Large-scale gene expression studies are widely used to identify genes that are differentially expressed between phenotypes relevant to disease. Often thousands of differentially expressed genes (DEGs) are found using this type of analysis, which complicates the interpretation of the data. In this project we treat DEGs as windows into the biological processes that underlie disease. In order to find these processes, we put DEGs into the context in which they perform their functions - through the interactions of their protein products. Protein-protein interactions can provide biological context to DEGs in the form of functional modules. These modules are groups of proteins that together perform cellular functions. In this thesis we have refined a functional module detection process that consists of two steps. Firstly, community detection methods are applied to protein interaction networks (PINs) to detect groups of interacting proteins, and secondly, the biological coherence of the proteins grouped together is evaluated to select communities that represent potential functional modules. Two features that are central to this work are the detection of modules at different scales of network organization, and CommWalker, a module evaluation method that we developed which is able to detect signals of poorly-studied functions. By integrating these methods into our functional module detection process, we were able to obtain a good coverage of potential functional modules. Testing for enrichment of DEGs on these functional modules can uncover biological processes that are involved in the contrasted phenotypes and merit further investigation. We have applied our pipeline to find differentially regulated functions between hypoxic and normoxic breast cancer cell lines, and between M1 and M2 macrophages. Our results generate biological hypotheses of cellular functions that are differentially regulated in the investigated phenotypes, and proteins that are involved in these functions. We were able to validate several proteins in enriched modules which did not correspond to DEGs that were input into the pipeline, which suggests our methodology can reveal new biological insight.

Molecular networks are commonly used to explore cellular organisation and disease mechanisms. Function is studied using molecular interaction networks, such as protein-protein networks. Although much biological insight has been gained using these models of molecular function, they are hindered by their reliance on available experimental data and an inability to capture the complexity of biological processes. Functional modules can be identified based on molecular network topology, making it essential that the edges accurately depict molecular interactions. However, these networks struggle to depict the temporal nature of interactions, giving the impression that all interactions are constant. This misrepresentation can result in functionally heterogeneous clusters. The notoriously inaccurate nature of experimental protein interaction data, along with variable conformity among network clusters and functional modules further impedes functional module extraction. Representation of genes by single nodes artificially merges the functions of pleiotropic genes, distorting the arrangement of function within molecular networks. This thesis therefore explores a more suitable model for representing function. Pathways are composed of sets of proteins that are known to interact within a particular cellular context, corresponding to a discernible biological function. Their representation of context dependent cellular activity makes them ideal for use as nodes within a new pathway level model. Using combinatorial algorithms a reduced redundancy pathway set was produced to represent global cellular systems. Enrichment analysis provides reliable functional annotations for each pathway node, attributing independent functions to pleiotropic genes. Edges are based on functional semantic similarity, generating a network representation of functional organisation. Both yeast and human biological systems are presented as functionally connected pathway networks. Pathway annotation and experimentation with semantic similarity measures provides insight into the cross-talk between biological processes. Pathway functional modules elucidate the intracellular implementation of processes. Disease modules highlight the effects of functional perturbations and disease mechanisms. The pathway model provides a complementary, high-level functional model that begins to bridge the gap between molecular data and phenotype. The utilisation of pathway data provides a large, well-validated data source, avoiding the inaccuracies inherent with molecular data. Pathway models better represent components of biological complexity such as pleiotropy and linear implementation of functions.

A pharmacometric analysis must be based on correct data to be valid. Source clinical data is rarely ready to be modelled as is, but rather needs to be reprogrammed to fit the format required by the pharmacometric modelling software. The reprogramming steps include selecting the subsets of data relevant for modelling, deriving new information from the source and adjusting units and encoding. Sometimes, the source data may also be flawed, containing vague definitions and missing or confusing values. In either setting, the source data needs to be reprogrammed to remedy this, followed by extensive quality control to capture any errors or inconsistencies produced along the way. The quality control is a lengthy task which is often performed manually, either by the scientists conducting the pharmacometric study or by independent reviewers. This project presents an automatic data quality control with the purpose of aiding the data curation process, as to minimize any potential errors that would otherwise have to be detected by the manual quality control. The automatic quality control is implemented as an R-package and is specifically tailored for the needs of Pharmetheus.

