Refine
Document Type
- Doctoral Thesis (11)
Language
- English (11) (remove)
Has Fulltext
- yes (11)
Is part of the Bibliography
- no (11)
Keywords
- Computational chemistry (3)
- Arzneimitteldesign (2)
- virtual screening (2)
- Alignment <Biochemie> (1)
- Bayes-Lernen (1)
- Coenzym (1)
- Cofaktor (1)
- De-Novo-Synthese (1)
- Deterministische Optimierung (1)
- Dimensionsreduktion (1)
Institute
- Biowissenschaften (6)
- Biochemie und Chemie (3)
- Pharmazie (2)
Virtual screening of potential bioactive substances using the support vector machine approach
(2005)
Die vorliegende Dissertation stellt eine kumulative Arbeit dar, die in insgesamt acht wissenschaftlichen Publikationen (fünf publiziert, zwei eingerichtet und eine in Vorbereitung) dargelegt ist. In diesem Forschungsprojekt wurden Anwendungen von maschinellem Lernen für das virtuelle Screening von Moleküldatenbanken durchgeführt. Das Ziel war primär die Einführung und Überprüfung des Support-Vector-Machine (SVM) Ansatzes für das virtuelle Screening nach potentiellen Wirkstoffkandidaten. In der Einleitung der Arbeit ist die Rolle des virtuellen Screenings im Wirkstoffdesign beschrieben. Methoden des virtuellen Screenings können fast in jedem Bereich der gesamten pharmazeutischen Forschung angewendet werden. Maschinelles Lernen kann einen Einsatz finden von der Auswahl der ersten Moleküle, der Optimierung der Leitstrukturen bis hin zur Vorhersage von ADMET (Absorption, Distribution, Metabolism, Toxicity) Eigenschaften. In Abschnitt 4.2 werden möglichen Verfahren dargestellt, die zur Beschreibung von chemischen Strukturen eingesetzt werden können, um diese Strukturen in ein Format zu bringen (Deskriptoren), das man als Eingabe für maschinelle Lernverfahren wie Neuronale Netze oder SVM nutzen kann. Der Fokus ist dabei auf diejenigen Verfahren gerichtet, die in der vorliegenden Arbeit verwendet wurden. Die meisten Methoden berechnen Deskriptoren, die nur auf der zweidimensionalen (2D) Struktur basieren. Standard-Beispiele hierfür sind physikochemische Eigenschaften, Atom- und Bindungsanzahl etc. (Abschnitt 4.2.1). CATS Deskriptoren, ein topologisches Pharmakophorkonzept, sind ebenfalls 2D-basiert (Abschnitt 4.2.2). Ein anderer Typ von Deskriptoren beschreibt Eigenschaften, die aus einem dreidimensionalen (3D) Molekülmodell abgeleitet werden. Der Erfolg dieser Beschreibung hangt sehr stark davon ab, wie repräsentativ die 3D-Konformation ist, die für die Berechnung des Deskriptors angewendet wurde. Eine weitere Beschreibung, die wir in unserer Arbeit eingesetzt haben, waren Fingerprints. In unserem Fall waren die verwendeten Fingerprints ungeeignet zum Trainieren von Neuronale Netzen, da der Fingerprintvektor zu viele Dimensionen (~ 10 hoch 5) hatte. Im Gegensatz dazu hat das Training von SVM mit Fingerprints funktioniert. SVM hat den Vorteil im Vergleich zu anderen Methoden, dass sie in sehr hochdimensionalen Räumen gut klassifizieren kann. Dieser Zusammenhang zwischen SVM und Fingerprints war eine Neuheit, und wurde von uns erstmalig in die Chemieinformatik eingeführt. In Abschnitt 4.3 fokussiere ich mich auf die SVM-Methode. Für fast alle Klassifikationsaufgaben in dieser Arbeit wurde der SVM-Ansatz verwendet. Ein Schwerpunkt der Dissertation lag auf der SVM-Methode. Wegen Platzbeschränkungen wurde in den beigefügten Veröffentlichungen auf eine detaillierte Beschreibung der SVM verzichtet. Aus diesem Grund wird in Abschnitt 4.3 eine vollständige Einführung in SVM gegeben. Darin enthalten ist eine vollständige Diskussion der SVM Theorie: optimale Hyperfläche, Soft-Margin-Hyperfläche, quadratische Programmierung als Technik, um diese optimale Hyperfläche zu finden. Abschnitt 4.3 enthält auch eine Diskussion von Kernel-Funktionen, welche die genaue Form der optimalen Hyperfläche bestimmen. In Abschnitt 4.4 ist eine Einleitung in verschiede Methoden gegeben, die wir für die Auswahl von Deskriptoren genutzt haben. In diesem Abschnitt wird der Unterschied zwischen einer „Filter“- und der „Wrapper“-basierten Auswahl von Deskriptoren herausgearbeitet. In Veröffentlichung 3 (Abschnitt 7.3) haben wir die Vorteile und Nachteile von Filter- und Wrapper-basierten Methoden im virtuellen Screening vergleichend dargestellt. Abschnitt 7 besteht aus den Publikationen, die unsere Forschungsergebnisse enthalten. Unsere erste Publikation (Veröffentlichung 1) war ein Übersichtsartikel (Abschnitt 7.1). In diesem Artikel haben wir einen Gesamtüberblick der Anwendungen von SVM in der Bio- und Chemieinformatik gegeben. Wir diskutieren Anwendungen von SVM für die Gen-Chip-Analyse, die DNASequenzanalyse und die Vorhersage von Proteinstrukturen und Proteininteraktionen. Wir haben auch Beispiele beschrieben, wo SVM für die Vorhersage der Lokalisation von Proteinen in der Zelle genutzt wurden. Es wird dabei deutlich, dass SVM im Bereich des virtuellen Screenings noch nicht verbreitet war. Um den Einsatz von SVM als Hauptmethode unserer Forschung zu begründen, haben wir in unserer nächsten Publikation (Veröffentlichung 2) (Abschnitt 7.2) einen detaillierten Vergleich zwischen SVM und verschiedenen neuronalen Netzen, die sich als eine Standardmethode im virtuellen Screening etabliert haben, durchgeführt. Verglichen wurde die Trennung von wirstoffartigen und nicht-wirkstoffartigen Molekülen („Druglikeness“-Vorhersage). Die SVM konnte 82% aller Moleküle richtig