445
13 Continuous Molecular Fields Approach Applied to Structure-Activity Modeling
with low coverage value lie in the regions of chemical space densely populated by
compounds from the dataset, whereas those with high coverage values lie in nonpopulated areas rather far from other compounds. One can see that starting from the
coverage 0.9 the mean absolute cross-validation error of compounds outside AD is
considerably higher than that of compounds within AD. This means that in this case
the DM corresponding to coverage 0.9 can be used as a threshold for defining the
AD of CMF 3D-QSAR model.
13.3.2 Virtual Screening via Combination CMF with 1-SVM
Technique
As it follows from our earlier publications, one-class classification (novelty detection) machine learning methods is a mathematical base of a new general approach
to conducting ligand-based virtual screening of chemical compounds [41, 42]. Although several dozens of different algorithms for building one-class classification
(novelty detection) models are known [43, 44], only those of them which are based
on using kernels are suited for working with continuous molecular fields. The OneClass Support Vector Machines ( 1-SVM) [12] machine learning method is one of
Fig. 13.1  Plot of the mean absolute cross-validation error (calculated for compounds outside AD)
vs coverage for the thrombin dataset
0.0
0.2
0.4
0.6
0.8
1.0
0.2
0.3
0.4
0.5
0.6
0.7
Coverage - Mean Absolute Error Plot
Coverage
Mean Absolute Error