Multiple-Instance Case-Based Learning for Predictive Toxicology

Predictive toxicology is the task of building models capable of determining, with a certain degree of accuracy, the toxicity of chemical compounds. Machine Learning (ML) in general, and lazy learning techniques in particular, have been applied to the task of predictive toxicology. ML approaches differ in which kind of chemistry knowledge they use but all rely on some specific representation of chemical compounds. In this paper we deal with one specific issue of molecule representation, the multiplicity of descriptions that can be ascribed to a particular compound. We present a new approach to lazy learning, based on the notion of multiple-instance, which is capable of seamlessly working with multiple descriptions. Experimental analysis of this approach is presented using the Predictive Toxicology Challenge data set.

• Type: Conference
• Year: 2004
• Where: KELSI
• Authors: Eva Armengol, Enric Plaza
• Keywords: Chemical Compounds, Informatics, KELSI 2004, Lazy Learning, Predictive Toxicology