Proteins participate in the majority of cellular processes. To determine the function of a protein it is not sufficient to solely know its sequence, its structure in isolation, or how it works individually. Additionally, we need to know how the protein interacts with other proteins in biological networks. This is because most of the proteins perform their main function through interactions. This thesis sets out to improve the understanding of protein-protein interaction networks (PPINs). For this, we propose three approaches: (1) Studying measures and methods used in social and complex networks. The methods, measures, and properties of social networks allow us to gain an understanding of PPINs via the comparison of different types of network families. We studied models that describe social networks to see which models are useful in describing biological networks. We investigate the similarities and differences in terms of the network community profile and centrality measures. (2) Studying PPINs and their role in evolution. We are interested in the relationship of PPINs and the evolutionary changes between species. We investigate whether the centrality measures are correlated with the variability and similarity in orthologous proteins. (3) Studying protein features that are important to evaluate, classify, and predict interactions. Interactions can be classified according to different characteristics. One characteristic is the energy (that is the attraction or repulsion of the molecules) that occurs in interacting proteins. We identify which type of energy values contributes better to predicting PPIs. We argue that the number of energetic features and their contribution to the interactions can be a key factor in predicting transient and permanent interactions. Contributions of this thesis include: (1) We identified the best community sizes in PPINs. This finding will help to identify important groups of interacting proteins in order to better understand their particular interactions. We furthermore find that the generative model describing biological networks is very different from the model describing social networks A generative model is a model for randomly generating observable data. We showed that the best community size for PPINs is around ten, different from the best community size for social and complex network (around 100). We revealed differences in terms of the network community profile and correlations of centrality measures; (2) We outline a method to test correlation of centrality measures with the percentage of sequence similarity and evolutionary rate for orthologous proteins. We conjecture that a strong correlation exists. While not obtaining positive results for our data. Therefore, (3) we investigate a method to discriminate energetic features of protein interactions that in turn will improve the PPIN data. The use of multiple data sets makes possible to identify the energy values that are useful to classify interactions. For each data set, we performed Random Forest and Support Vector Machine with linear, polynomial, radial, and sigmoid kernels. The accuracy obtained in this analysis reinforces the idea that energetic features in the protein interface help to discriminate between transient and permanent interactions. / Graduate / 0984

Proteins and their function play one of the most essential roles in various biological processes. The study of PPI is of considerable importance. PPI network data are of great scientific value, however, they are incomplete and experimental identification is time and money consuming. Available computational methods perform well on model organisms’ PPI prediction but perform poorly for a novel organism. Due to the incompleteness of interaction data, it is challenging to train a model for a novel organism. Also, millions to billions of interactions need to be verified which is extremely compute-intensive. We aim to improve the performance of predicting whether a pair of proteins will interact, with only two sequences as input. And also efficiently predict a PPI network with a proteome of sequences as input. We hypothesize that information about cellular locations where proteins are active and proteins' 3D structures can help us to significantly improve predict performance. To overcome the lack of experimental data, we use predicted structures by AlphaFold2 and cellular locations by DeepGoPlus. We believe that proteins belonging to disjoint biological components have very little chance to interact. We manually choose several disjoint pairs and further confirmed it by experimental PPI. We generate new no-interaction pairs with disjoint classes to update the D-SCRIPT dataset. As result, the AUPR has improved by 10% compared to the D-SCRIPT dataset. Besides, we pre-filter the negatives instead of enumerating all the potential PPI for de-novo PPI network prediction. For E.coli, we can pass around a million negative interactions. To combine the structure and sequence information, we generate a graph for each protein. A graph convolution network using Self-Attention Graph Pooling in Siamese architecture is used to learn these graphs for PPI prediction. In this way, we can improve around 20% in AUPR compared to our baseline model D-SCRIPT.

