April 2 - 4, 2001. San Diego, California, USA, 221st American Chemical Society (ACS) National Meeting

Empirical Intelligence in Physical Property Prediction

Robert S. DeWitte and Eduard Kolovanov

Abstract

Neural Nets, Decision Trees, Genetic Algorithms and other forms of "mid-tier" artificial intelligence have attracted a lot of attention in the past decade for their ability to generate insightful and useful models of all sorts of phenomena, including the physical properties of molecular species. One should be wary, however, of unpredictable behaviour in these models, particularly in cases where input falls outside the domain (loosely speaking) of the training set. Since it is not clear, a priori, how positive behaviour within the domain has been achieved, it may come at the expense of negative behaviour outside the domain. While it is true that every prediction methodology is fundamentally limited by the "experience set" defined by training samples, empirical models parameterized on such training sets have the fundamental advantage of well defined behaviour. This stems from the fact that the essential physics embedded in the empirical model controls the behaviour over all input conditions. Such "Empirical Intelligence" is the basis of the Advanced Chemistry Development Suite of Physical Property predictors. This talk will highlight the advantages of this approach by focusing on the applications of Physical Property Prediction in Drug Discovery Research. Each application demands different rigour from a prediction method: some are well met by any reasonably accurate approach; others demand a systematic methodology based on parameterized physical models. It is in this latter case that "Empirical Intelligence" is more robust.

Download the presentation in PDF format (400 Kb) or MS Power Point 97/2000 format (897 Kb ZIP file).