Skip to main content
. 2013 Sep 23;29(24):3135–3142. doi: 10.1093/bioinformatics/btt554

Table 1.

Results of LOO tests using SVM classifiers based on different feature vectors

Class Featurea C|γb IVA versus non-effectorc
IVB versus non-effector
Sn Sp Acc MCC Sn Sp Acc MCC
Single AAC(20D) 10, 0.3 0.633 0.967 0.900 0.667 0.694 0.940 0.887 0.655
DPC(400D) 10, 0.0125 0.533 0.992 0.900 0.663 0.674 0.932 0.877 0.625
PSSM(400D) 10, 0.02 0.733 0.983 0.933 0.782 0.894 0.975 0.958 0.874
PSSM_AC(200D, G = 10) 10, 0.1 0.667 0.967 0.907 0.691 0.829 0.957 0.929 0.790
Combined AAC+DPC(420D) 10, 0.02 0.567 1.000 0.913 0.715 0.688 0.932 0.882 0.637
PSSM+AAC(420D) 10, 0.02 0.767 0.975 0.933 0.784 0.897 0.976 0.959 0.878
PSSM_AC+AAC(220D) 10, 0.05 0.667 0.975 0.913 0.712 0.848 0.966 0.941 0.824
PSSM+PSSM_AC(600D) 10, 0.00125 0.733 0.958 0.913 0.720 0.897 0.964 0.949 0.852
PSSM+PSSM_AC+AAC(620D) 10, 0.00125 0.700 0.983 0.927 0.759 0.897 0.964 0.950 0.853

aAAC: amino acid composition; DPC: dipeptide composition; PSSM: PSSM composition; PSSM_AC: auto covariance transformation of PSSM profiles. The figure in the bracket refers to the dimensions of the features.

bC and γ are the cost and the gamma parameter of the SVM, respectively. They were optimized based on a 10-fold cross-validation grid search.

c120 non-effectors were selected randomly in the non-effector dataset as negative training samples for keeping the ratio of effectors to non-effectors at ∼1:4.