Inferring high-confidence human protein-protein interactions

BMC Bioinformatics

Table 3 Evaluation of different ranking schemes

Reference set	IDBOS		Hypergeometric		Occurrence		Random rank
	<A>	Gain	<A>	Gain	<A>	Gain	<A>
Far-Western blotting	0.409	184%	0.313	118%	0.178	24%	0.144
Isothermal titration calorimetry	0.542	215%	0.469	173%	0.286	66%	0.172
Nuclear magnetic resonance	0.632	157%	0.534	117%	0.331	34%	0.246
Surface plasmon resonance	0.567	160%	0.496	128%	0.320	47%	0.218
PDB complex PPIs	0.842	258%	0.746	217%	0.566	141%	0.235
Mouse homologous PPIs	0.318	222%	0.293	198%	0.207	110%	0.099

The average accuracy < A > and gain associated with interaction detection based on shuffling (IDBOS), hypergeometric test ranking, and occurrence-ranking schemes were estimated using Equations 3 and 4. The value associated with the randomly ranked data set does not represent a random protein-protein interaction (PPI) data set as only the rank, and not the interactions themselves, were randomized. The tested reference sets contain PPIs derived from 1) far-Western blotting, 2) isothermal titration calorimetry, 3) nuclear magnetic resonance, 4) surface plasmon resonance experiments, 5) known protein crystal complexes contained in the Protein Data Bank (PDB), and 6) those derived from experimentally determined mouse PPIs. We tested whether the numerical differences in values of < A > for each pairwise comparison within each reference set were due to chance using the t-test, and rejected this hypothesis with a p-value of <10^-6.

ISSN: 1471-2105