PredGPI: a GPI-anchor predictor

BMC Bioinformatics

Table 1 Comparison between PredGPI and other available predictors

Predictor	TP	FP	Cov (%)	Acc (%)	FP rate (%)	MCC
PredGPI	112	15	77.2	88.2	0.14	0.823
FragAnchor	102	37	70.3	73.4	0.35	0.725
BIG-PI	79	33	53.4	70.5	0.31	0.609
DGPI	117	250	79.1	31.9	2.35	0.492
GPI-SOM	126	182	85.1	40.9	1.7	0.583
MemType-2L (Shangai server*)	74	189	51.0	28.1	1.8	0.368
MemType-2L (Harvard server) *	≤ 107	≥ 60	≤ 73.8	≤ 64.1	≥ 0.56	≤ 0.683

Performances are evaluated on 145 positive and 10,630 negative examples contained in GPI-Set and Non-GPI-Set, respectively. PredGPI performances were evaluated using the jack-knife procedure. It's worth noticing that many of the tested proteins may have been used for the training of other predictors.
Abbreviations: TP = True Positives, FP = False Positives; the number of sequences is listed; Cov = Coverage, Acc = Accuracy, FP rate = False Positives over the total number of negative examples, MCC = Matthews Correlation Coefficient.
*MemType-2L is available in two versions: the Shangai server http://www.csbio.sjtu.edu.cn/bioinf/ and the Harvard server http://sbgrid.org/chou/bioinf/MemType/. For sake of completeness we used both.
** The Harvard server of MemType-2L gave an answer only for 143 out of the 145 positive examples comprised in GPI-Set. 105 sequences are correctly predicted in this set. Moreover the server gave an answer only for 4,265 out of the 10,630 negative examples comprised in Non-GPI-Set. The number of mispredictions in this set is equal to 60. The limits to the indexes scoring the performance of this server are computed in the best case, which is by considering all the non predicted proteins as correctly predicted.

ISSN: 1471-2105