The healthy function of complex tissues is dependent on a complex combination of cell types properly working together to maintain homeostasis. Diseases or stressful conditions frequently alter the normal mix of cell types found in a healthy tissue, either directly or by eliciting an immune response. The cell type composition of these tissues is, therefore, of natural interest to both researchers and clinicians.

However, quantifying cell type populations has proven to be a challenging and often expensive task. Traditional methods suffer from several limitations and have potential to introduce bias. FACS sorting has been a common approach for many years, but remains slow and expensive, making it difficult to apply to large studies. Single-cell methods are emerging and may become more cost effective in the future, but still present a prohibitive financial barrier for many labs. Moreover, both these technologies fail to capture cells with unusual morphologies. Neurons, myocytes, and adipocytes are too large, unusually shaped, or fragile to be reliably estimated by these methods.

As gene expression data has become more ubiquitous, interest in computational cell type quantification methods have gained interest and popularity. These approaches, termed cell type deconvolution, utilize knowledge of cell type specific gene expression to estimate cell type abundances in samples of unknown composition. However, gene expression deconvolution is a challenging problem, and accurate predictions are sensitive to a number of factors. Many approaches have emerged, but struggle to maintain accurate predictions when faced with novel data from varying platforms, tissue types, or species.

I have developed the Gene Expression Deconvolution Tool (GEDIT), a flexible, robust deconvolution tool that aims to overcome limitations still present in the field. GEDIT is designed to be flexible applicable to a wide range of cell types, platforms, and species. GEDIT utilizes novel techniques for selecting signature genes, which identifies genes with cell type specific expression patterns and improves the speed and accuracy of results. A transformation is also applied, in order to control for the effect of highly expressed genes and further improve quality of results. Lastly, GEDIT applies a linear regression to model the observed. I have applied GEDIT to a number of datasets, including the entire GTEx database.

In addition, I am also performing a large-scale benchmarking project, in which I compare 8 current benchmarking tools (with more to be added) on several datasets of known proportions. This includes a large clinical dataset, with over 5,000 blood samples taken directly from healthy individuals. Cell type quantification for this data has been carried out by physical means, specifically cell electrical impedance counting. This project is comprehensively evaluating the performance of these tools when used with several mixture and reference datasets.

In the functional genomics era, systems-level analysis of genomic and proteomic data quickens the pace of gene-function discovery. I created a program and collected cross-referenced database for the rice genome. This program utilizes sequence similarity as a way to annotate all genes with the gene ontology (GO) terms and performs a functional analysis on genes in rice, an economically important crop with limited functional analysis tools. For demonstration purpose, application of the tool to one gene list created from cytokinin oxidase/ dehydrogenase activity reveals two GO terms that suggest the relevance for cytokinin associated terms as cytokinins was used to generate the gene list. To my knowledge, while there are tools available on the web to interpret the biological function of a gene list, none are set up for use with rice protein sequences. My program overcomes this limitation by permitting the functional analysis of rice amino acid sequences. The example gene list and code are available to reproduce the results.

In multicellular organisms, there are many different cell types possessing the same genetic information, each performing a particular role. Many players, such as transcription factors, nucleosome positioning, histone post-translational modifications and non-coding RNAs, contribute to regulate the expression of specific genes, defining the specific cell state. Once the expression pattern is established during development, it is faithfully maintained throughout the life of an organism. In many organisms, a key component of this epigenetic regulation is a covalent modification of cytosines (5meC) in the genome, known as DNA methylation. DNA methylation is associated with repression of transcription if present in gene promoters and it can suppress the transcription of aberrant intragenic transcripts. The focus of my doctoral research has been the development of methods to map the distribution of 5meC genome-wide and understand how DNA methyltransferases are recruited to their targets. Since mammalian genomes are large, current approaches to map the methylation of the entire genome are expensive. Several methods have been developed to assess the methylation status of part of the genome. Some of them are based on enrichment probes, others on enzymatic digestion. Chapters 1 and 2 are based on methods that assess the methylation status of part of the genome. Reduced-Representation Bisulfite Sequencing (RRBS) captures the majority of CpG islands and promoters. Since only 1% of the genome is assessed with this technique, costs associated with sequencing are dramatically reduced. In chapter 1, this technique is used to discover methylation levels at specific CpG sites associated with complex disease traits. Most of the CpG sites in the genome are methylated and do not have variable methylation levels between different cell types, suggesting that some of the fragments isolated by RRBS do not provide useful information. In order to overcome this limitation, we improved an existing method, Methylation-sensitive Restriction Enzyme-seq (MRE-seq), which enriches for regions poorly methylated (approximately 20% of the genome). This method, called MRE-BS (MRE-Bisulfite Sequencing), is described in chapter 2. The costs are similar to RRBS, but the development of a multiple regression model has allowed us to estimate differential methylation between two samples across 60% of the genome.