klassifizieren. Die Klassifizierung war zudem robuster als mit dreilagigen feedforward-ANN bei der Verwendung verschiedener Anzahlen an Hidden-Neuronen. In diesem Projekt haben wir verschiedene Deskriptoren zur Beschreibung der Moleküle berechnet: Ghose-Crippen Fragmentdeskriptoren [86], physikochemische Eigenschaften [9] und topologische Pharmacophore (CATS) [10]. Die Entwicklung von weiteren Verfahren, die auf dem SVM-Konzept aufbauen, haben wir in den Publikationen in den Abschnitten 7.3 und 7.8 beschrieben. Veröffentlichung 3 stellt die Entwicklung einer neuen SVM-basierten Methode zur Auswahl von relevanten Deskriptoren für eine bestimmte Aktivität dar. Eingesetzt wurden die gleichen Deskriptoren wie in dem oben beschriebenen Projekt. Als charakteristische Molekülgruppen haben wir verschiedene Untermengen der COBRA Datenbank ausgewählt: 195 Thrombin Inhibitoren, 226 Kinase Inhibitoren und 227 Faktor Xa Inhibitoren. Es ist uns gelungen, die Anzahl der Deskriptoren von ursprünglich 407 auf ungefähr 50 zu verringern ohne signifikant an Klassifizierungsgenauigkeit zu verlieren. Unsere Methode haben wir mit einer Standardmethode für diese Anwendung verglichen, der Kolmogorov-Smirnov Statistik. Die SVM-basierte Methode erwies sich hierbei in jedem betrachteten Fall als besser als die Vergleichsmethoden hinsichtlich der Vorhersagegenauigkeit bei der gleichen Anzahl an Deskriptoren. Eine ausführliche Beschreibung ist in Abschnitt 4.4 gegeben. Dort sind auch verschiedene „Wrapper“ für die Deskriptoren-Auswahl beschrieben. Veröffentlichung 8 beschreibt die Anwendung von aktivem Lernen mit SVM. Die Idee des aktiven Lernens liegt in der Auswahl von Molekülen für das Lernverfahren aus dem Bereich an der Grenze der verschiedenen zu unterscheidenden Molekülklassen. Auf diese Weise kann die lokale Klassifikation verbessert werden. Die folgenden Gruppen von Moleküle wurden genutzt: ACE (Angiotensin converting enzyme), COX2 (Cyclooxygenase 2), CRF (Corticotropin releasing factor) Antagonisten, DPP (Dipeptidylpeptidase) IV, HIV (Human immunodeficiency virus) protease, Nuclear Receptors, NK (Neurokinin receptors), PPAR (peroxisome proliferator-activated receptor), Thrombin, GPCR und Matrix Metalloproteinasen. Aktives Lernen konnte die Leistungsfähigkeit des virtuellen Screenings verbessern, wie sich in dieser retrospektiven Studie zeigte. Es bleibt abzuwarten, ob sich das Verfahren durchsetzen wird, denn trotzt des Gewinns an Vorhersagegenauigkeit ist es aufgrund des mehrfachen SVMTrainings aufwändig. Die Publikationen aus den Abschnitten 7.5, 7.6 und 7.7 (Veröffentlichungen 5-7) zeigen praktische Anwendungen unserer SVM-Methoden im Wirkstoffdesign in Kombination mit anderen Verfahren, wie der Ähnlichkeitssuche und neuronalen Netzen zur Eigenschaftsvorhersage. In zwei Fällen haben wir mit dem Verfahren neuartige Liganden für COX-2 (cyclooxygenase 2) und dopamine D3/D2 Rezeptoren gefunden. Wir konnten somit klar zeigen, dass SVM-Methoden für das virtuelle Screening von Substanzdatensammlungen sinnvoll eingesetzt werden können. Es wurde im Rahmen der Arbeit auch ein schnelles Verfahren zur Erzeugung großer kombinatorischer Molekülbibliotheken entwickelt, welches auf der SMILES Notation aufbaut. Im frühen Stadium des Wirstoffdesigns ist es wichtig, eine möglichst „diverse“ Gruppe von Molekülen zu testen. Es gibt verschiedene etablierte Methoden, die eine solche Untermenge auswählen können. Wir haben eine neue Methode entwickelt, die genauer als die bekannte MaxMin-Methode sein sollte. Als erster Schritt wurde die „Probability Density Estimation“ (PDE) für die verfügbaren Moleküle berechnet. [78] Dafür haben wir jedes Molekül mit Deskriptoren beschrieben und die PDE im N-dimensionalen Deskriptorraum berechnet. Die Moleküle wurde mit dem Metropolis Algorithmus ausgewählt. [87] Die Idee liegt darin, wenige Moleküle aus den Bereichen mit hoher Dichte auszuwählen und mehr Moleküle aus den Bereichen mit niedriger Dichte. Die erhaltenen Ergebnisse wiesen jedoch auf zwei Nachteile hin. Erstens wurden Moleküle mit unrealistischen Deskriptorwerten ausgewählt und zweitens war unser Algorithmus zu langsam. Dieser Aspekt der Arbeit wurde daher nicht weiter verfolgt. In Veröffentlichung 6 (Abschnitt 7.6) haben wir in Zusammenarbeit mit der Molecular-Modeling Gruppe von Aventis-Pharma Deutschland (Frankfurt) einen SVM-basierten ADME Filter zur Früherkennung von CYP 2C9 Liganden entwickelt. Dieser nichtlineare SVM-Filter erreichte eine signifikant höhere Vorhersagegenauigkeit (q2 = 0.48) als ein auf den gleichen Daten entwickelten PLS-Modell (q2 = 0.34). Es wurden hierbei Dreipunkt-Pharmakophordeskriptoren eingesetzt, die auf einem dreidimensionalen Molekülmodell aufbauen. Eines der wichtigen Probleme im computerbasierten Wirkstoffdesign ist die Auswahl einer geeigneten Konformation für ein Molekül. Wir haben versucht, SVM auf dieses Problem anzuwenden. Der Trainingdatensatz wurde dazu mit jeweils mehreren Konformationen pro Molekül angereichert und ein SVM Modell gerechnet. Es wurden anschließend die Konformationen mit den am schlechtesten vorhergesagten IC50 Wert aussortiert. Die verbliebenen gemäß dem SVM-Modell bevorzugten Konformationen waren jedoch unrealistisch. Dieses Ergebnis zeigt Grenzen des SVM-Ansatzes auf. Wir glauben jedoch, dass weitere Forschung auf diesem Gebiet zu besseren Ergebnissen führen kann.
