SVM Prediction Using Adaptive Binary Kernels

• Activity Number: 392
• Type: Invited
• Date/Time: Wednesday, August 6, 2003 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Joseph S. Verducci*+ and Michael A. Fligner and Paul Blower
• Companies: The Ohio State University and The Ohio State University and Leadscope Inc.
• Keywords: classification ; discrimination ; statistical learning ; high-dimensional data

Abstract

Chemical databases often encode the structure of molecules in long binary strings called fingerprints. A general goal is to use these fingerprints to predict some specific biological activity of a molecule, such as its ability to kill certain cancer cells. It is well known that different classes of chemicals interact with cells via different mechanisms, and that small structural differences within a class can produces large changes in biological activity. Under these circumstances, simple implementations of support vector machines do not perform well. However, recent work (Wilton et al. 2002) suggests that some specialized kernel smoothers may work well in distinguishing biologically active molecules. Our approach is first to form localized regions using the Jacard-Tanimoto metric, which is sensitive to the relative number of mismatched features, and then use "weighted triples" kernels within local neighborhood. These utilize low order interactions of binary features in creating the underlying kernel function. The technique is illustrated by identifying key feature-combinations of a subclass of colchicines with high activity against H23 lung cancer cells.