Homology-based prediction of interactions between proteins using Averaged One-Dependence Estimators

BMC Bioinformatics

Table 2 Evaluation of true prediction performance on Dset2

Method	TP	FP	TN	FN	Sp (%)	Re (%)	Pr (%)	F
PSOPIA (θ = 0.293, the higheset F)	143	5,026	1,766,423	4,152	99.72	3.33	2.77	0.030
PSOPIA (θ = 0.670)	24	151	1,771,298	4,271	99.99	0.56	13.71	0.012
PSOPIA (θ = 0.890)	4	31	1,771,418	4,291	99.99	0.09	11.43	0.002
(I) BIPS, only filtered by the template interactions
(A) Template: Taxonomy ID = 9609 (human)	19	680	1,765,404	3,710	99.96	0.51	2.72	0.009
(B) Template: all species	19	833	1,765,005	3,705	99.95	0.51	2.23	0.008
(II) BIPS, filtered by known DDIs (iPfam or 3DID)
(A) Template: Taxonomy ID = 9609 (human)	5	60	1,771,096	4,261	99.99	0.12	7,69	0.002
(B) Template: all species	5	72	1,771,059	4,256	99.99	0.12	6.49	0.002
(III) BIPS, filtered by known DDIs (iPfam or 3DID) and GO; biological process, cellular component or molecular function
(A) Template: Taxonomy ID = 9609 (human)	3	47	1,771,245	4,284	99.99	0.07	6.00	0.001
(B) Template: all species	3	56	1,771,216	4,280	99.99	0.07	5.08	0.001

For PSOPIA trained on Dset1 (a data set independent of Dset2), the best threshold value, 0.995, which gave the highest F-measure in the 10-fold CV, was used to classify a pair of proteins as interacting or non-interacting. For BIPS, the default values in homologue conditions were used: joint identities ≥ 80%, joint e-values ≥ 1.0 × e⁻¹⁰, and template sequence coverage ≥ 80% (see [21] for more details of these parameters). In addition to the filtering by the template interactions only (I), two additional filters were applied: (II) filtered by known DDIs in iPfam or 3DID and (III) filtered by known DDIs and GO annotations (biological process, cellular component or molecular function). Furthermore, two template interactions, (A) only from human (taxonomy ID = 9609) and (B) from all species, were also considered.

ISSN: 1471-2105