Nonparametric, Parametric and Semiparametric Models for Screening and Decoding Pools of Chemical Compounds

Abstract

During High Throughput Screening (HTS), large collections of chemical compounds are tested for potency with respect to one or more assays. In reality, only a very small fraction of the compounds in a collection will be potent enough to act as lead molecules in later drug discovery phases. Testing all compounds is neither cost-effective nor desirable. Based on the belief that chemical structure is highly related to potency of compounds, structure activity relationships (SARs) can be very helpful for selecting a handful of chemical compounds for testing.

This work investigates SARs using four different statistical methods. The first uses a latent class cell-based method. The second benefits from a fractional factorial design for optimizing the cell-based method to significantly increase hit rates. The third improves HTS efficiency by considering pooling experiments for chemical compounds in the presence of interaction and dilution. Rather than testing one compound at a time, chemical compounds are mixed together and tested by groups. Likelihood models are built and hit rates are shown to be higher than for traditional methods. The fourth solves the estimation problem in a pooling experiment by treating the pooling data as missing at random. Semiparametric models are implemented and estimators are shown to be more efficient than likelihood methods based on the same data.