Advances in Consensus Modeling for ADME/Tox Prediction 

The attrition rate of drug candidates in pharmaceutical R&D is a well known problem. In silico ADME/Tox prediction has become an increasing important method to help rank order the drug candidates in the hopes of improving the odds of success earlier in the R&D process. A number of publications and conference proceedings have extolled the virtues of consensus modeling for ADME/Tox, which heretofore has been a manual, laborious process. Consensus modeling involves the use of multiple models simultaneously to improve the result over any of the individual models in the prediction. A system is described that automates the creation and validation of consensus models for both real variable predictors (for example, log P) as well as Boolean variable or classification predictors (for example, mutagenicity).

An integrated N-fold cross-validation process is included for real variable consensus models, and an integrated randomization cross-check process is included for Boolean variables. Both provide statistics that allow the quality of the consensus prediction to be assessed. In addition, two new concepts for consensus modeling will be demonstrated. The first, substructure-localized consensus models (LOCOMOs), allows consensus models to be trained for specific classes of compounds, improving the overall accuracy of prediction. The second, model normalization, involves the use of scaling factors and constants to better map a model to a specific set of experimental data. In all cases, once a consensus model is trained and validated, it can be used with a higher degree of confidence to screen large libraries of compounds.