Analysis of Multitarget Activities and Assay Interference Characteristics of Pharmaceutically Relevant Compounds

Abstract

The availability of large amounts of data in public repositories provide a useful source of knowledge in the field of drug discovery. Given the increasing sizes of compound databases and volumes of activity data, computational data mining can be used to study different characteristics and properties of compounds on a large scale. One of the major source of identification of new compounds in early phase of drug discovery is high-throughput screening where millions of compounds are tested against many targets. The screening data provides opportunities to assess activity profiles of compounds.<br /> This thesis aims at systematically mining activity data from publicly available sources in order to study the nature of growth of bioactive compounds, analyze multitarget activities and assay interference characteristics of pharmaceutically relevant compounds in context of polypharmacology. In the first study, growth of bioactive compounds against five major target families is monitored over time and compound-scaffold-CSK (cyclic skeleton) hierarchy is applied to investigate structural diversity of active compounds and topological diversity of their scaffolds. The next part of the thesis is based on the analysis of screening data. Initially, extensively assayed compounds are mined from the PubChem database and promiscuity of these compounds is assessed by taking assay frequencies into account. Next, DCM (dark chemical matter) or consistently inactive compounds that have been extensively tested are systematically extracted and their analog relationships with bioactive compounds are determined in order to derive target hypotheses for DCM. Further, PAINS (pan-assay interference compounds) are identified in the extensively tested set of compounds using substructure filters and their assay interference characteristics are studied. Finally, the limitations of PAINS filters are addressed using machine learning models that can distinguish between promiscuous and DCM PAINS. Structural context dependence of PAINS activities is studied by assessing predictions through feature weighting and mapping.