|
(M/FF239) Framework for Incorporating QSAR Model Predictions in Decision-Making Process. McDowell, Robert1, Jaworska, Joanna2, 1 2 ABSTRACT- Application of QSAR's in decision-making has been restricted by lack of an appropriate metric for their reliability and uncertainty. Bayesian revision of model predictions has recently been advocated as the appropriate method for incorporating prior knowledge and model performance parameters to quantity the reliability of specific model predictions. This paper extends that concept by applying the concepts of formal decision theory to provide a systematic, widely used methodology to incorporate all aspects of the decision problem in a comprehensive framework. In the context of utilizing QSAR predictions in chemical use decisions, the decision problem is framed by eight components: possible actions by the decision-maker; possible states of the chemical in question; prior probabilities for these states (probabilities the decision-maker holds before obtaining a prediction or forecast); predictions or forecasts of a chemical's true state provided by QSAR model); reliability parameters of the QSAR model (sensitivity and specificity); consequences or net costs and benefits for each decision-maker action/true chemical state combination (referred to as the "payoff matrix"); a choice criterion; and, strategies for future action conditional on model predictions. This Bayesian decision framework incorporates these components and provides a means for rational decision-making and identifying optimal strategies. In addition, the economic value of additional information or improvements in the QSAR model that reduce uncertainty in model output can be computed in this framework. The process is illustrated with a hypothetical decision problem regarding using a chemical in an industrial process based on output from CATABOL, a biodegradation QSAR model. Key words: QSAR, decision theory, CATABOL, Bayes theorem |