In the past decade, automatisation has led to an immense increase of data in biology. Next generation sequencing techniques will produce a vast amount of sequences across all species in the coming years. In many cases, identifying the function and biological role of a protein from its sequence can be a complicated and time-intensive task. The identification of a protein's interaction partners is a tremendous help for understanding the biological context in which it is involved. In order to fully characterise a protein-protein interaction (PPIs), it is necessary to know the three-dimensional structure of the interacting partners. Despite optimisation efforts from projects such as the Protein Structure Initivative, determining the structure of a protein through crystallography remains a time- and cost-intensive procedure. The primary aim of the research described in this dissertation was to produce a World Wide Web resource that facilitates visual exploration and validation (or questioning) of data derived from functional genomics experiments, by building upon existing structural information about direct physical PPIs. Secondary aims were (i) to demonstrate the utility of the new resource, and (ii) its application in biological research. We created a database that emphasises specifically the intersection between the PPIs-results emerging from the structural biology and functional genomics communities. The BISC database holds BInary SubComplexes and Modellable Interactions in current functional genomics databases (BICS-MI). It is publicly available at hyyp://bisc.cse.ucsc.edu. BISC is divided in three sections that deliver three types of information of interest to users seeking to investigate or browse PPIs. The template section (BISCHom and BISCHet) is devoted to those PPIs that are characterised in structural detail, i.e. binary SCs extracted from experimentally determined three-dimensional structures. BISCHom and BISCHet contain the homodimeric (13,583 records) and heterodimeric (5612 records) portions of these, respectively. Besides interactive, embedded Jmol displays emphasising the interface, standard information and links are provided, e.g. sequence information and SPOP classification for both partners, interface size and energy scores (PISA). An automated launch of the MolSurfer program enables the user to investigate electrostatic and hydrophobic correlation between the partners, at the inter-molecular interface. The modellable interactions section (BISC0MI) identifies potentially modellable interactions in three major functional genomics interaction databases (BioGRID), IntAct, HPRD). To create BISC-MI all PPIs that are amenable to automated homology modelling based on conservative similarity cut-offs and whose partner protein sequences have recrods in the UniProt database, have been extracted. The modellable interaction services (BISC-MI Services) section offers, upon user request, modelled SC-structures for any PPIs in BISC-MI. This is enabled through an untomated template-based (homology) modelling protocol using the popular MODELLER program. First, a multiple sequence alignment (MSA) is generated using MUSCLE, between the target and homologous proteins collected from UniProt (only reviewed proteins from organisms whose genome has been completely sequenced are included to find putative orthologs). Then a sequence-to-profile alignment is generated to integrate the template structure in the MSA. All models are produced upon user request to ensure that the most recent sequence data for the MSAs are used. Models generated through this protocol are expected to be more accurate generally than models offered by other automated resources that rely on pairwise alignments, e.g. ModBase. Two small studies were carried out to demonstrate the usability and utility of BISC in biological research. (1) Interaction data in functional genomics databases often suffers from insufficient experimental and reporting standards. For example, multiple protein complexes are typically recorded as an inferred set of binary interactions. Using the 20S core particle of the yeast proteasome as an example, we demonstrate how the BISC Web resource can be used as a starting point for further investigation of such inferred interactions. (2) Malaria, a mosquito-borne disease, affects 3500-500 million people worldwide. Still very little is known about the malarial parasites' genes and their protein functions. For Plasmodium falciparum, the most lethal among the malaria parasites, only one experimentally derived medium scale PPIs set is available. The validity of this set has been doubted in the the malarial biologist community. We modelled and investigated eleven binary interactions from this set using the BISC modelling pipeline. Alongside we compared the BISC models of the individual partners to those obtained from ModBase.

No description available.

Protein-protein interactions (PPIs) are key drug targets and recent breakthroughs in this area are providing insight into the types of molecules needed to selectively and potently inhibit a target traditionally seen as untractable. The rules that have been used to design classic substratecompetitive drugs (for example Lipinski's rule of five) may not apply in this new field in the same way. Here I present work performed in three systems that are well-validated drug targets for oncogenesis: the CDK2/cyclin A complex, the PLK1 Polobox domain and MDM2. In each case the site of the protein-protein interaction is defined and understood and the rationale for pharmaceutical intervention is clear. I use these as a model system to evaluate the characteristics of drugs that target protein-protein interaction sites and present work on the development of inhibitors as potential leads for subsequent drug development. In Chapter 1 I introduce the problems, challenges and rewards of PPI drug development; in Chapter 2 I present co-crystal structures of MDM2 with isoindolinone inhibitors; in Chapter 3 I detail attempts to co-crystallise the Plk1 Polobox with inhibitors and screen potential inhibitors; in Chapter 4 I present the results of screening to identify inhibitors of Cyclin A recruitment; and in Chapter 5 I discuss other strategies for inhibition of the CDK2/cyclin A complex, including results with a covalent inhibitor. Through these projects I have been able to demonstrate the wide applicability of the PPI inhibition approach, identify key features of drugs able to inhibit PPIs and contribute to drug design in each system.

