Fig. 7
The weighted string kernel WSeqKernel versus other string kernel methods. We compare the performance of five different distances between protein sequences for detecting remote homologies. Those distances include the distances based on the weighted string kernel defined in this work, with (β,k max )=(0.2,10), the weights set according to the Mean Weight scheme, and the BL62 substitution matrix and Optim10 substitution matrices, and three other string kernel distances, Subseq [23], Spectrum [24], and WDegree, a weighted string kernel with different weights [25, 26]) “True” relationships are defined according to CATH topologies. The analysis is performed on CATH40e4 that contains 10619 sequences corresponding to 1363 folds