Computational Methods for the Integration of Biological Activity and Chemical Space

One general aim of medicinal chemistry is the understanding of structure-activity relationships of ligands that bind to biological targets. Advances in combinatorial chemistry and biological screening technologies allow the analysis of ligand-target relationships on a large-scale. However, in order to extract useful information from biological activity data, computational methods are needed that link activity of ligands to their chemical structure. In this thesis, it is investigated how fragment-type descriptors of molecular structure can be used in order to create a link between activity and chemical ligand space. First, an activity class-dependent hierarchical fragmentation scheme is introduced that generates fragmentation pathways that are aligned using established methodologies for multiple alignment of biological sequences. These alignments are then used to extract consensus fragment sequences that serve as a structural signature for individual biological activity classes. It is also investigated how defined, chemically intuitive molecular fragments can be organized based on their topological environment and co-occurrence in compounds active against closely related targets. Therefore, the Topological Fragment Index is introduced that quantifies the topological environment complexity of a fragment in a given molecule, and thus goes beyond fragment frequency analysis. Fragment dependencies have been established on the basis of common topological environments, which facilitates the identification of activity class-characteristic fragment dependency pathways that describe fragment relationships beyond structural resemblance. Because fragments are often dependent on each other in an activity class-specific manner, the importance of defined fragment combinations for similarity searching is further assessed. Therefore, Feature Co-occurrence Networks are introduced that allow the identification of feature cliques characteristic of individual activity classes. Three differently designed molecular fingerprints are compared for their ability to provide such cliques and a clique-based similarity searching strategy is established. For molecule- and activity class-centric fingerprint designs, feature combinations are shown to improve similarity search performance in comparison to standard methods. Moreover, it is demonstrated that individual features can form activity-class specific combinations. Extending the analysis of feature cliques characteristic of individual activity classes, the distribution of defined fragment combinations among several compound classes acting against closely related targets is assessed. Fragment Formal Concept Analysis is introduced for flexible mining of complex structure-activity relationships. It allows the interactive assembly of fragment queries that yield fragment combinations characteristic of defined activity and potency profiles. It is shown that pairs and triplets, rather than individual fragments distinguish between different activity profiles. A classifier is built based on these fragment signatures that distinguishes between ligands of closely related targets. Going beyond activity profiles, compound selectivity is also analyzed. Therefore, Molecular Formal Concept Analysis is introduced for the systematic mining of compound selectivity profiles on a whole-molecule basis. Using this approach, structurally diverse compounds are identified that share a selectivity profile with selected template compounds. Structure-selectivity relationships of obtained compound sets are further analyzed