Skip to main content

Table 2 Comparison of kernels for ℓ-mer content with their AA-property enhanced counterparts.

From: Exploiting physico-chemical properties in string kernels

Method auROC50 #Wins
Spectrum (ℓ = 5) 15.2% 7/54
Spectrum-RBF (ℓ = 5, σ = 1) 42.1% 45/54
Mismatch (ℓ = 5, m = 1) 42.3% 13/54
Mismatch-RBF (ℓ = 5,m = 1, σ = 1) 43.6% 36/54
Profile (ℓ = 5, τ = 7.5) 82.1% 3/54
Profile-RBF (ℓ = 5,τ = 7.5, σ = 100) 82.2% 10/54
  1. Comparison of the three kernels proposed in [11, 21, 22], with their AA-property enhanced counterparts for remote homology detection of 54 protein families. auROC50 is the average auROC50 score and #Wins the number of families for which each method outperforms its counterpart (Spectrum vs. Spectrum-RBF, Mismatch vs. Mismatch-RBF, Profile vs. Profile-RBF). The kernels taking advantage of AA properties lead to a higher average accuracy in all three cases (p-Values: 6.92 10−8 for spectrum, 0.0045 for mismatch, and 1.0 for profile kernels). For ℓ and τ we use the published parameter settings. For σ we chose the best result among σ = {0.1,1,10,100,1000}.