L'objectif de cette thèse est de développer des outils de prédiction d'interactions entre protéines qui puissent être appliqués en particulier sur le réseau d’interaction autour de la protéine CFTR, qui est impliquée dans la mucoviscidose. Le développement de méthodes de prédiction in silico peut s'avérer utile pour suggérer aux biologistes de nouvelles cibles d'interaction. Nous proposons une nouvelle méthode pour la prédiction de liens dans un réseau. Afin de bénéficier de l'information des données non étiquetées, nous nous plaçons dans le cadre de l'apprentissage semi-supervisé. Nous abordons ce problème de prédiction comme une tâche d'apprentissage d'un noyau de sortie. Un noyau de sortie est supposé coder les proximités existantes entres les nœuds du graphe et l'objectif est d'approcher ce noyau à partir de descriptions appropriées en entrée. L'utilisation de l'astuce du noyau dans l'ensemble de sortie permet de réduire le problème d'apprentissage à celui d'une fonction d'une seule variable à valeurs dans un espace de Hilbert. En choisissant les fonctions candidates pour la régression dans un espace de Hilbert à noyau reproduisant à valeur opérateur, nous développons, comme dans le cas de fonctions à valeurs scalaires, des outils de régularisation. Nous établissons en particulier des théorèmes de représentation, qui permettent de définir de nouveaux modèles de régression. Nous avons testé l'approche développée sur des données artificielles, des problèmes test ainsi que sur un réseau d'interaction chez la levure et obtenu de très bons résultats. Puis nous l'avons appliquée à la prédiction d'interactions entre protéines dans le cas d'un réseau construit autour de CFTR. / The aim of this thesis is to develop tools for predicting interactions between proteins that can be applied to the human proteins forming a network with the CFTR protein. This protein, when defective, is involved in cystic fibrosis. The development of in silico prediction methods can be useful for biologists to suggest new interaction targets. We propose a new method to solve the link prediction problem. To benefit from the information of unlabeled data, we place ourselves in the semi-supervised learning framework. Link prediction is addressed as an output kernel learning task, referred as Output Kernel Regression. An output kernel is assumed to encode the proximities of nodes in the target graph and the goal is to approximate this kernel by using appropriate input features. Using the kernel trick in the output space allows one to reduce the problem of learning from pairs to learning a single variable function with output values in a Hilbert space. By choosing candidates for regression functions in a reproducing kernel Hilbert space with operator valued kernels, we develop tools for regularization as for scalar-valued functions. We establish representer theorems in the supervised and semi-supervised cases and use them to define new regression models for different cost functions. We first tested the developed approach on transductive link prediction using artificial data, benchmark data as well as a protein-protein interaction network of the yeast and we obtained very good results. Then we applied it to the prediction of protein interactions in a network built around the CFTR protein.

Protein aggregation is a frequently cited problem during the development of liquid protein formulations, which is especially problematic since each protein exhibits different aggregation behaviour. Aggregation can be controlled by judicious choice of solution conditions, such as salt and buffer type and concentration, pH, and small molecule additives. However, finding conditions is still a trial and error process. In order to improve formulation development, a fundamental understanding of how excipients impact upon protein aggregation would significantly contribute to the development of stable protein therapeutics. The underlying mechanisms that control effects of excipients on protein behaviour are poorly understood. This dissertation is directed at understanding how excipients alter the conformational and colloidal stability of proteins and the link to aggregation. This knowledge can be used for finding novel ways of either predicting or preventing/inhibiting protein aggregation. Experiments using static and dynamic light scattering, intrinsic fluorescence, turbidity and electrophoretic light scattering were conducted to study the effect of solution conditions such as pH, salt type and concentration on protein aggregation behaviour for three model systems: lysozyme, insulin and a monoclonal antibody. Emphasis is placed on understanding the effects of solution additives on protein-protein interactions and the link to aggregation. This understanding has allowed the rational development of stable formulations with novel additives, such as arginine containing dipeptides and polycations.