We investigate the utility of modern kernel-based machine learning methods for ligand-based virtual screening. In particular, we introduce a new graph kernel based on iterative graph similarity and optimal assignments, apply kernel principle component analysis to projection error-based novelty detection, and discover a new selective agonist of the peroxisome proliferator-activated receptor gamma using Gaussian process regression. Virtual screening, the computational ranking of compounds with respect to a predicted property, is a cheminformatics problem relevant to the hit generation phase of drug development. Its ligand-based variant relies on the similarity principle, which states that (structurally) similar compounds tend to have similar properties. We describe the kernel-based machine learning approach to ligand-based virtual screening; in this, we stress the role of molecular representations, including the (dis)similarity measures defined on them, investigate effects in high-dimensional chemical descriptor spaces and their consequences for similarity-based approaches, review literature recommendations on retrospective virtual screening, and present an example workflow. Graph kernels are formal similarity measures that are defined directly on graphs, such as the annotated molecular structure graph, and correspond to inner products. We review graph kernels, in particular those based on random walks, subgraphs, and optimal vertex assignments. Combining the latter with an iterative graph similarity scheme, we develop the iterative similarity optimal assignment graph kernel, give an iterative algorithm for its computation, prove convergence of the algorithm and the uniqueness of the solution, and provide an upper bound on the number of iterations necessary to achieve a desired precision. In a retrospective virtual screening study, our kernel consistently improved performance over chemical descriptors as well as other optimal assignment graph kernels. Chemical data sets often lie on manifolds of lower dimensionality than the embedding chemical descriptor space. Dimensionality reduction methods try to identify these manifolds, effectively providing descriptive models of the data. For spectral methods based on kernel principle component analysis, the projection error is a quantitative measure of how well new samples are described by such models. This can be used for the identification of compounds structurally dissimilar to the training samples, leading to projection error-based novelty detection for virtual screening using only positive samples. We provide proof of principle by using principle component analysis to learn the concept of fatty acids. The peroxisome proliferator-activated receptor (PPAR) is a nuclear transcription factor that regulates lipid and glucose metabolism, playing a crucial role in the development of type 2 diabetes and dyslipidemia. We establish a Gaussian process regression model for PPAR gamma agonists using a combination of chemical descriptors and the iterative similarity optimal assignment kernel via multiple kernel learning. Screening of a vendor library and subsequent testing of 15 selected compounds in a cell-based transactivation assay resulted in 4 active compounds. One compound, a natural product with cyclobutane scaffold, is a full selective PPAR gamma agonist (EC50 = 10 +/- 0.2 muM, inactive on PPAR alpha and PPAR beta/delta at 10 muM). The study delivered a novel PPAR gamma agonist, de-orphanized a natural bioactive product, and, hints at the natural product origins of pharmacophore patterns in synthetic ligands.
