The Impact of Extraneous Descriptors on Computational Models

• Activity Number: 595
• Type: Invited
• Date/Time: Thursday, August 4, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Physical and Engineering Sciences
• Author(s): Kjell Johnson*+
• Companies: Pfizer Inc.
• Keywords: predictive modeling ; computational chemistry ; extraneous descriptors ; model performance

Abstract

An objective of cheminformatics is to understand the relationship between chemical structures and endpoints of interest such as target activity or safety. This understanding can help scientists modify molecules to improve activity or decrease safety risks. Chemical structures are often described by their physical properties, or by their representations in either 2 or 3-dimensions. Numerous software packages can generate different types of these descriptors, and thousands of descriptors can easily be computed for any set of compounds.

In an attempt to find the most predictive model, practitioners often try many different types of descriptors. A common practice is to include more descriptors in a model in an attempt to find the most predictive model, even if only a small fraction of the descriptors are relevant to the response.

This presentation will explore the effect of the inclusion of extraneous descriptors on the predictive ability of many traditional and modern modeling methods (LM, PLS, GBM, RF, NN, SVM, etc.). While the predictive ability of some methods is only slightly affected by extraneous descriptors, the ability of other models is significantly decreased.