Automated Visualization and Discovery of Predictively Useful Biological Response Profiles in Complex Data Sets. Louis Anthony Cox, Jr., Cox Associates, 503 Franklin Street, Denver, CO 80218; and Michael Bird, Exxon Biomedical Sciences, Inc., Mettlers Road, CN 2350, East Millstone, NJ 08875
Risk analysts must frequently confront the following practical problems: (i) How to integrate multiple, heterogeneous sources of evidence, e.g., from in vitro, in vivo, and QSAR studies; (ii) How to draw sound and useful risk inferences from data sets that are incomplete (e.g., containing missing or censored observations), inaccurate, and/ or inconsistent; and (iii) How to visualize and explore complex, multiway interactions among multiple risk factors, such as concentration, hours-per-day, and number of days of exposure, and responses at many different anatomic sites. We present an approach for simultaneously overcoming these problems using an integrated set of ideas from information theory, Monte-Carlo resampling of small data sets, and artificial intelligence rule induction from structured data bases. This approach, called the STEM methodology (since it characterizes risks in terms of Source, Target, Effect, and Mechanism) has proved powerful for discovering predictively useful patterns in chemical health effects data bases. We illustrate it using two examples: a STEM analysis of an immunotoxicity data base and a STEM analysis of bioassay data for isoprene. STEM's ability to make useful, cost-effective risk predictions even when data are costly or unavailable is demonstrated. The value of experimental information, measured in terms of reductions in expected prediction error rates, is quantified as a part of STEM's approach to dealing with uncertainties in risk analysis. Finally, several unexpected patterns that STEM discovered in the isoprene data set are presented, including highly asymmetric effects on tumor risks of exposure concentration and duration, as well as positive associations between lung and liver tumors. These discoveries, which were subsequently confirmed by more traditional statistical methods, illustrate the relevance of the STEM approach for analyzing dose-response patterns for multi-site carcinogens.