This work investigated the applicability of global pairwise sequence alignment to the detection of functional analogues in virtual screening. This variant of sequence comparison was developed for the identification of homologue proteins based on amino acid or nucleotide sequences. Because of the significant differences between biopolymers and small molecules several aspects of this approach for sequence comparison had to be adapted. All proposed concepts were implemented as the ‘Pharmacophore Alignment Search Tool’ (PhAST) and evaluated in retrospective experiments on the COBRA dataset in version 6.1. The aim to identify functional analogues raised the necessity for identification and classification of functional properties in molecular structures. This was realized by fragment-based atom-typing, where one out of nine functional properties was assigned to each non-hydrogen atom in a structure. These properties were pre-assigned to atoms in the fragments. Whenever a fragment matched a substructure in a molecule, the assigned properties were transferred from fragment atoms to structure atoms. Each functional property was represented by exactly one symbol. Unlike amino acid or nucleotide sequences, small drug-like molecules contain branches and cycles. This was a major obstacle in the application of sequence alignment to virtual screening, since this technique can only be applied to linear sequences of symbols. The best linearization technique was shown to be Minimum Volume Embedding. To the best of knowledge, this work represents the first application of dimensionality reduction to graph linearization. Sequence alignment relies on a scoring system that rates symbol equivalences (matches) and differences (mismatches) based on functional properties that correspond to rated symbols. Existing scoring schemes are applicable only to amino acids and nucleotides. In this work, scoring schemes for functional properties in drug-like molecules were developed based on property frequencies and isofunctionality judged from chemical experience, pairwise sequence alignments, pairwise kernel-based assignments and stochastic optimization. The scoring system based on property frequencies and isofunctionality proved to be the most powerful (measured in enrichment capability). All developed scoring systems performed superior compared to simple scoring approaches that rate matches and mismatches uniformly. The frameworks proposed for score calculations can be used to guide modifications to the atom-typing in promising directions. The scoring system was further modified to allow for emphasis on particular symbols in a sequence. It was proven that the application of weights to symbols that correspond to key interaction points important to receptor-ligand-interaction significantly improves screening capabilities of PhAST. It was demonstrated that the systematic application of weights to all sequence positions in retrospective experiments can be used for pharmacophore elucidation. A scoring system based on structural instead of functional similarity was investigated and found to be suitable for similarity searches in shape-constrained datasets. Three methods for similarity assessment based on alignments were evaluated: Sequence identity, alignment score and significance. PhAST achieved significantly higher enrichment with alignment scores compared to sequence identity. p-values as significance estimates were calculated in a combination of Marcov Chain Monte Carlo Simulation and Importance Sampling. p-values were adapted to library size in a Bonferroni correction, yielding E-values. A significance threshold of an E-value of 1*10-5 was proposed for the application in prospective screenings. PhAST was compared to state-of-the-art methods for virtual screening. The unweighted version was shown to exhibit comparable enrichment capabilities. Compound rankings obtained with PhAST were proven to be complementary to those of other methods. The application to three-dimensional instead of two-dimensional molecular representations resulted in altered compound rankings without increased enrichment. PhAST was employed in two prospective applications. A screening for non-nucleoside analogue inhibitors of bacterial thymidin kinase yielded a hit with a distinct structural framework but only weak activity. The search for drugs not member of the NSAID (non-steroidal anti-inflammatory drug) class as modulators of gamma-secretase resulted in a potent modulator with clear structural distiction from the reference compound. The calculation of significance estimates, emphasizing on key interactions, the pharmacophore elucidation capabilities and the unique compound rannkings set PhAST apart from other screening techniques.
Development of a computational method for reaction-driven de novo design of druglike compounds
(2010)
A new method for computer-based de novo design of drug candidate structures is proposed. DOGS (Design of Genuine Structures) features a ligand-based strategy to suggest new molecular structures. The quality of designed compounds is assessed by a graph kernel method measuring the distance of designed molecules to a known reference ligand. Two graph representations of molecules (molecular graph and reduced graph) are implemented to feature different levels of abstraction from the molecular structure. A fully deterministic construction procedure explicitly designed to facilitate synthesizability of proposed structures is realized: DOGS uses readily available synthesis building blocks and established reaction schemes to assemble new molecules. This approach enables the software to propose not only the final compounds, but also to give suggestions for synthesis routes to generate them at the bench. The set of synthesis schemes comprises about 83 chemical reactions. Special focus was put on ring closure reactions forming drug-like substructures. The library of building blocks consists of about 25,000 readily available synthesis building blocks. DOGS builds up new structures in a stepwise process. Each virtual synthesis step adds a fragment to the growing molecule until a stop criterion (upper threshold for molecular mass or number of synthesis steps) is fulfilled. In a theoretical evaluation, a set of ~1,800 molecules proposed by DOGS is analyzed for critical properties of de novo designed compounds. The software is able to suggest drug-like molecules (79% violate less than two of Lipinski’s ‘rule of five’). In addition, a trained classifier for drug-likeness assigns a score >0.8 to 51% of the designed molecules (with 1.0 being the top score). In addition, most of the DOGS molecules are deemed to be synthesizable by a retro-synthesis descriptor (77% of molecules score in the top 10% of the decriptor’s value range). Calculated logP(o/w) values of constructed molecules resemble a unimodal distribution centred close to the mean of logP(o/w) values calculated for the reference compounds. A structural analysis of selected designs reveals that DOGS is capable of constructing molecules reflecting the overall topological arrangement of pharmacophoric features found in the reference ligands. At the same time, the DOGS designs represent innovative compounds being structurally distinct from the references. Synthesis routes for these examples are short and seem feasible in most cases. Some reaction steps might need modification by using protecting groups to avoid unwanted side reactions. Plausible bioisosters for known privileged fragments addressing the S1 pocket of trypsin were proposed by DOGS in a case study. Three of them can be found in known trypsin inhibitors as S1-adressing side chains. The software was also tested in two prospective case studies to design bioactive compounds. DOGS was applied to design ligands for human gamma-secretase and human histamine receptor subtype 4 (hH4R). Two selected designs for gamma-secretase were readily synthesizable as suggested by the software in one-step reactions. Both compounds represent inverse modulators of the target molecule. In a second case study, a ligand candidate selected for hH4R was synthesized exactly following the three-step synthesis plan suggested by DOGS. This compound showed low activity on the target structure. The concept of DOGS is able to deliver synthesizable and bioactive compounds. Suggested synthesis plans of selected compounds were readily pursuable. DOGS can therefore serve as a valuable idea generator for the design of new pharmacological active compounds.
