When applied to drug discovery, modern computational systems can provide insight into the highly complex systems underlying drug activity and predict compounds or targets of interest. Many tools have been developed for computer aided drug discovery (CADD), focusing on small molecule ligands, protein targets, or both. The aim of this thesis is the improvement of CADD tools for describing small molecule properties and application of CADD to several stages of drug discovery regarding two targets for the treatment of obesity and related diseases: the neuropeptide Y4 receptor (Y4R) and the melanocortin-4 receptor (MC4R). In the first chapter, the major categories of CADD are outlined, including descriptions for many of the popular tools and examples where these tools have directly contributed to the discovery of new drugs. Following the introduction, several improvements for encoding stereochemistry and signed property distribution are introduced and tested in scenarios meant to simulate applications in virtual high-throughput screening. Y4R and MC4R are both class A G-protein coupled receptors (GPCRs) with endogenous peptide ligands that play critical roles in the signaling of satiety and energy metabolism. So far, no structures from either receptor family have been experimentally elucidated. CADD was combined with high-throughput screening (HTS) to discover the first small molecule positive allosteric modulators (PAMs) of Y4R. Secondly, CADD techniques were used to model the interaction of Y4R and pancreatic polypeptide based on experimental results that elucidate specific binding contacts. Similar SB-CADD approaches were used to model the interaction of MC4R with its high affinity peptide agonist α-MSH. Due to its role in monogenic forms of obesity, these models were used to predict which residues directly participate in binding and correlate mutated residues with their potential role in the binding site.

Die Bestimmung der Ähnlichkeit von molekularen Strukturen und das Clustern solcher Strukturen gemäß Ähnlichkeit sind zwei zentrale Fragen im Wirkstoffdesign. Die Arbeit beschreibt im ersten Teil zwei neue Verfahren zum Vergleich von Molekülen auf 3-dimensionale Ähnlichkeit. Der erste Algorithmus benutzt als Eingabe nur die Koordinaten der Atome der zu vergleichenden Moleküle. Wir können zeigen, daß eine rein geometrische Zielfunktion in der Lage ist, Wirkungsähnlichkeit von Substanzen vorherzusagen, und daß der Algorithmus geeignet ist, Ähnlichkeiten zu finden, die mit bisherigen, einfacheren Methoden nicht gefunden werden konnten. Das zweite Verfahren nutzt zusätzlich noch die Bindungsstruktur der Eingabemoleküle. Es ist flexibel, d.h. alle Konformere der Moleküle werden simultan behandelt. Wir erhalten ein sehr schnelles Verfahren, das bei geeigneter Parametereinstellung auch beweisbar optimale Lösungen liefert. Für praktisch relevante Anwendungen erreichen wir erstmals Laufzeiten, die selbst das Durchsuchen großer Datenbanken ermöglichen. Im zweiten Teil beschreiben wir zwei Methoden, eine Menge von molekularen Strukturen so zu organisieren, daß die Suche nach geometrisch ähnlichen deutlich schneller durchgeführt werden kann als durch lineare Suche. Nach Analyse der Daten mit graphentheoretischen Methoden finden hierarchische Verfahren und repräsentantenbasierte Ansätze ihre Anwendung. Schließlich geben wir einen neuen Algorithmus zum Biclustern von Daten an, einem Problem, das bei der Analyse von Genexpressionsdaten eine wichtige Rolle spielt. Mit graphentheoretischen Methoden konstruieren wir zunächst deterministisch Obermengen von Lösungen, die danach heuristisch ausgedünnt werden. Wir können zeigen, daß dieser neue Ansatz bisherige, vergleichbare z.T. deutlich überbietet. Seine prinzipielle Einfachheit läßt anwendungsbezogene Modifikationen leicht zu. / Two important questions in drug design are the following: "How to compute the similarity of two molecules?" and "How to cluster molecules by similarity?" In the first part we describe two different approaches to compare molecules for 3D-similarity. The first algorithm just uses the 3D coordinates of the atoms as input. We show that this algorithm is able to detect similar activity or similar adverse reaction, even with a simple purely geometry based scoring function. Compared to previous simpler approaches more interesting hits are found. The connectivity structures of the molecular graphs are used by the second algorithm as additional input. This fully flexible approach -- conformers of the molecules are treated simultaneously -- may even find provably optimal solutions. Parameter settings for practically relevant instances allow running times that make it possible to even search large databases. The second part describes two methods to search a database of molecular structures. After analyzing the data with graph theoretical methods two algorithms for two different ranges of similarity are designed. Scanning the database for structures similar to a given query can be accelerated considerably. We use hierarchical methods and dominating set techniques. Finally we propose a new biclustering algorithm. Biclustering problems recently appeared mainly in the context of analysing gene expression data. Again graph theoretical methods are our main tools. In our model biclusters correspond to dense subgraphs of certain bipartite graphs. In a first phase the algorithm deterministically finds supersets of solution candidates. Thinning out these sets by heuristical methods leads to solutions. This new algorithm outperforms former comparable methods and its simple structure make it easy to customize it for practical applications.

When applied to drug discovery, modern computational systems can provide insight into the highly complex systems underlying drug activity and predict compounds or targets of interest. Many tools have been developed for computer aided drug discovery (CADD), focusing on small molecule ligands, protein targets, or both. The aim of this thesis is the improvement of CADD tools for describing small molecule properties and application of CADD to several stages of drug discovery regarding two targets for the treatment of obesity and related diseases: the neuropeptide Y4 receptor (Y4R) and the melanocortin-4 receptor (MC4R). In the first chapter, the major categories of CADD are outlined, including descriptions for many of the popular tools and examples where these tools have directly contributed to the discovery of new drugs. Following the introduction, several improvements for encoding stereochemistry and signed property distribution are introduced and tested in scenarios meant to simulate applications in virtual high-throughput screening. Y4R and MC4R are both class A G-protein coupled receptors (GPCRs) with endogenous peptide ligands that play critical roles in the signaling of satiety and energy metabolism. So far, no structures from either receptor family have been experimentally elucidated. CADD was combined with high-throughput screening (HTS) to discover the first small molecule positive allosteric modulators (PAMs) of Y4R. Secondly, CADD techniques were used to model the interaction of Y4R and pancreatic polypeptide based on experimental results that elucidate specific binding contacts. Similar SB-CADD approaches were used to model the interaction of MC4R with its high affinity peptide agonist α-MSH. Due to its role in monogenic forms of obesity, these models were used to predict which residues directly participate in binding and correlate mutated residues with their potential role in the binding site.