Protein-protein interactions are essential for the survival of all living cells, allowing for processes such as cell signalling, metabolism and cell division to occur. Yet in humans there are only >38k annotated interactions of an interactome estimated to range between 150k to 600k interactions and out of a potential 300M protein pairs.Experimental methods to define the human interactome generate high quality results, but are expensive and slow. Computational methods play an important role to fill the gap.To further this goal, the prediction of human protein-protein interactions was investigated by the development of new predictive modules and the analysis of diverse datasets within the framework of the previously established PIPs protein-protein interaction predictor Scott and Barton 2007. New features considered include the semantic similarity of Gene Ontology annotating terms, clustering of interaction networks, primary sequences and gene co-expression. Integrating the new features in a naive Bayesian manner as part of the PIPs 2 predictor resulted in two sets of predictions. With a conservative threshold, the union of both sets is >300k predicted human interactions with an intersect of >94k interactions, of which a subset have been experimentally validated. The PIPs 2 predictor is also capable of making predictions in organisms that have no annotated interactions. This is achieved by training the PIPs 2 predictor based on a set of evidence and annotated interactions in another organism resulting in a ranking of protein pairs in the original organism of interest. Such an approach allows for predictions to be made across the whole proteome of poorly characterised organism, rather than being limited only to proteins with known orthologues. The work described here has increased the coverage of the human interactome and introduced a method to predict interactions in organisms that have previously had limited or no annotated interactions. The thesis aims to provide a stepping stone towards the completion of the human interactome and a way of predicting interactions in organisms that have been less well studied, but are often clinically relevant.

Understanding how proteins fold and bind is interesting since these processes are central to most biological activity. Protein folding and protein-protein interaction are by themselves very complex but using a good and robust system to study them could ease some of the hurdles. In this thesis I have tried to answer some of the fundamental questions of protein folding and binding. I chose to work with PDZ domains, which are protein domains consisting of 90-100 amino acids. They are found in more than 400 human proteins and function mostly as protein-protein interaction units. These proteins are very stable, easy to express and purify and their folding reaction is reversible under most laboratory conditions. I have characterized the interaction of PSD-95 PDZ3 domain with its putative ligand under different experimental conditions and found out that its binding kinetics is sensitive to salt and pH.  I also demonstrated that the two conserved residues R318 and H372 in PDZ3 are responsible for the salt and pH effect, respectively, on the binding reaction. Moreover, I determined that for PSD 95 PDZ3 coupling of distal residues to peptide binding was better described by a distance relationship and there was a very weak evidence of an allosteric network. Further, I showed that another PDZ domain, SAP97 PDZ2 undergoes conformational change upon ligand binding. Also, I characterized the binding mechanism of a dimeirc ligand/PDZ1-2 tandem interaction and showed that despite its apparent complexity the binding reaction is best described by a square scheme. Additionally, I determined that for the SAP 97 PDZ/HPV E6 interaction that all three PDZ domains each bind one molecule of the E6 protein and that a set of residues in the PDZ2 of SAP 97 could operate in an unexpected long-range manner during E6 interaction. Finally, I showed that perhaps all members in the PDZ family could fold via a three state folding mechanism. I characterized the folding mechanism of five different PDZ domains having similar overall fold but different primary structure and the results indicate that all five fold via an intermediate with two transition states. Transition state one is rate limiting at low denaturant concentration and vice versa for transition state two. Comparing and characterizing the structures of the transition states of two PDZ domains using phi value analysis indicated that their early transition states are less similar as compared to their late transition states.

The Membrane Yeast Two-Hybrid (MYTH) technology was used in this study to find protein-protein interactors of Saccharomyces cerevisiae ATP binding cassette (ABC) transporters Nft1p, Pdr10p, Pdr18p and Vmr1p. There were 23 interactors for Nft1p, 22 interactors for Pdr10p, 4 interactors for Pdr18p and 1 interactor for Vmr1p. The 43 unique interactors belong to a wide variety of functional categories. There were 11 interactors involved in metabolism, 9 interactors involved in transport, 8 interactors with unknown function, 4 interactors involved in trafficking and secretion, 3 interactors involved in protein folding, 2 interactors involved in stress response, and 1 interactor in each of the following categories: cell wall assembly, cytoskeleton maintenance, nuclear function, protein degradation, protein modification and protein synthesis. Follow up experiments also showed that Pdr15p and Pdr18p play an important role in zinc homeostasis because deletion of these ABC transporters results in sensitivity to zinc shock.