This study focuses on structural features of a particular GPCR type, the family C GPCRs. Structure- and ligand-based approaches were adopted for prediction of novel mGluR5 binding ligand and their binding modes. The objectives of this study were: 1. An analysis of function and structural implication of amino acids in the TM region of family C GPCRs. 2. The prediction of the TM domain structure of mGluR5. 3. The discovery of novel selective allosteric modulators of mGluR5 by virtual screening. 4. The prediction of a ligand binding mode for the allosteric binding site in mGluR5. GPCRs are a super-family of structurally related proteins although their primary amino acid sequence can be diverse. Using sequence information a conservation analysis of family C GPCRs should be applied to reveal characteristic differences and similarities with respect function, folding and ligand binding. Using experimental data and conservation analysis the allosteric binding site of mGluR5 should be characterized regarding NAM and PAM and selective ligand binding. For further evaluation experimental knowledge about family A GPCRs as well as conservation between vertebrate rhodopsins was planned to be compared to results obtained for family C GPCRs (Section 4.1 Conservation analysis of family C GPCRs). Since no receptor structure is available for any family C GPCR, discussion of conserved sequence positions between family A and C GPCRs requires the prediction of a receptor structure for mGluR5 using a family A receptor as template. In order to predict the mGluR5 structure a sequence alignment to a GPCR template protein will have to be proposed and GPCR specific features considered in structure calculation (Section 4.1.4 Structure prediction of mGluR5). The obtained structure was intended to be involved in ligand binding mode prediction of newly discovered active molecules. For discovery of novel selective mGluR modulators several ligand-based virtual screening protocols were adapted and evaluated. Prediction models were derived for selection of possibly active molecules using a diverse collection of known mGluR binding ligands. For that purpose a data collection of known mGluR binding ligands should be established and this reference collection analyzed with respect to different ligand activity classes, NAM or PAM and selective modulators. The prediction of novel NAMs and PAMs using several combinations of 2D-, 3D-, pharmacophore or molecule shape encoding methods with machine learning techniques and similarity determining methods should be tested in a prospective manner (Section 4.2 Virtual screening for novel mGluR modulators). In collaboration with Merz Pharmaceuticals (Merz GmbH & Co. KGaA, Frankfurt am Main, Germany) the modulating effect of a few hundred molecules should be approved in a functional cell-based assay. With the objective to predict a binding mode of the discovered active molecules, molecule docking should be applied using the allosteric binding site of the modeled mGluR5 structure (Section 4.2.4 Modeling of binding modes). Predicted ligand binding modes are to be correlated to conservation profiles that had resulted from the sequence-based entropy analysis and information from mutation experiments, and shall be compared to known ligand binding poses from crystal structures of family A GPCRs.
Reggie-1 (flotillin-2) and reggie-2 (flotillin-1) are membrane microdomain proteins which are associated with the membrane by means of acylation. They influence different cellular signaling processes, such as neuronal, T-cell and insulin signaling. Upon stimulation of the EGF receptor, reggie-1 becomes phosphorylated and undergoes tyrosine 163 dependent translocation from the plasma membrane to endosomal compartments. In addition, reggie-1 was shown to influence actindependent processes. Reggie-2 has been demonstrated to affect caveolin- and clathrin-independent endocytosis. Both proteins form homo- and hetero-oligomers, but the function of these oligomers has remained elusive. Moreover, it has not been clarified if functions of reggie-1 are also influenced by reggie-2 and vice versa. The first aim of the study was to further investigate the interplay and the heterooligomerization of reggie proteins and their functional effects. Both reggie proteins were individually depleted by means of siRNA. In different siRNA systems and various cell lines, reggie-1 depleted cells showed reduced protein amounts of reggie-1 and reggie-2, but reggie-2 knock down cells still expressed reggie-1 protein. The decrease of reggie-2 in reggie-1 depleted cells was only detected at protein but not at mRNA level. Furthermore, reggie-2 expression could be rescued by expression of siRNA resistant wild type reggie-1-EGFP constructs, but not by the soluble myristoylation mutant G2A. This mutant was also not able to associate with endogenous reggie-1 or reggie-2, which demonstrates that membrane association of reggie-1 is necessary for hetero-oligomerization. In addition, fluorescence microscopy studies and membrane fractionations showed that correct localization of overexpressed reggie-2 was dependent on co-overexpressed reggie-1. Thus, hetero-oligomerization is crucial for membrane association of reggie-2 and for its protein stability or protein expression. Moreover, the binding of reggie-2 to reggie-1 required tyrosine 163 of reggie-1 which was previously shown to be important for endosomal translocation of reggie-1. Since reggie-2 was implicated to function in clathrin- and caveolin-independent endocytosis pathways, the effect of reggie-2 depletion on reggie-1 endocytosis was investigated. Indeed, reggie-1 was dependent on reggie-2 for endosomal localization and EGF-induced endocytosis. By FRET-FLIM analysis it could be shown that reggie heterooligomers are dynamic in size or conformation upon EGF stimulation. Thus, it can be concluded that reggie proteins are interdependent in different aspects, such as protein stability or expression, membrane association and subcellular localization. In addition, these results demonstrate that the hetero-oligomers are dynamic and reggie proteins influence each other in terms of function. A further aim was the characterization of reggie-1 and reggie-2 function in actindependent processes, where so far only reggie-1 was known to play a role. Depletion of either of the proteins reduced cell migration, cell spreading and the number of focal adhesions in steady state cells. Thus, also reggie-2 affects actin-dependent processes. Further investigation of the focal adhesions during cell spreading revealed that depletion of reggie-1 displayed different effects as compared to reggie-2 knock down. Reggie-1 depleted cells had elongated cell-matrix-adhesions and showed reduced activation of FAK and ERK2. On the other hand, depletion of reggie-2 resulted in a restricted localization of focal adhesion at the periphery of the cell and decreased ERK2 phosphorylation, but it did not affect FAK autophosphorylation. Hence, reggie proteins influence the regulation of cell-matrix-adhesions differently. A link between reggie proteins and focal adhesions is the actin cross-linking protein -actinin. The interaction of -actinin with reggie-1 could be verified by means of co-immunoprecipitations and FRET-FLIM analysis. Reggie-1 binds -actinin especially in membrane ruffles and in other locations where actin remodeling takes place. Moreover, -actinin showed a different localization pattern during cell spreading in reggie-1 depleted cells, as compared to the control cells. These results provide further insights into the function of both reggie proteins. Their interplay and hetero-oligomerization was shown to be crucial for their role in endocytosis. In addition, both reggie proteins influence actin-dependent processes and differentially affect focal adhesion regulation.
