Challenges in predictive ADMET: imperfect data in an imperfect world

COMP 190

Stephen R Johnson, Computer-Assisted Drug Design, Computer-Assisted Drug Design, Bristol-Myers Squibb, P.O. Box 4000, Princeton, NJ 08543-4000
Great effort is being allocated to the in silico and in vitro prediction of ADMET related activities. There is a growing industry consensus, however, that these efforts fall somewhat short in quantitative predictive ability. This presentation presents an overview of the factors contributing to these problems, including biological complexity, assay reproducibility, and in vitro-in vivo relationships. We will present a simple method to understand the extent to which a model can be expected to be predictive based on the underlying data used to generate it. Finally, an example of a model generated using noisy data will be used to illustrate many of these points. The relative utility of the model will be demonstrated for use in compound design.