October 21, 2004, 1^st Annual PhysChem Symposium: Early ADME and Medicinal Chemistry

Techniques for Improving LogP Prediction

Karim Kassam, Fedor Gorohov, Eduard Kolovanov

Abstract

Predicting LogP has been a topic of study for over forty years. During this time many approaches have been put forward, including the popular fragmental approach which is the basis for the cLogP and ACD/LogP approaches. The properties of this approach give it excellent reliability within the space of well characterized chemical structures. In an earlier publication (Petrauskas and Kolovanov, 2000), the secondary algorithms within ACD/Labs were explained, which allow the algorithm a measure of extrapolation beyond the fragments specifically within the algorithm's experience. Nonetheless, there are circumstances where the degree of novelty in the chemical class is so high that these extrapolations are beyond the limit of reasonable caution. This presentation will describe how to recognize these situations, and two techniques for using targeted measurements to train the algorithm for these new classes of compounds: System Training for the case of small numbers of representative compounds; Accuracy Extender for the case of a statistical sampling of class members.

Download the presentation in MS PowerPoint (802 Kb ZIP file) or Adobe Acrobat format (742 Kb PDF file).

Relevant Products: LogP, Solubility, LogD