Algorithms for incorporating prior topological information in HMMs: application to transmembrane proteins

BMC Bioinformatics

Table 2 Results of the independent test on a dataset of 26 transmembrane proteins with known three-dimensional structures. The proteins were chosen not to have significant sequence identity (<30%) with the proteins used to train the methods: HMM-TM, UMDHMM^TMHP, TMHMM and HMMTOP. The methods that allow the incorporation of experimental information are listed separately

Method	Q	C	SOV	Correctly predicted TM segments (%)	Correctly predicted Topologies (%)
Methods that allow the incorporation of experimental information
HMM-TM	0.899	0.780	0.942	21/26 (80.77%)	21/26 (80.77%)
TMHMM	0.899	0.782	0.956	19/26 (73.08%)	17/26 (65.38%)
HMMTOP	0.881	0.744	0.925	19/26 (73.08%)	18/26 (69.23%)
Phobius †	0.894	0.773	0.907	15/26 (57.69%)	13/26 (50%)
Methods that do not allow the incorporation of experimental information
MEMSAT	0.890	0.762	0.928	16/26 (61.54%)	13/26 (50%)
UMDHMM^TMHP	0.896	0.777	0.947	23/26 (88.46%)	22/26 (84.61%)
S-TMHMM †	0.899	0.781	0.957	21/26 (80.77%)	20/26 (76.92%)
PRO-TMHMM*†	0.870	0.718	0.916	16/26 (61.54%)	15/26 (57.69%)
PRODIV-TMHMM*†	0.897	0.778	0.946	19/26 (73.08%)	19/26 (73.08%)

* The methods using evolutionary information are denoted with an asterisk.
† These predictors were trained on sets containing sequences similar to the ones included in the test set.

ISSN: 1471-2105