Visualizing spatial assay data in anatomical images is vital for understanding biological processes in cell, tissue, and organ organizations. Technologies requiring this functionality include traditional one-at-a-time assays, and bulk and single-cell omics experiments, including transcript, protein and metabolite profiling. To address this need, this Ph.D. project developed the \textit{spatialHeatmap} software environment. As an integrated multipurpose visualization and analysis toolbox, \textit{spatialHeatmap} provides flexible solutions for these needs by allowing users to visualize spatial biological data in adequately formatted anatomical images from public collections or custom images, and extend them with a series of large scale data mining graphics. \textit{spatialHeatmap} colors the spatial features (\textit{e.g.} tissues) annotated in the images according to the quantitative abundance levels of measured biomolecules (\textit{e.g.} mRNAs) using a color key. This core functionality of the package is called a spatial heatmap (SHM) plot. Single-cell data can be co-visualized in composite plots that combine SHMs with embedding plots of high-dimensional data. The resulting spatial context information is essential for gaining insights into the tissue-level organization of single-cell data, or vice versa. Functionalities for integrated visualization of cellular compositions of bulk data with deconvolution algorithms are also included. Alternatively, a new co-clustering method automates the association of unlabeled single-cell data with the corresponding source tissues. Spatially resolved single-cell data from more recently developed spatial sequencing technologies can also be co-visualized with the corresponding anatomical SHMs. Additional core functionalities include large-scale identification of biomolecules with spatially selective abundance patterns, and clusters of biomolecules sharing similar abundance profiles.

\textit{spatialHeatmap} is a new software for visualizing spatial assay data in anatomical images with extensions for large-scale data mining and co-visualization of single-cell data. It integrates many novel functionalities into a generic spatial visualization toolbox that is flexible, extensible, and object-oriented by design. A graphical and a command-line interface are provided to support the needs of both non-expert and computational users, respectively. It is distributed as a free, open-source Bioconductor package (\href{https://bioconductor.org/packages/release/bioc/html/spatialHeatmap.html}{https://bioconductor.org/packages/ spatialHeatmap}) that users can install on personal computers, shared servers, or cloud systems. Because of its rich functionalities and broad applicability, \textit{spatialHeatmap} is expected to be widely adopted by the spatial biology community.

This dissertation is about the development of gene expression signature (GES) search methods and their application in drug discovery, specifically for promoting healthy aging. GES searching is a powerful technology facilitating the identification of drugs for treating diseases and drug repurposing. This is achieved by identifying drugs in GES databases inducing signatures similar to query GESs obtained from diseased samples or drug treatments. The new connections are useful for developing pharmacological interventions. This dissertation is divided into the following three components. First, I developed the signatureSearch R/Bioconductor package that integrates existing and novel methods for GES searching and functional enrichment analysis (FEA). Subsequently, I tested the performance of different GES search methods. They represent the first systematic performance tests of these methods in the field. Second, I applied signatureSearch to the human healthy aging field to reveal insights into longevity associated (LA) drugs and their targets by searching the Integrated Network-based Cellular Signatures (LINCS) database. For this, I assessed the performance of LINCS drugs, inducing GESs representative for their mechanism of action (MOA), by computing for each MOA a recall score based on the GES similarity of the corresponding drugs. The obtained recall scores were used to prioritize LA drugs in the downstream discovery. LA MOA categories along with the corresponding drugs were identified by querying LINCS with GESs of drugs present in LINCS and DrugAge, and scoring the enrichment of each MOA. The corresponding LA pathways were identified via global mapping of targets of LA drugs and MOA categories. Next, I searched LINCS with the GESs from 11 well-studied LA drugs and one longevity phenotype. To identify LA pathways, the targets of the newly identified LA drugs were used for FEA. The results from the three steps were integrated and then interrogated with a combinatorial approach to select the most reliable set of LA drugs and targets. Collectively, this study identified a list of drugs, targets and pathways useful for pharmacological lifespan extension strategies. Third, I developed several data packages to incorporate in signatureSearch detailed annotations of drugs and targets from different community databases.

De novo genome and transcriptome assemblies of next generation sequences (NGS) are important for many genomics applications of unsequenced organisms. Both assembly types present many challenges owing to (i) the large amount of data to process, (ii) sequencing errors and (iii) the complexity of transcriptomes and genomes. The latter is a result of alternative splice events, variable and incomplete representations of transcripts in RNA-Seq libraries, while repetitive regions in genomes complicate their assembly. Usually, de novo assemblies result in thousands of short and incomplete transcripts (transfrags) or genomic sequences (contigs or scaffolds) while requiring a large amount of processing time and memory. However, with decreasing NGS costs reference genomes of many species have become available recently that can be used to guide and improve de novo assemblies. Here, we introduce two reference assisted algorithms BRANCH and AlignGraph. BRANCH improves transcriptome assemblies guided by genomic contigs from the same species or reference genes from a closely related species. AlignGraph improves genome assemblies with help provided by a closely related reference genome. In addition, we introduce the short read clustering algorithm SEED that is useful as a preprocessing tool in de novo assemblies by reducing their time and memory requirements.

