Analysis of Biological Screening Data and Molecular Selectivity Profiles Using Fingerprints and Mapping Algorithms

Abstract

<p>The identification of promising drug candidates is a major milestone in the early stages of drug discovery and design. Among the properties that have to be optimized before a drug candidate is admitted to clinical testing, potency and target selectivity are of great interest and can be addressed very early. Unfortunately, optimization–relevant knowledge is often limited, and the analysis of noisy and heterogeneous biological screening data with standard methods like QSAR is hardly feasible. Furthermore, the identification of compounds displaying different selectivity patterns against related targets is a prerequisite for chemical genetics and genomics applications, allowing to specifically interfere with functions of individual members of protein families. In this thesis it is shown that computational methods based on molecular similarity are suitable tools for the analysis of compound potency and target selectivity. Originally developed to facilitate the efficient discovery of active compounds by means of virtual screening of compound libraries, these ligand–based approaches assume that similar molecules are likely to exhibit similar properties and biological activities based on the <i>similarity property principle</i>. Given their holistic approach to molecular similarity analysis, ligand–based virtual screening methods can be applied when little or no structure– activity information is available and do not require the knowledge of the target structure.<br /> The methods under investigation cover a wide methodological spectrum and only rely on properties derived from one– and two–dimensional molecular representations, which renders them particularly useful for handling large compound libraries. Using biological screening data, these virtual screening methods are shown to be able to extrapolate from experimental data and preferentially detect potent compounds. Subsequently, extensive benchmark calculations prove that existing 2D molecular fingerprints and dynamic mapping algorithms are suitable tools for the distinction between compounds with differential selectivity profiles. Finally, an advanced dynamic mapping algorithm is introduced that is able to generate target–selective chemical reference spaces by adaptively identifying most–discriminative molecular properties from a set of active compounds. These reference spaces are shown to be of great value for the generation of predictive target–selectivity models by screening a biologically annotated compound library. </p>