Flexible SAR Prediction System using KNIME

Abstract

This research consists of 2 subjects. The first is the development of SAR prediction system. New notation has been introduced to linear fragments, such as branching and variable chain length. Descriptor selection step uses Relief algorithm from a group of correlated fragments. Prediction model is based on the cascade model. A few rules have been selected based on the rule priority definition. No prediction has been done when no rules are applicable to a compound. Results are judged by AUC of ROC. Application to rodent carcinogenicity prediction showed better AUC than those given by Naive Bayes and Random Forests methods. The second part reveals the development of prediction system on KNIME environment. We have converted the above prediction system onto KNIME. The visual workflow has enabled easy understanding of the system. We could substitute a program to a KNIME node, and a python code has been implemented by a KNIME Jython node. The resulting system has given us a flexible SAR prediction environment and we can easily compare prediction results by a variety of methods.