Improving prediction of burial state of residues by exploiting correlation among residues

BMC Bioinformatics

Table 1 Prediction accuracy of ACRF and BRNN on the four datasets α,β,α/β,α+β

Methods	α	β	α/β	α+β
LR	0.821 ±0.005	0.801 ±0.004	0.808 ±0.003	0.809 ±0.005
BRNN	0.825 ±0.004	0.805 ±0.003	0.812 ±0.004	0.812 ±0.006
ACRF	0.833 ±0.006	0.813 ±0.005	0.818 ±0.003	0.822 ±0.005
ACRF-CN	0.806 ±0.006	0.785 ±0.004	0.787 ±0.003	0.794 ±0.006
ACRF-CN-SC	0.805 ±0.004	0.782 ±0.005	0.783 ±0.005	0.789 ±0.007
ACRF-CN-SC-SS	0.801 ±0.004	0.769 ±0.004	0.773 ±0.005	0.784 ±0.005

For the sake of fair comparison, ACRF and BRNN use identical feature sets. To investigate the effects of different features on prediction accuracy, we evaluated a set of variants of ACRF, including ACRF-CN with contact number removed, ACRF-CN-SC with both contact number and sequence conservation removed, and ACRF-CN-SC-SS with contact number, sequence conservation, and secondary structure information removed from the ACRF model

ISSN: 1471-2105