Ant to the classification problem. We proposed a feature-selection procedure combining filter and wrapper methods to select a subset of feature elements which can make the classifier achieve best prediction performance. In the first step, F-score was used to measure the discriminative power of each feature element between the positive and negative sets, which is defined as follows, (i(z) { i )2 z( ({) {i )2 x x xi x , ??nz X (z) X ({) 1 1 n{ (z) 2 ({) 2 (x {i ) z x (x {i ) x nz {1 k 1 k,i n{ {1 k 1 k,iMaterials and Methods Data collectionData used in this study were retrieved from GeneDB [19] by June 2012. To ensure data quality, we took the information of “Curation” and “Gene Ontology” from GeneDB into account, and only selected the proteins with consistent supporting information. Finally, 156 T. brucei proteins were collected as flagellar proteins of high quality based on the comprehensive annotation from GeneDB. To generate a negative dataset for the classification, we extracted T. brucei proteins containing annotation for `cellular component’ from GeneDB together with the mitochondrial proteins collected in our previous study [9]. This set was filtered by removing the entries either annotated as flagellar MedChemExpress ML-240 related or with low confidence such as “by similarity”, “potential” and “probable”. We retained 652 proteins as nonflagellar proteins with high confidence. To obtain a non-redundant dataset, BLASTclust [20] was used to remove redundant proteins with sequence identity higher than 30 , and 8 flagellar and 60 non-flagellar proteins were discarded from the collected dataset. Thus, 148 flagellar and 592 non-flagellar 24272870 proteins were finally used as our positive and negative sets, respectively. Systematic IDs of these positive and negative samples are listed in Table S1. We randomly selected 3/4th of the positive and negative data as the training set. The remaining data were used as the test set. To assess the performance and stability of the prediction model, we repeated the random sampling process fifty times, and 58-49-1 obtained 50 groups of training and test sets.F (i)Feature constructionWe examined a number of features which are potentially useful for the identification of flagellar proteins based on the general understanding of protein subcellular localization. The initial features can be grouped into five categories: (a) basic sequence attributes such as sequence length, amino acid composition and dipeptide composition; (b) physicochemical and biochemical properties, such as extinction coefficient, instability index, aliphatic index, and various amino acid propensities obtained from AAindex (http://www.genome.ad.jp/aaindex) [21]; (c) structural properties such as secondary structural content [22], unfoldabilityi i where xi , x(z) and x({) are the average value of the ith feature over the whole, positive and negative datasets, respectively; x(z) k,i and x({) are the ith feature of the kth protein in the positive and k,i negative datasets, respectively; and nz and n{ are the numbers of proteins in the positive and negative datasets, respectively. The larger an F-score is, the more discriminative the feature is. In the first round, feature elements with F-scores above a pre-selected threshold were retained and used in the next step feature selection. The F-score threshold was selected based on the distribution of the sorted F-scores of all feature elements, and the cross-validationTFPP: Trypanosome Flagellar Protein Predictoraccura.Ant to the classification problem. We proposed a feature-selection procedure combining filter and wrapper methods to select a subset of feature elements which can make the classifier achieve best prediction performance. In the first step, F-score was used to measure the discriminative power of each feature element between the positive and negative sets, which is defined as follows, (i(z) { i )2 z( ({) {i )2 x x xi x , ??nz X (z) X ({) 1 1 n{ (z) 2 ({) 2 (x {i ) z x (x {i ) x nz {1 k 1 k,i n{ {1 k 1 k,iMaterials and Methods Data collectionData used in this study were retrieved from GeneDB [19] by June 2012. To ensure data quality, we took the information of “Curation” and “Gene Ontology” from GeneDB into account, and only selected the proteins with consistent supporting information. Finally, 156 T. brucei proteins were collected as flagellar proteins of high quality based on the comprehensive annotation from GeneDB. To generate a negative dataset for the classification, we extracted T. brucei proteins containing annotation for `cellular component’ from GeneDB together with the mitochondrial proteins collected in our previous study [9]. This set was filtered by removing the entries either annotated as flagellar related or with low confidence such as “by similarity”, “potential” and “probable”. We retained 652 proteins as nonflagellar proteins with high confidence. To obtain a non-redundant dataset, BLASTclust [20] was used to remove redundant proteins with sequence identity higher than 30 , and 8 flagellar and 60 non-flagellar proteins were discarded from the collected dataset. Thus, 148 flagellar and 592 non-flagellar 24272870 proteins were finally used as our positive and negative sets, respectively. Systematic IDs of these positive and negative samples are listed in Table S1. We randomly selected 3/4th of the positive and negative data as the training set. The remaining data were used as the test set. To assess the performance and stability of the prediction model, we repeated the random sampling process fifty times, and obtained 50 groups of training and test sets.F (i)Feature constructionWe examined a number of features which are potentially useful for the identification of flagellar proteins based on the general understanding of protein subcellular localization. The initial features can be grouped into five categories: (a) basic sequence attributes such as sequence length, amino acid composition and dipeptide composition; (b) physicochemical and biochemical properties, such as extinction coefficient, instability index, aliphatic index, and various amino acid propensities obtained from AAindex (http://www.genome.ad.jp/aaindex) [21]; (c) structural properties such as secondary structural content [22], unfoldabilityi i where xi , x(z) and x({) are the average value of the ith feature over the whole, positive and negative datasets, respectively; x(z) k,i and x({) are the ith feature of the kth protein in the positive and k,i negative datasets, respectively; and nz and n{ are the numbers of proteins in the positive and negative datasets, respectively. The larger an F-score is, the more discriminative the feature is. In the first round, feature elements with F-scores above a pre-selected threshold were retained and used in the next step feature selection. The F-score threshold was selected based on the distribution of the sorted F-scores of all feature elements, and the cross-validationTFPP: Trypanosome Flagellar Protein Predictoraccura.