Skip to main content
. 2006 Nov 16;35(Database issue):D232–D236. doi: 10.1093/nar/gkl812

Figure 2.

Figure 2

Cumulative results of a benchmark test PCB00033. The underlying dataset is a small subset of SCOP comprising of 55 classification tasks (corresponding to 8 all-α, 15 all-β, 30 α/β and 2 other classes). The numbers represent average AUC values [0,1] obtained by receiver operator curve (ROC) analysis (18). This value is high for good classifiers and is close to 0.5 for random classification. The classification methods include 1NN—Nearest neighbor (30), RF—Random forest (16), SVM—Support Vector Machines (14), ANN—Artificial neural networks (15) and LogReg—Logistic regression (17). The comparison methods include BLAST (8), SW—Smith–Waterman (9), NW—Needleman–Wunsch (10), LZW—Lempel–Ziv compression distance and PPMZ—partial match compression distance (11). The Smith–Waterman algorithm performs better than the other comparison algorithms, especially when used in conjunction with SVM.