Abstract

The quality of in vitro data used to build in silico absorption, distribution, metabolism, and toxicity (ADMET) models is, in many cases, inconsistent. The paucity of data from single laboratory sources has led to the mixing of data sets with varying experimental conditions and to the coverage of restricted chemical space in models which are purported to be of general applicability. In order to overcome these shortcomings, a method, Metropolis/Monte Carlo adaptive ranking simulation (MARS) has been developed. This aims to estimate “optimal flexible threshold points” in order to achieve better correlation between any in silico ADMET model and any discrete qualitative experimental data. The MARS method covers three key factors: the predictive model, the experimental procedure for the assay, and the chemical series or scaffold. When large and general solubility data sets (>650 compounds) are analyzed against commercially available in silico models, using MARS, an improvement in κ statistics up to 16.2% is obtained. When particular chemical series are addressed, improvements up to 46.0% are seen on κ statistics. This coefficient then allows an investigation into the effectiveness of a classifier by assessing the improvement over chance. These improvements in ranking estimations allow more predictive decision-making for virtual libraries.