Descriptor vector redesign by neuro-fuzzy analysis

Virtual screening is an important challenge for bioinformaticians and molecular biologists. Finding the most active ligands that bind to target molecules by intelligent methods reduces the computational efforts in the drug discovery process dramatically. Recently, we proposed a neuro-fuzzy based approach to virtual screening by generating interval rules, each one with a class assignment for enriched regions of active ligands. The underlying input data is based on so called descriptor vector calculations. Descriptors are designed by experts, and they represent properties of the molecules. It is not assumed that all descriptors in a usually high dimensional descriptor vector are needed for classification. We demonstrate the usefulness of a neuro-fuzzy system for rule-based classification and feature selection. On the one hand the classification results can be used for efficient virtual screening, and on the other hand the rules can be used by experts for the ongoing drug design process. By feature analysis of the rules descriptor vectors can be redesigned. In the future more tools are needed that are capable not only of achieving several analysis results but giving also rule-based hints to the experts of how to plan their further research activities.