Chapter 3 focuses on how de novo DNA methyltransferases are recruited to their target sites. This work has shown that the murine DNMT3b is guided by histone post-translational modifications, both in yeast and primordial germ-cells. In general, DNMT3b is absent from regions marked by H3K4me3 and it is recruited in gene-bodies by H3K36me3.

The last chapter focuses on the presence of 5meC in messenger RNA, the function of which is still unknown. We discovered several hundred putative methylation sites that are associated with predicted secondary structures in mRNAs. This finding might be explained either by the recruitment of RNA methyltransferases (RMTs) by structural motives or by the limitations of the method utilized. More experiments are needed to understand what signal is needed for the specific recruitment of RMTs to their target sites and what is the function of this mark.

Various epigenetic clocks have been constructed using DNA methylation data, using regression models to estimate age from DNA methylation patterns. To overcome the constraints imposed by these epigenetic clocks, the Epigenetic Pacemaker (EPM) clock is built to predict an individual’s epigenetic state in an unbiased non-linear manner. The EPM clock models the initial methylation value and rate of change in methylation at each methylation locus, enabling an intuitive interpretation of coefficients of selected sites. Since the EPM model is computationally heavy, selecting informative loci in the preprocessing step is necessary. We selected model sites using either a novel randomized ridge regression selection method or the Pearson Correlation Coefficient (PCC) method. The PCC metric achieved higher performance and was used as the site selection method when applying EPM clock to a schizophrenia data set. In this data set, age acceleration predicted by EPM model was positively correlated with schizophrenia status and sex as a male. By experimenting with different EPM models, we conclude that a full model using all samples to build an EPM model generates stable epigenetic state predictions. Building EPM model using more sites with higher PCC values correlated with phenotype traits is more informative.

With the development of next-generation sequencing technologies, we can detect numerous genetic variants associated with many diseases or complex traits over the past decades. Genome-wide association studies (GWAS) have been one of the most effective methods to identify those variants. It discovers disease-associated variants by comparing the genetic information between controls and cases. This approach is simple and effective and has been used by many studies.

Before performing GWAS, we need to detect the genetic variants of the sample population. A subset of these variants, however, may have poor sequencing quality due to limitations in NGS or variant callers. In genetic studies that analyze a large number of sequenced individuals, it is critical to detect and remove those variants with poor quality as they may cause spurious findings. Here, I will present ForestQC, an efficient statistical tool for performing quality control on variants identified from NGS data by combining a traditional filtering approach and a machine learning approach, which outperforms widely used methods by considerably improving the quality of variants to be included in the analysis.

Once this association is identified, the next step is to understand the genetic mechanism of rare variants on how the variants influence diseases, especially whether or how they regulate gene expression as they may affect diseases through gene regulation. However, it is challenging to identify the regulatory effects of rare variants because it often requires large sample sizes and the existing statistical approaches are not optimized for it. To improve statistical power, I will introduce a new approach, LRT-q, based on a likelihood ratio test that combines effects of multiple rare variants in a nonlinear manner and has higher power than previous approaches. I apply LRT-q to the GTEx dataset and find many novel biological insights.