Analysis of coding principles in the olfactory system and their application in cheminformatics
(2007)
Unser Geruchssinn vermittelt uns die Wahrnehmung der chemischen Welt. Im Laufe der Evolution haben sich in unserem olfaktorischen System Mechanismen entwickelt, die wahrscheinlich optimal auf die Erfüllung dieser Aufgabe angepasst sind. Die Analyse dieser Verarbeitungsstrategien verspricht Einblicke in effiziente Algorithmen für die Kodierung und Verarbeitung chemischer Information, deren Entwicklung und Anwendung dem Kern der Chemieinformatik entspricht. In dieser Arbeit nähern wir uns der Entschlüsselung dieser Mechanismen durch die rechnerische Modellierung von funktionellen Einheiten des olfaktorischen Systems. Hierbei verfolgten wir einen interdisziplinären Ansatz, der die Gebiete der Chemie, der Neurobiologie und des maschinellen Lernens mit einbezieht.
The aim of the thesis was to identify structure activity relationships (SAR) in the primary screening data of high-throughput screening (HTS) assays. The strategy was to perform a hierarchical clustering of the molecules, assign the primary screening data to the created clusters and derive models from the clusters. The models should serve to identify singletons, clusters enriched with actives, not confirmed hits and false-negatives. Two hierarchical clustering algorithms, NIPALSTREE and hierarchical k-means have been developed and adapted for this purpose, respectively. A graphical user interface (GUI) has been implemented to extract SAR from the clustering results. Retrospective and prospective applications of the clustering approach were performed. SAR models were created by combining the clustering results with different chemoinformatic methods. NIPALSTREE projects a data set onto one dimension using principle component analysis. The data set is sorted according to the scoring vector and split at the median position into two subsets. The algorithm is applied recursively onto the subsets. The hierarchical k-means recursively separates a data set into two clusters using the k-means algorithm. Both algorithms are capable of clustering large data sets with more than a million data points. They were validated and compared to each other on the basis of different structural classes. NIPALSTREE provided with the loading vectors first insights into SAR whereas the hierarchical k-means yielded superior results. A GUI was developed allowing the display of and the navigation in the clustering results. Functionalities were integrated to analyse the clusters in the dendrogram, molecules in a cluster, and physicochemical properties of a molecule. Measures were developed to identify clusters enriched with actives, to characterize singletons and to analyse selectivity and specificity. Different protease inhibitors of the COBRA database were examined using the hierarchical k-means algorithm. Supported by similarity searches and nearest neighbour analyses thrombin inhibitor singletons were quickly isolated and displayed in the dendrogram. By scaling enrichment factors to the logarithm of the dendrogram level, clusters enriched with different structural classes of factor Xa inhibitors were simultaneously identified. The observed co-clustering of other protease inhibitors provided a deeper insight into selectivity and specificity and shows the utility of the approach for constructing focussed screening libraries. Specificity was analyzed by extracting and clustering relative frequencies of the protease inhibitors from the clusters of dendrogram level 7. A unique ligand based point of view on the pocketome of the protease enzymes was obtained. To identify not confirmed hits and false-negatives in the primary screening data of HTS assays, three assays were retrospectively analysed with the hierarchical k-means algorithm. A rule catalogue was developed judging hits in terminal clusters based on the cluster size, the percent control values of the entries in a cluster, the overall hit rate, the hit rate in the cluster and the environment of a cluster in the dendrogram. It resulted in the identification of a high proportion of not confirmed hits and provided for each hit a rating in context of related non-hits. This allows prioritizing compounds for follow-up studies. Non-hits and hits were retrieved from terminal clusters containing hits. Molecules bearing false-negative scaffolds were co-extracted and enriched. To minimize the number of false-positives in the extracted lists, Bayesian regularized artificial neutral network classification models were trained with the data. Applying the models marked improvement of enrichment factors for the false-negatives was obtained. It proofs the scaffold-hopping potential of the approach. NIPALSTREE, the hierarchical k-means algorithm and self-organising maps were prospectively applied to identify novel lead candidates for dopamine D3 receptors. Compounds with novel scaffolds and low nanomolar binding affinity (65 nM, compound 42) were identified. To provide a deeper insight into the SAR of these molecules, different alternative computational methods were employed. Support vector-based regression and partial least squares were examined. Predictive models for dopamine D2 and D3 receptor binding affinity values were obtained. Important features explaining SAR were extracted from the models. The prospective application of the models to the diverse and novel virtual screening data was of limited success only. Docking studies were performed using a homology model of the dopamine D3 receptor. The visual inspection of the binding modes resulted in the hypothesis of two alternative binding pockets for the aryl moiety of dopamine D3 receptor antagonists. A pharmacophore model was created simultaneously requiring both aryl moieties. Virtual screening with the model identified a nanomolar hit (65 nM, compound 59) corroborating the hypothesis of the two binding pockets and providing a new lead structure for dopamine D3 receptors. The presented data shows that the combined approach of hierarchically clustering a data set in combination with the subsequent usage of the clusters for model generation is suited to extract SAR from screening data. The models are successful in identifying singletons, clusters enriched with actives, not confirmed hits and false-negative scaffolds.
