Determining the best set of molecular descriptors for a Toxicity classification problem

Abstract

The safety norms for drug design are very strict with at least three stages of trials. One test, early on in the trials, is about the cardiotoxicity of the molecules, that is, whether the compound blocks any heart channel. Chemical libraries contain millions of compounds. Accurate a priori and in silico classification of non-blocking molecules, can reduce the screening for an effective drug, by half. The compound has to be checked for other risk factors alongside its therapeutic effect; these tests can also be done using a computer. Actual screening in a research laboratory is very expensive and time consuming. To enable the computer modelling, the molecules are provided in Simplified Molecular Input Line Entry (SMILE) format. In this study, they have been decoded using the chem-informatics development kit written in the Java language. The kit is accessed in the R statistical software environment through the rJava package, that is further wrapped in the rcdk package. The strings representing the molecular structure, are parsed by the rcdk functions, to provide structure-activity descriptors, that are known, to be good predictors of biological activity. These descriptors along with the known blocking behaviour of the molecule, constitute the input to the Decision Tree, Random Forest, Gradient Boosting, Support-Vector-Machine, Logistic Regression, and Artificial Neural Network algorithms. This paper reports the results of the data analysis project with shareware tools, to determine the best subset of molecular descriptors, from the large set that is available.