Cairo University, Faculty of Computers & Information Information Technology Department 5 Ahmed Zewal St., Orman, Giza, office box# 12613 (email)
Most of the current Machine Learning applications in cheminformatics
are black box applications. Support vector machine and neural networks are the
most used classification techniques in prediction of the mutagenic activity of compounds.
The problem of these techniques is that the rules/reasons of prediction are
unknown. The rules could show the most important features/descrpitors of the compounds
and the relations among them. This article proposes a model for generating
the rules that governs prediction through the rough set theory. These rules, which
based on two levels of selection for the highly discriminating power features, are
visualized by lattice generated using the formal concept analysis approach. That
is, better understanding of the reasons that leads to the mutagenic activity can be
obtained. The resulted lattice shows that lipophilicity, number of nitrogen atoms,
and electronegativity are the most important parameters in mutagenicity detection.
Moreover, experimental results are compared against previous researches for validating
the proposed model.