The goal of this thesis was the development, evaluation and application of novel virtual screening approaches for the rational compilation of high quality pharmacological screening libraries. The criteria for a high quality were a high probability of the selected molecules to be active compared to randomly selected molecules and diversity in the retrieved chemotypes of the selected molecules to be prepared for the attrition of single lead structures. For the latter criterion the virtual screening approach had to perform “scaffold hopping”. The first molecular descriptor that was explicitly reported for that purpose was the topological pharmacophore CATS descriptor, representing a correlation vector (CV) of all pharmacophore points in a molecule. The representation is alignment-free and thus renders fast screening of large databases feasible. In a first series of experiments the CATS descriptor was conceptually extended to the three-dimensional pharmacophore-pair CATS3D descriptor and the molecular surface based SURFCATS descriptor. The scaling of the CATS3D descriptor, the combination of CATS3D with different similarity metrics and the dependence of the CATS3D descriptor on the threedimensional conformations of the molecules in the virtual screening database were evaluated in retrospective screening experiments. The “scaffold hopping” capabilities of CATS3D and SURFCATS were compared to CATS and the substructure fingerprint MACCS keys. Prospective virtual screening with CATS3D similarity searching was applied for the TAR RNA and the metabotropic glutamate receptor 5 (mGlur5). A combination of supervised and unsupervised neural networks trained on CATS3D descriptors was applied prospectively to compile a focused but still diverse library of mGluR5 modulators. In a second series of experiments the SQUID fuzzy pharmacophore model method was developed, that was aimed to provide a more general query for virtual screening than the CATS family descriptors. A prospective application of the fuzzy pharmacophore models was performed for TAR RNA ligands. In a last experiment a structure-/ligand-based pharmacophore model was developed for taspase1 based on a homology model of the enzyme. This model was applied prospectively for the screening for the first inhibitors of taspase1. The effect of different similarity metrics (Euc: Euclidean distance, Manh: Manhattan distance and Tani: Tanimoto similarity) and different scaling methods (unscaled, scaling1: scaling by the number of atoms, and scaling2: scaling by the added incidences of potential pharmacophore points of atom pairs) on CATS3D similarity searching was evaluated in retrospective virtual screening experiments. 12 target classes of the COBRA database of annotated ligands from recent scientific literature were used for that purpose. Scaling2, a new development for the CATS3D descriptor, was shown to perform best on average in combination with all three similarity metrics (enrichment factor ef (1%): Manh = 11.8 ± 4.3, Euc = 11.9 ± 4.6, Tani = 12.8 ± 5.1). The Tanimoto coefficient was found to perform best with the new scaling method. Using the other scaling methods the Manhattan distance performed best (ef (1%): unscaled: Manh = 9.6 ± 4.0, Euc = 8.1 ± 3.5, Tani = 8.3 ± 3.8; scaling1: Manh = 10.3 ± 4.1, Euc = 8.8 ± 3.6, Tani = 9.1 ± 3.8). Since CATS3D is independent of an alignment, the dependence of a “receptor relevant” conformation might also be weaker compared to other methods like docking. Using such methods might be a possibility to overcome problems like protein flexibility or the computational expensive calculation of many conformers. To test this hypothesis, co-crystal structures of 11 target classes served as queries for virtual screening of the COBRA database. Different numbers of conformations were calculated for the COBRA database. Using only a single conformation already resulted in a significant enrichment of isofunctional molecules on average (ef (1%) = 6.0 ± 6.5). This observation was also made for ligand classes with many rotatable bonds (e.g. HIV-protease: 19.3 ± 6.2 rotatable bonds in COBRA, ef (1%) = 12.2 ± 11.8). On average only an improvement from using the maximum number of conformations (on average 37 conformations / molecule) to using single conformations of 1.1 fold was found. It was found that using more conformations actives and inactives equally became more similar to the reference compounds according to the CATS3D representations. Applying the same parameters as before to calculate conformations for the crystal structure ligands resulted in an average Cartesian RMSD of the single conformations to the crystal structure conformations of 1.7 ± 0.7 Å. For the maximum number of conformations, the RMSD decreased to 1.0 ± 0.5 Å (1.8 fold improvement on average). To assess the virtual screening performance and the scaffold hopping potential of CATS3D and SURFACATS, these descriptors were compared to CATS and the MACCS keys, a fingerprint based on exact chemical substructures. Retrospective screening of ten classes of the COBRA database was performed. According to the average enrichment factors the MACCS keys performed best (ef (1%): MACCS = 17.4 ± 6.4, CATS = 14.6 ± 5.4, CATS3D = 13.9 ± 4.9, SURFCATS = 12.2 ± 5.5). The classes, where MACCS performed best, consisted of a lower average fraction of different scaffolds relative to the number of molecules (0.44 ± 0.13), than the classes, where CATS performed best (0.65 ± 0.13). CATS3D was the best performing method for only a single target class with an intermediate fraction of scaffolds (0.55). SURFCATS was not found to perform best for a single class. These results indicate that CATS and the CATS3D descriptors might be better suited to find novel scaffolds than the MACCS keys. All methods were also shown to complement each other by retrieving scaffolds that were not found by the other methods. A prospective evaluation of CATS3D similarity searching was done for metabotropic glutamate receptor 5 (mGluR5) allosteric modulators. Seven known antagonists of mGluR5 with sub-micromolar IC50 were used as reference ligands for virtual screening of the 20,000 most drug-like compounds – as predicted by an artificial neural network approach – of the Asinex vendor database (194,563 compounds). Eight of 29 virtual screening hits were found with a Ki below 50 µM in a binding assay. Most of the ligands were only moderately specific for mGluR5 (maximum of > 4.2 fold selectivity) relative to mGluR1, the most similar receptor to mGluR5. One ligand