins increases the performance of single predictors. The percentage of proteins predicted as natively unfolded by S_{SU} in the three kingdoms are: 4.1% for Bacteria, 1.0% for Archaea and 20.0% for Eukarya; compatible with previous determinations. Evidence is given of a scaling law relating the number of natively unfolded proteins with the total number of proteins in a genome; a first estimate of the critical exponent is 1.95 +- 0.21
arXiv commentThe title of the paper has been changed to convey the focus of the work. The text has been slightly modified to clarify the subject. A more extended version of the paper has been published in BMC Bioinformatics 2010, 11:198 (21 April 2010)