<p> Every organism contains a few hundred to thousands of proteins. A protein is made of a sequence of molecular building blocks named amino acids. Amino acids will be referred to as residues. Every protein performs one or more functions in the cell. In order for a protein to do its job, it requires to bind properly to other partner proteins. Many genetic diseases such as cancer are caused by mutations (changes) of specific residues which cause disturbances in the functions of those proteins. The problem of prediction of protein binding site is a crucial topic in computational biology. A protein is usually made up of 50 to a few thousand residues. A contact site can occur within a protein or with other proteins. By having a robust and accurate model for identifying residues that are involved in the binding site, scientists can investigate the impact of critical mutations and residues that can cause genetic diseases. </p><p> The main focus of this thesis is to propose a machine learning model for predicting the binding site between two proteins. By extracting structural information from a protein, we can have additional knowledge of binding sites. This structural information can be converted into a penalty matrix for a graphical model to be learned from the protein sequence. The second part of this thesis is mostly focused on motion planning algorithms for proteins and simulation of the protein pathway changes using a Monte Carlo based method. Later, by applying a novel geometry based scoring function, we cluster the intermediate conformations into corresponding subsets that may indicate interesting intermediate states.</p><p>

I propose the use of the tandem fluorescent timer protein as a reporter of dynamic gene regulation. The tandem fluorescent timer is a fusion of two fluorophores with different maturation kinetics whose fluorescence ratio is a reporter of protein age. Traditional approaches to live single-cell tracking of dynamic gene expression involve the use of destabilized fluorescent reporters. The reduced stability of these reporters improve performance but also result in reduced signal and an increased signal to noise ratio. I first develop a platform to test reporter performance by designing and implementing an inducible synthetic network orthogonally in S. cerevisiae cells and by developing a microfluidics-enabled live cell-tracking pipeline. To test the performance of different reporters, I develop an algorithm to decode the underlying regulatory dynamic signal of a fluorescence profile. I then simulate the fluorescence output of my platform under dynamic regulatory signaling to demonstrate the potential reporter performance of a stable timer protein. Finally, I conduct live cell-tracking experiments of yeast cells expressing the timer under a periodic signal to test in vivo performance of the tandem fluorescent timer. I demonstrate that compared to a traditional stable fluorescent reporter, the tandem fluorescent timer is more accurate when tracking faster periodic signals and it is more robust to global fluctuations.

Bioinformatics is a unique field in that it incorporates many different disciplines, including biology, computer science, and statistics, to study biological data. There is a vast array of techniques that utilize bioinformatics, including pangenomics, RNASeq, whole genome metagenomics, and 16S sequencing. To study bacterial interactions, we used a model system of species interactions, Myxococcus xanthus. M. xanthus is a soil bacterium that is a known predator of other bacteria. It has one of the largest repertoires of two component systems (TCS) to respond to external stresses. TCS are a pair of proteins, one that senses environmental stress (histidine kinase, HK) and another that usually acts as a transcriptional regulator (response regulators, RR). We studied a class of RRs, NtrC-like, reliant on an alternative sigma factor, sigma54. The oligomerization of NtrC-like RRs is regulated to modulate activation of the protein, which would change the bacterium’s ability to respond to its environment. We studied HsfA, a NtrC-like RR that regulates specialized metabolites. Specialized metabolites are used in bacterial interactions. In predation interactions they are used to kill prey. Our goal was to find genes that might be involved in specialized metabolite production that would aid in predation. We used prediction tools to find putative binding sites of HsfA to find potentially new metabolites. We used two motifs to attempt to predict if the oligomerization of these response regulators is positively or negatively regulated. We found that the presence of a motif in the receiver domain to be associated with negative regulation of oligomerization, but further studies are needed to experimentally confirm this finding. One environment in which bacterial interactions occur is in the gut. The gut microbiome is the consortium of organisms and their genomic content in the gastrointestinal tract. The gut microbiome is sensitive to aspects of a person’s lifestyle, such as diet and medication. Here we studied the effect of two different diets and two drugs on the gut microbiome. Risperidone, an antipsychotic used to treat schizophrenia and bipolar disorder, has been shown to cause obesity and diabetes. We studied the effect of diet and risperidone usage on weight gain and the microbiome using a C57Bl/6J female mouse model. Our results show that diet has a strong impact on the microbial composition of the gut in response to risperidone. As many mental health patients stop and restart their medication, we examined the effect of stopping and restarting risperidone on the microbiome. When risperidone is stopped the microbiome reverts to a state similar to the control group but diverges into a different microbial composition upon restarting treatment. Interestingly, mice did not gain significantly more weight than their control group upon the second risperidone treatment. Further studies are needed to examine the functional changes occurring with the stop and restart of risperidone to determine the mechanism of mice resisting weight gain during the second round of treatment. Captopril is used to treat hypertension, a very common disease in the United States. Here we studied the effect of captopril on weight gain, metabolic phenotypes, and the gut microbiome. Our results showed that captopril caused an increase in resting metabolic rate (RMR) in mice. This occurred through an increase in energy expenditure. This increase in RMR had the effect of captopril-treated mice being resistant to weight gain. Our group has previously shown that the gut microbiome can directly affect RMR. Therefore, we studied the gut microbiome of captopril-treated mice. We observed a shift in their gut microbiome to organisms Akkermansia muciniphila and Lactobacillus, associated with lean body mass. Captopril therefore has the potential to be a better medication to treat patients with both hypertension and obesity. Further studies are needed to determine the effect of captopril on the microbiome in a hypertension mouse model.

