EnsembleFam: towards more accurate protein family prediction in the twilight zone

BMC Bioinformatics

Table 2 Performance comparison of different methods on the twilight zone sequences, i.e. sequences having less than \(40\%\) identity is shown in this table

Dataset	Method	predCount = 1	predCount = 2	predCount = 3	predCount = 4	predCount = 5	predCount \(> 5\)
Identity: \(0 < x \le 30\)
COG-500-1074	EnsembleFam	72.07	81.00	82.82	84.96	85.33	85.27
	pHMM	69.54	73.75	55.51	70.62	70.85	73.55
	DeepFam	57.14	54.52	49.90	46.92	43.64	35.94
COG-250-1796	EnsembleFam	72.84	77.07	81.02	82.14	84.66	86.45
	pHMM	75.39	73.82	73.84	71.02	67.44	72.43
	DeepFam	32.44	32.54	30.24	29.53	30.02	28.68
COG-100-2892	EnsembleFam	75.24	79.55	81.21	80.63	82.05	88.95
	pHMM	63.44	59.69	53.45	48.16	47.42	57.57
	DeepFam	27.30	26.13	25.54	27.62	24.83	25.36
Identity: \(30 < x \le 40\)
COG-500-1074	EnsembleFam	90.96	94.51	95.88	96.16	97.08	97.84
	pHMM	62.22	61.20	88.95	87.38	85.19	85.85
	DeepFam	58.45	58.32	59.39	58.41	58.37	54.81
COG-250-1796	EnsembleFam	91.54	95.19	95.52	95.95	96.62	97.73
	pHMM	63.05	89.41	89.05	87.74	84.82	83.69
	DeepFam	47.09	48.38	50.12	51.09	50.73	48.78
COG-100-2892	EnsembleFam	92.92	95.23	96.04	96.35	96.81	97.99
	pHMM	87.07	87.78	86.08	84.04	80.16	81.69
	DeepFam	38.73	42.62	46.07	48.33	49.30	45.32

The best results are highlighted in bold font. The dataset is divided into six subgroups based on the number of predictions made by EnsembleFam. Using the column “predCount = 5” as an example, the accuracy in this table is computed as follows. For a protein, if EnsembleFam makes 5 function predictions for it, and one of these is correct, the protein is counted as correct in the column “predCount = 5”; if all 5 function predictions are incorrect, the protein is counted as a wrong prediction. For the same protein, regardless of how many function predictions are made by pHMM, as long as one of these is correct, the protein is counted as correct in the column “predCount = 5”; otherwise, the protein is counted as incorrect in the column. As for DeepFam, which makes exactly one prediction for each protein, the same protein is counted as correct in the column “predCount = 5” if and only if the sole DeepFam prediction for it is correct. All the accuracy value showed in the table is the average of 3-fold cross-validation

ISSN: 1471-2105