Statistical learning and virtual screening in drug discovery

Hugh Chipman, Canada Research Chair in Mathematical Modeling, Dept of Mathematics and Statistics, Acadia University, Wolfville, N.S. High-throughput screening of compounds for biological activity is often an important first step in the drug discovery process. From a statistical learning perspective, the results of screening process can be used to construct a model. Using various descriptors of molecular structure as inputs, we seek to predict activity. These descriptors can be easily calculated, but the activity is the outcome of more expensive screening procedures. Screening results for a part of the library constitute a training set, which can be used to build a model to predict activity. This model enables "virtual screening" in which activity is predicted rather than measured. This talk describes a number of recent models developed for such virtual screening, including mixture discriminant analysis, decision trees, nearest neighbours, and ensemble models.