Modern Electronic Health Record (EHR) systems are now integral to healthcare. Having evolved from hospital billing and laboratory systems in the 80's, EHR systems have grown considerably as we learn to represent more and more aspects of patient encounter, diagnosis and treatment digitally. EHR user interfaces, however, lag considerably behind their consumer-electronics counterparts in usability, most notably with respect to customizability. This limitation is especially evident in the implementation of audible alerts that are coupled to sensors or timing devices in intensive-care settings. The most current standard, (ISO/IEC 60601-1-8) has been designed for alerts that are intended to signal situations of varying priorities: however, it is not universally implemented, and has been criticized for the difficulty that healthcare providers have in discriminating between individual alarms, and for the failure to incorporate prior research with respect to "sense of urgency" as it applies to alarm efficacy. In the present work, however, we consider that there are more effective means to allow a user to identify an alarm correctly than "sense of urgency" response. This thesis focuses on the problem of correct identification of alerts: what happens when a human subject is allowed to create or designate (i.e., personalize) one's own alerts? Given the ubiquity, low costs and commoditization of consumer-electronics devices, we believe that it is just a matter of time before such devices become the norm in critical care and replace existing, special-purpose devices for information delivery at the point of patient care. We built a tool, PASA (Personalized Alert Study Application), that would allow us to capture and edit sounds and orchestrate studies that would contrast any two types of sounds. PASA facilitated a study where study participant's responses to "personalized" sounds were contrasted with sounds that meet the ISO/IEC 60601-1-8:2012 standard. We performed two sub-studies that contrasted responses to two banks of 6-alerts and 10-alerts. The 6-alert study was repeated with the same subjects after two weeks without training to measure recall. We observed that accuracy, reaction time, and retention were significantly improved with the personalized sounds. For example, the median errors for the 6-alert baseline study were 4 for personalized vs. 27 for standard alerts. For the 6-alert repeat study it was 7 vs. 43. The median for the 10-alert study was 1 for personalized vs. 55 for standard alerts. Accuracy for recognition, while remaining constant for personalized alerts, degraded considerably for standardized alerts as the number of alerts increased from 6 to 10. We conclude that personalization of alerts may improve information delivery and reduce cognitive overload on the health care provider. This potential positive effect at the point of patient care merits further studies in a clinical or simulated clinical setting.

