Comparison of Two Multiple-tree Algorithms on High-throughput Screening Data from Drug Discovery: RandomForest and Partitionator

• Activity Number: 46
• Type: Topic Contributed
• Date/Time: Sunday, August 8, 2004 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Katja S. Remlinger*+ and Jacqueline M. Hughes-Oliver
• Companies: North Carolina State University and North Carolina State University
• Keywords: classification trees ; data-mining ; dimension reduction ; variable importance ; accumulation curves

Abstract

In drug discovery, large chemical libraries are screened to identify active compounds. Screening the entire library is not very cost- or time-efficient. Methods are needed that can predict the biological activity of a compound based on that compound's chemical structure. Data-mining techniques are good candidates for this task. They perform well on large datasets, and they are very flexible. This paper compares two multiple tree algorithms, RandomForest and Partitionator, on a dataset from drug discovery that was used in the KDD Cup 2001. We first give a brief description of both algorithms, point out their differences, and then compare their performance on the KDD dataset. Both algorithms achieve weighted accuracies on predicting the test set activities that are in the top 5% of all competitor results. Furthermore, we propose three different approaches to define the screening order of test set compounds that are suitable for multiple-tree algorithms.