Quantitative Structure-Activity-Relationship Modeling Using Leo Breiman's Random Forest

• Activity Number: 86
• Type: Contributed
• Date/Time: Monday, August 4, 2003 : 8:30 AM to 10:20 AM
• Sponsor: Section on Physical and Engineering Sciences
• Author(s): Christopher H. Tong*+ and Vladimir B. Svetnik and Andy I. Liaw
• Companies: Merck & Co., Inc. and Merck and Company and Merck & Co., Inc.
• Keywords: Random Forest ; classification ; regression ; tree ; data mining ; drug discovery

Abstract

Quantitative structure-activity-relationship (QSAR) models relate a measure of a molecule's biological activity with its chemical structure. QSAR continues to be a very active area of research in chemistry and statistics. Unfortunately, most of the existing tools fail to incorporate multiple mechanisms of action (MOA), and therefore have seen limited application in drug discovery. Recent progress, specifically in ensemble learning for classification and regression, has revealed new opportunities. We present one example, Leo Breiman's Random Forest, with which we had some success. Its advantages include the ability to handle multiple MOAs, assessment of variable importance, and a meaningful measure of molecular similarity. Application of random forest is illustrated by an example of QSAR for P-glycoprotein transport, using a publicly available data set. Our results compare favorably with previous methods. We give particular attention to variable selection and demonstrate potential pitfalls of a naive approach. We also discuss the R language interface, available on CRAN, that we have developed for Breiman and Cutler's Random Forest Fortran software.