In recent years, machine learning has emerged as an important discipline. However, despite the popularity of machine learning techniques, data in the form of discrete structures are not fully exploited. For example, when data appear as graphs, the common choice is the transformation of such structures into feature vectors. This procedure, though convenient, does not always effectively capture topological relationships inherent to the data; therefore, the power of the learning process may be insufficient. In this context, the use of kernel functions for graphs arises as an attractive way to deal with such structured objects. On the other hand, several entities in computational biology applications, such as gene products or proteins, may be naturally represented by graphs. Hence, the demanding need for algorithms that can deal with structured data poses the question of whether the use of kernels for graphs can outperform existing methods to solve specific computational biology problems. In this dissertation, we address the challenges involved in solving two specific problems in computational biology, in which the data are represented by graphs. First, we propose a novel approach for protein function prediction by modeling proteins as graphs. For each of the vertices in a protein graph, we propose the calculation of evolutionary profiles, which are derived from multiple sequence alignments from the amino acid residues within each vertex. We then use a shortest path graph kernel in conjunction with a support vector machine to predict protein function. We evaluate our approach under two instances of protein function prediction, namely, the discrimination of proteins as enzymes, and the recognition of DNA binding proteins. In both cases, our proposed approach achieves better prediction performance than existing methods. Second, we propose two novel semantic similarity measures for proteins based on the gene ontology. The first measure directly works on the gene ontology by combining the pairwise semantic similarity scores between sets of annotating terms for a pair of input proteins. The second measure estimates protein semantic similarity using a shortest path graph kernel to take advantage of the rich semantic knowledge contained within ontologies. Our comparison with other methods shows that our proposed semantic similarity measures are highly competitive and the latter one outperforms state-of-the-art methods. Furthermore, our two methods are intrinsic to the gene ontology, in the sense that they do not rely on external sources to calculate similarities.

In 2003, Pomeroy et al. published a research study that described a gene expression based prediction of central nervous system embryonal tumour (CNS) outcome. Over a half of decade, many models and approaches have been developed based on experimental data consisting of 99 samples with 7,129 genes. The way, how meaningful knowledge from these models can be extracted, and how this knowledge for further research is still a hot topic. This thesis addresses this and has developed an information method that includes modelling of interactive patterns, important genes discovery and visualisation of the obtained knowledge. The major goal of this thesis is to discover important genes responsible for CNS tumour and import these genes into a well structured knowledge framework system, called Brain-Gene-Ontology. In this thesis, we take the first step towards finding the most accurate model for analysing the CNS tumour by offering a comparative study of global, local and personalised modelling. Five traditional modelling approaches and a new personalised method – WWKNN (weighted distance, weighted variables K-nearest neighbours) – are investigated. To increase the classification accuracy and one-vs.-all based signal to- noise ratio is also developed for pre-processing experimental data. For the knowledge discovery, CNS-based ontology system is developed. Through ontology analysis, 21 discriminate genes are found to be relevant for different CNS tumour classes, medulloblastoma tumour subclass and medulloblastoma treatment outcome. All the findings in this thesis contribute for expanding the information space of the BGO framework.

The advent of microarray technology, allowing measurement of gene expression levels for thousands of genes in parallel, has made possible experiments designed to investigate the genetic control of variation in gene expression level (described in the literature as ???genetical genomics??? or ???eQTL??? experiments). Published results from these studies, in yeast and in mice, show that genetic variation is an important factor in gene regulation, and furthermore that individual polymorphisms modify the expression level of many genes. The concern of this thesis is the bioinformatic analyses of the expression level and genotype data sets that are the raw material for these studies. In particular this thesis addresses the two issues of detection of artefactual effects, and maximizing the information that can be extracted from the data. It is shown that while a polymorphism affecting the expression of many genes may be readily detected, care must be taken to determine whether the detected effect is genuinely one of genetic control of expression level, rather than the effect of correlations in measured expression level not of genetic cause. A significance test is devised to distinguish between these cases. The detection of artefactual correlation is explored further in the reanalysis of the published data from a large yeast study. A critique is given of the permutation method used to ascribe genetic control as the cause of inter gene expression level correlation. The presence of some degree of artefactual correlation is shown, and novel methods are presented for identifying such artefacts. To extend the analyses that may be applied to eQTL data, an algorithm is presented for determining secondary eQTLs for gene expression level (as opposed to a single primary QTL), along with a significance test for the putative QTL found. The technique is demonstrated on a large public data set. In addition to the use for which they are intended, the data sets generated for eQTL studies provide opportunities for additional analyses. In this thesis a method is developed for calculating a genome wide map of meiotic recombination frequency from the genotype data for multiple segregant strains. The method is demonstrated on the published genotype data generated for a large yeast eQTL study.