SEED joins sequences into clusters that can differ by up to three mismatches and three overhanging residues from their virtual center. It is based on a modified spaced seed method, called block spaced seeds. Its clustering component operates on the hash tables by first identifying virtual center sequences and then finding all their neighboring sequences that meet the similarity parameters. SEED can cluster 100 million short read sequences in <4 h with a linear time and memory performance. When using SEED as a preprocessing tool on genome/transcriptome assembly data, it was able to reduce the time and memory requirements of the Velvet/Oasis assembler, for the datasets used in this study, by 60-85% and 21-41%, respectively. In addition, the assemblies contained longer contigs than non-preprocessed data as indicated by 12-27% larger N50 values. Compared to other clustering tools, SEED showed the best performance in generating clusters of NGS data similar to true cluster results with a 2- to 10-fold better time performance. While most of SEED's utilities fall into the preprocessing area of NGS data, our tests also demonstrate its efficiency as stand-alone tool for discovering clusters of small RNA sequences in NGS data from unsequenced organisms.

BRANCH's input includes assembled RNA reads, genomic sequences (e.g. contigs) and the RNA reads themselves. It uses a customized version of BLAT to align the transfrags and RNA reads to the genomic sequences. After identifying exons from the alignments, it defines a directed acyclic graph and maps the transfrags to paths on the graph. It then joins and extends the transfrags by applying an algorithm that solves a combinatorial optimization problem, called the Minimum weight Minimum Path Cover with given Paths. In performance tests on real data from Caenorhabditis elegans and Saccharomyces cerevisiae, assisted by genomic contigs from the same species, BRANCH improved the sensitivity and precision of transfrags generated by Velvet/Oases or Trinity by 5.1-56.7% and 0.3-10.5%, respectively. These improvements added 3.8-74.1% complete transcripts and 3.8-8.3% proteins to the initial assembly. Similar improvements were achieved when guiding the BRANCH processing of a transcriptome assembly from a more complex organism (mouse) with genomic sequences from a related species (rat).

AlignGraph is an algorithm for extending and joining de novo assembled contigs or scaffolds guided by closely related reference genomes. It aligns paired-end (PE) reads and pre-assembled contigs or scaffolds to a close reference. From the obtained alignments, it builds a novel data structure, called the paired-end multi-positional de Bruijn graph. The incorporated positional information from the alignments and PE reads allows us to extend the initial assemblies, while avoiding incorrect extensions and early terminations. In our performance tests, AlignGraph was able to substantially improve the contigs and scaffolds from several assemblers. For instance, 28.7-62.3% of the contigs of Arabidopsis thaliana and human could be extended, resulting in improvements of common assembly metrics, such as an increase of the N50 of the extendable contigs by 89.9-94.5% and 80.3-165.8%, respectively. In another test, AlignGraph was able to improve the assembly of a published genome (Arabidopsis strain Landsberg) by increasing the N50 of its extendable scaffolds by 86.6%. These results demonstrate AlignGraph's efficiency in improving genome assemblies by taking advantage of closely related references.

This dissertation presents three projects that are described in individual chapters. The first two projects are about software development for reproducible research involving big data, and the third project is about large-scale discovery of drug-like natural compounds. The following briefly summarizes each chapter.

Chapter 1 introduces the general research area of each project. These introductions focus on the status of the field, current challenges and future needs. At the end of each project introduction, an overview is given about the specific solutions proposed and implemented by the research of this dissertation.

Chapter 2 describes the design and development of systemPipeR. This generic software environment provides flexible utilities for designing, building and running automated end-to-end analysis workflows for a wide range of research applications. Important functionalities include a uniform workflow interface across different data analysis applications, automated report generation, and support for running both R and command-line software, such as big data analysis tools, on personal computers, HPC clusters, and cloud systems.

The third chapter introduces systemPipeShiny. This web system extends the systemPipeR workflow environment with a versatile graphical user interface provided by a Shiny App. It allows non-R users to run many of systemPipeR's workflow design, control, and visualization functionalities interactively without requiring command-line knowledge. It also integrates highly interactive visualization functionalities and a graphics workbench.