Recent studies have shown that omics data can be used for automatic disease diagnosis with machine learning algorithms. I will introduce an accurate and automated machine learning pipeline for the diagnosis of atopic dermatitis (AD) based on transcriptome and microbiota data. I will demonstrate that this classifier can accurately differentiate subjects with AD and healthy individuals. It also identifies a set of genes and microorganisms that are predictive for AD. I will show that they are directly or indirectly associated with AD.

3’ RNA sequencing provides an alternative to whole transcript analysis. However, we do not know a priori the relative advantage of each method. Thus, a comprehensive comparison between the whole transcript and the 3’ method is needed to determine their relative merits. Single cell RNA sequencing (scRNA-seq) enables the profiling of the transcriptomes of individual cells. Cell type identification is one of the major goals in scRNA-seq. Current methods for assigning cell types have several limitations, such as unwanted sources of variation that influence clustering and a lack of canonical markers for certain cell types. Thus, new methods need to be developed. We first used two commercially available library preparation kits, the KAPA Stranded mRNA-seq kit (traditional method) and the Lexogen QuantSeq 3’ mRNA-seq kit (3’ method), to determine the advantages and disadvantages of these two approaches. We found that the 3’ RNA-seq method detected more short transcripts than the whole transcript method. With regard to differential expression analysis, we found that the whole transcript method detected more differentially expressed genes, regardless of the level of sequencing depth. We then developed ACTINN (Automated Cell Type Identification using Neural Networks), which employs a neural network to predicts cell types for scRNA-seq datasets. We trained and tested ACTINN on multiple datasets, the results showed that ACTINN is fast and accurate, and should therefore be a useful tool to complement existing scRNA-seq pipelines. Lastly, we performed scRNA-seq to study gene networks associated with host defense comparing lesions from reversal reaction vs. lepromatous lesions from leprosy patients. We constructed an antimicrobial ecosystem by integrating the IFNG and IL1B antimicrobial targets with the cell-cell co-abundance in lesions, which revealed that the interaction of dendritic cells, macrophages, T cells, keratinocytes and fibroblasts contributes to the capacity of granulomas to eliminate the pathogen in leprosy.

DNA methylation patterns change with increasing age, therefore contributing to some age related diseases. Epigenetic drift over time might result in measurable differences between biological and chronological age, so epigenetic changes are shown to be reflective of an individual's lifestyle. Consequently, the estimation of the epigenetic state can help with medical and biological research, and may act as functional biomarkers of disease before clinical threshold is reached. A variety of epigenetic clocks have been constructed with regression models to approach age estimation from individuals’ DNA methylation patterns. This study involves the use of the Epigenetic Pacemaker Model (EPM), an implementation of a fast conditional expectation maximization algorithm to model non-linear epigenetic trait associations directly without transformation of the phenotype of interest. EPM models the initial methylation value as well as the rate of change at each locus, therefore allows an intuitive interpretation of the selected sites. Since EPM is computationally heavy, it will be useful to build a succinct model for the users who would like to use EPM in their studies. Therefore, the aim of the project is to make EPM more accessible for users to analyze their epigenetic data, at the same time conserving the accuracy in the epigenetic data prediction. The project is primarily based on python code, and uses the R shiny to develop a website app in order for users from other places to upload their data and implement the EPM graph displaying the logarithmic trend of the relationship between epigenetic state and chronological age.

Besides DNA sequences, the genes are regulated by epigenomic mechanisms. Advances in high-throughput sequencing technologies have enabled the generation of huge amount of epigenomic data sets. Those epigenomic big data then requires the application of sophisticated computational approaches and statistical algorithms. My dissertation then focuses on mining and integrating epigenomic big data to inform novel biological mechanisms behind those datasets. The first research project compares the binding patterns of pluripotent regulatory factors, i.e. Oct4, Sox2, Klf4 and c-Myc, between human and mouse in induced pluripotent stem cells. The result suggests the genome-wide regulatory mechanisms are conserved between those two species, but the detailed transcriptional mechanisms are diverged. The second research project analyzes the temporal expression data from embryonic stem cells to cardiomyocytes. The results in this project then identify regulators, including transcription factors and long intergenic non-coding RNAs, which are strongly associated with the cardiogenesis differentiation process. The third research project integrates datasets of DNA methylation and histone modification in 35 human cell types. The result shows histone modifications, especially for H3K4me3, are highly predictable of DNA methylation. As a summary, my dissertation analyzes and integrates epigenomic big data in biological context related with embryonic stem cells and induced pluripotent cells, provides and discoveries novel insights to understand epigenetic regulation of gene expression.