exhibited even a better Ki for mGluR1 than for mGluR5 (mGluR5: Ki > 100 µM, mGluR1: Ki = 14 µM). All hits had different scaffolds than the reference molecules. It was demonstrated that the compiled library contained molecules that were different from the reference structures – as estimated by MACCS substructure fingerprints – but were still considered isofunctional by both CATS and CATS3D pharmacophore approaches. Artificial neural networks (ANN) provide an alternative to similarity searching in virtual screening, with the advantage that they incorporate knowledge from a learning procedure. A combination of artificial neural networks for the compilation of a focused but still structurally diverse screening library was employed prospectively for mGluR5. Ensembles of neural networks were trained on CATS3D representations of the training data for the prediction of “mGluR5-likeness” and for “mGluR5/mGluR1 selectivity”, the most similar receptor to mGluR5, yielding Matthews cc between 0.88 and 0.92 as well as 0.88 and 0.91 respectively. The best 8,403 hits (the focused library: the intersection of the best hits from both prediction tasks) from virtually ranking the Enamine vendor database (ca. 1,000,000 molecules), were further analyzed by two self-organizing maps (SOMs), trained on CATS3D descriptors and on MACCS substructure fingerprints. A diverse and representative subset of the hits was obtained by selecting the most similar molecules to each SOM neuron. Binding studies of the selected compounds (16 molecules from each map) gave that three of the molecules from the CATS3D SOM and two of the molecules from the MACCS SOM showed mGluR5 binding. The best hit with a Ki of 21 µM was found in the CATS3D SOM. The selectivity of the compounds for mGluR5 over mGluR1 was low. Since the binding pockets in the two receptors are similar the general CATS3D representation might not have been appropriate for the prediction of selectivity. In both SOMs new active molecules were found in neurons that did not contain molecules from the training set, i. e. the approach was able to enter new areas of chemical space with respect to mGluR5. The combination of supervised and unsupervised neural networks and CATS3D seemed to be suited for the retrieval of dissimilar molecules with the same class of biological activity, rather than for the optimization of molecules with respect to activity or selectivity. A new virtual screening approach was developed with the SQUID (Sophisticated Quantification of Interaction Distributions) fuzzy pharmacophore method. In SQUID pairs of Gaussian probability densities are used for the construction of a CV descriptor. The Gaussians represent clusters of atoms comprising the same pharmacophoric feature within an alignment of several active reference molecules. The fuzzy representation of the molecules should enhance the performance in scaffold hopping. Pharmacophore models with different degrees of fuzziness (resolution) can be defined which might be an appropriate means to compensate for ligand and receptor flexibility. For virtual screening the 3D distribution of Gaussian densities is transformed into a two-point correlation vector representation which describes the probability density for the presence of atom-pairs, comprising defined pharmacophoric features. The fuzzy pharmacophore CV was used to rank CATS3D representations of molecules. The approach was validated by retrospective screening for cyclooxygenase 2 (COX-2) and thrombin ligands. A variety of models with different degrees of fuzziness were calculated and tested for both classes of molecules. Best performance was obtained with pharmacophore models reflecting an intermediate degree of fuzziness. Appropriately weighted fuzzy pharmacophore models performed better in retrospective screening than CATS3D similarity searching using single query molecules, for both COX-2 and thrombin (ef (1%): COX-2: SQUID = 39.2., best CATS3D result = 26.6; Thrombin: SQUID = 18.0, best CATS3D result = 16.7). The new pharmacophore method was shown to complement MOE pharmacophore models. SQUID fuzzy pharmacophore and CATS3D virtual screening were applied prospectively to retrieve novel scaffolds of RNA binding molecules, inhibiting the Tat-TAR interaction. A pharmacophore model was built up from one ligand (acetylpromazine, IC50 = 500 µM) and a fragment of another known ligand (CGP40336A), which was assumed to bind with a comparable binding mode as acetylpromazine. The fragment was flexible aligned to the TAR bound NMR conformation of acetylpromazine. Using an optimized SQUID pharmacophore model the 20,000 most druglike molecules from the SPECS database (229,658 compounds) were screened for Tat-TAR ligands. Both reference inhibitors were also applied for CATS3D similarity searching. A set of 19 molecules from the SQUID and CATS3D results was selected for experimental testing. In a fluorescence resonance energy transfer (FRET) assay the best SQUID hit showed an IC50 value of 46 µM, which represents an approximately tenfold improvement over the reference acetylpromazine. The best hit from CATS3D similarity searching showed an IC50 comparable to acetylpromazine (IC50 = 500 µM). Both hits contained different molecular scaffolds than the reference molecules. Structure-based pharmacophores provide an alternative to ligand-based approaches, with the advantage that no ligands have to be known in advance and no topological bias is introduced. The latter is e.g. favorable for hopping from peptide-like substrates to drug-like molecules. A homology model of the threonine aspartase taspase1 was calculated based on the crystal structures of a homologous isoaspartyl peptidase. Docking studies of the substrate with GOLD identified a binding mode where the cleaved bond was situated directly above the reactive N-terminal threonine. The predicted enzyme-substrate complex was used to derive a pharmacophore model for virtual screening for novel taspase1 inhibitors. 85 molecules were identified from virtual screening with the pharmacophore model as potential taspase1- inhibitors, however biochemical data was not available before the end of this thesis. In summary this thesis demonstrated the successful development, improvement and application of pharmacophore-based virtual screening methods for the compilation of molecule-libraries for early phase drug development. The highest potential of such methods seemed to be in scaffold hopping, the non-trivial task of finding different molecules with the same biological activity.