Chapter 4 is a large-scale discovery project of bioactive natural compounds. First, the known structures and annotations of about 0.5 million natural compounds and about 4,000 known drugs were organized in a rational database. Second, the assembled database was used for virtual screening, where a series of supervised and non-supervised machine learning methods was applied to predict drug-like natural product candidates. This also identified drugs that have natural compound alternatives with identical structures and vice versa. Third, known drug-target annotations were used to systematically identify which molecular mechanisms and disease processes are perturbable by bioactive natural compounds. This study identified a large panel of new bioactive natural compounds with interesting applications in human health, including healthy aging and anticancer.

Background

Results

Conclusions

Characterizing the functional, structural, and evolutionary relationships of biological sequences is an important task in modern genomics and computational biology. Most of these applications involve the assembly of sequence families by similarity searching, subsequent formation of multiple sequence alignments (MSAs) and downstream phylogenetic analyses. Especially, MSAs play a central role in this modeling workflow. Thus, the quality of the MSAs is of critical importance for its success. In this study I present an approach to improve the quality of MSAs by using a sequence family revision approach that can automatically remove false positive candidates from sequence families and then recompute an improved MSA. The approach is able to combine sequence-level scores from two MSA scoring methods, norMD and GUIDANCE. It automatically selects an optimized score threshold for removing sequences from MSAs. To test the performance of this method, I developed several automated procedures to add to curated MSAs from the CDD database controlled numbers of randomly selected nonmember sequences. Then I performed Receiver Operating Characteristic (ROC) analysis on the classification results incorporating automatic threshold selection approaches. Surprisingly, the sequence-level scores, provided by the two MSA scoring methods, were less successful than a simple all-against-all BLAST-based pairwise alignment scoring approach. However, I was able to improve one of the MSA scoring methods by extending it with a dynamic threshold selection approach. The extended method outperformed the performance of the BLAST-based method in detecting false positives in sequence families.

One of the primary goals of bioinformatics is the identification of the function

of genes. The most reliable way of doing this is through experimentation. However,

this is a very slow and expensive process. While this is necessary in

the beginning and will continue to be necessary for special cases,

it becomes impractical when

one considers the number of different genes encoded in the genomes of every living

organism. A faster way is to instead identify the function of genes

by comparing them to the smaller set of genes with known function. This

comparison may be based on many different kinds of data, including sequence

similarity and gene expression data.

The goal of this dissertation is to provide tools to aid in the identification

of the function of unknown genes. To that end, we first present

a study in which gene expression data was used to annotate many unknown

genes by clustering the expression data. We then present a tool for

clustering gene expression data while also identifying short areas

of high sequence similarity (motifs) among members of the clusters.

Finally, we present a tool for identifying the functionally relevant

sub-sections of protein sequences. These sub-sections can then be used to find

proteins containing similar sub-sections, even though the rest

of the protein may be quite different. This tool can thus find

more distantly related proteins sharing functionally relevant features.

In this dissertation, I present several strategies to leverage experimental data towards a quantitative understanding of small molecule bioactivity that can inform the discovery of drug candidates. I review two cheminformatics tools I co-developed, present a comprehensive cross-target analysis of large public small molecule bioactivity data, and present a mathematical model of leptin transport across the mammalian blood-brain barrier.

First I present ChemMine Tools, a web service which provides both programmable and interactive online interfaces to a diverse set of analysis tools useful for analyzing small molecule structural data. ChemMine Tools allows users to import a set of small molecule structures, compute pairwise compound similarities, search for similar compounds, cluster compounds by structure or physical properties, and compute physicochemical properties.

The second software tool is bioassayR, a software package for large scale cross-target analysis of small molecule bioactivity profiles. bioassayR systematically analyzes data from thousands of screening experiments to identify target selective drug candidates and druggable protein targets. By simultaneously leveraging data from both custom small molecule screening efforts and public databases, bioassayR helps identify regions of the genome and proteome accessible to small molecule probes, elucidate novel mechanisms of action for bioactive molecules, and predict off-target effects which currently lead to a high attrition rate in drug discovery efforts.

The systematic cross-target analysis of public bioactivity data uses the bioassayR tool to analyze data from PubChem BioAssay. This groups small molecules into three groups based on an increasing number of active targets in these data- highly selective, family selective, and promiscuous binding, and find that FDA approved drugs are strongly represented among the family selective group. I also show that the compound-target space can be organized into biclusters, where shared activity tends to occur across protein targets sharing common Molecular Function Gene Ontology (MF GO) terms.

The leptin transport kinetic model extends current mathematical models of receptor endocytosis to transcytosis, and behaves similar to the experimentally observed dynamics of this system. A computational model is provided which allows for in-silico perturbation, to predict the potential effects of pathological states, or therapeutic small molecules.

Motivation

Results

Availability

Background

Results

Conclusions