Epigenetics and proteomics have emerged as powerful fields with the potential to transform our understanding of the relationship between lifestyle factors and health outcomes. This paper utilizes data from an exploratory, cross-sectional study conducted by Prosper DNA Inc. to identify proteomic and epigenetic biomarkers that can predict clinical lifestyle traits related to individual health. The study collected data from two groups, one representing healthy individuals and the other of unhealthy individuals with specific lifestyle characteristics. Imputation, Principal Component Analysis (PCA), and correlation analysis were performed on the proteomic and methylation data, which resulted in 10 principal components (PCs) each and 3 correlation matrices. Linear regression models were developed using proteomic or methylation PCs as predictors for each lifestyle trait. The models were evaluated using Leave-One-Out-Cross-Validation and Pearson correlation coefficient (r) to determine significance and confirm the accuracy of the models. 19 significant proteomic models and 28 significant methylation models were identified. Notably, 15 models across the proteomic and methylation models were directly associated with the original selection criteria, such as body mass index (BMI), fitness, and dietary habits. Through a correlation analysis, PC3, PC9, and PC9 of the proteomic PCs and MPC1, MPC3, and MPC4 of the methylation PCs were selected for PC loading analysis based on the strength and significance of the correlation between a given PC and the traits associated with the biomarkers. The most influential proteins and CpG sites were extracted from the loading and passed through STRING and Cistrome, respectively, to gather protein network information and transcription factor information. For the proteomic PCs, the proteins were involved in networks relating to platelet activation/coagulation, inflammatory response, and the regulation of proteolytic activity. For the methylation PCs, the transcription factors that bind to the CpG sites showed some relation to adipogenesis and tumor suppression. Future steps should include deeper analyses of the individual proteins, CpG sites, and transcription factors to provide concrete validation for the trends observed in this study.

DNA methylation, the addition of a methyl group at the fifth carbon of the pyrimidine ring resulting in 5-methylcytosine (5MC), is influenced by the cellular environment; continued exposure to environmental stimuli will result in detectable DNA methylation changes. These detectable changes have been leveraged to develop DNA methylation based predictive models for age and health. DNA methylation is commonly assayed through the use of bisulfite conversion, where unmethylated cytosine is deaminated to form uracial while 5MC is unchanged, followed by high throughput sequencing. Processing of bisulfite sequencing data is computationally demanding due to the asymmetrical nature of bisulifte sequencing data. Chapter 1 introduces a bisulfite sequencing processing platform, BSBolt, that is a signficant improvement over previous tools in terms of processing time and alignment accuracy. BSBolt and targeted bisulfite sequencing are utilized in chapter 2 to look into epigenetic suppression of transgenic t-cell receptor (TCR) expression in adoptive cell transfer therapy. The chapter 2 study shows that accumulation of DNA methylation over the viral vector promoter used to introduce the TCR sequence is associated with decreased expression despite persistence of the TCR sequence over time. Chapter 3 introduces an evolutionary framework, the epigenetic pacemaker (EPM), for modeling epigenetic aging. The EPM is a departure from the penalized regression based approaches broadly used in the field. The EPM attempts to minimize error across the observed methylation profiles rather than age prediction error. This approach allows the EPM to model nonlinear epigenetic aging across human lifespan as shown in chapter 4. Chapter 5 compares the EPM to penalized regression approaches and shows the EPM is more sensitive for detecting biological signals associated with epigenetic aging.