Abstract
This work presents a method to utilize the everโexpanding corporate collections of CYP450 inhibition data to forecast the future risk of compounds not yet synthesized. The global/local fusion method differs from existing QSAR methods, in that each prediction is derived from a customโbuilt QSAR model, constructed onโtheโfly, using a customized training set assembled for each prediction. It uses a consensus of global and local descriptorโbased models along with pharmacophoreโbased fingerprint similarity to form a prediction and to assess the uncertainty of the prediction on a caseโbyโcase basis. We also present a new forward prediction testing and validation scheme in which the corporate dataset is split chronologically, and predictions for a molecule are based on the pool of existing data available before the molecule is registered and tested. The validation accuracy of the CYP2D6 and CYP3A4 models approaches the underlying accuracy of the data, about 0.4 log IC~50~ units standard error (or nearly 70% r^2^ correlation) for the most confident predictions, and extends to about 0.6 log IC~50~ units standard error (or under 30% r^2^ correlation) for the least confident predictions. As a classification model for CYP2D6 and CYP3A4 activity, the validation accuracy is about 79% for predicted actives and 85% for predicted inactives, which is consistent with existing published models.