Solubility prediction by recursive partitioning.

Abstract

To build and test a computational model for predicting small molecule solubility, to improve the cost-effectiveness of the selection of vendor compounds suitable for nuclear magnetic resonance (NMR) screening. A simple recursive partitioning decision tree-based classification model was generated utilizing "off-the-shelf" commercial software from Accelrys Inc., with a training set of 1992 compounds based on a series of calculated topologic and physical properties. The predictive ability of the decision tree was then assessed by employing it to classify a test set of 2851 vendor compounds, and the classification was subsequently used to guide the purchase of 686 compounds for the purpose of NMR screening. When the decision tree was used to guide purchasing, the percentage of "acceptable" compounds suitable for NMR screening doubled compared with the use of a simple cLogP cutoff, improving the successful selection rate from 25% to 50%. A simple recursive partitioning decision tree may successfully be used to improve cost-effectiveness by reducing the wastage associated with the unnecessary purchase of vendor compounds unsuitable for NMR screening because of insolubility.