Mammalian hematopoiesis, the process by which blood cells develop from a common progenitor cell, is evolutionarily conserved and unfolds continuously over the lifetime of an organism. Development and differentiation, as in hematopoiesis, are mediated through the tightly controlled activities of transcription factor proteins, which regulate the expression levels of various cell type-defining genes. In turn, these transcription factors are modulated by genes that alter chromatin structure, thereby influencing the fundamental expressivity of genes. This cyclical and dynamic process, operating in a state of disequilibrium, forms the complex gene regulatory network underpinning hematopoiesis. Sequencing-based genomics technologies have empowered assays that measure various aspects of gene regulation, including gene expression, chromatin accessibility, and chromatin occupancy. This thesis is composed of six chapters. The first chapter of this thesis reviews key concepts required to appreciate the challenges of studying gene regulation. The second chapter focuses on developing methods to analyze the data generated from gene expression measurements at single-cell resolution. Insights from single-cell measurements have allowed us to deconvolute complex biological processes with unprecedented resolution. However, single-cell measurements provide only a transient snapshot of a cell’s state, destroying the cell upon measurement. Cell destruction is inherently linked to measurement and thus obfuscates a causal linkage between observed states within a population of cells. This fundamental disconnection presents a challenge to unraveling dynamic processes such as hematopoiesis, where temporal resolution is crucial. While single-cell omics measurements are generally limited to the capture of independent cell snapshots, it was realized that sufficient sampling depth along a dynamic cell process enables the reconstruction of that process. Thus, many methods were developed to reconstruct cell trajectories from single-cell data. These methods drove the field to converge on effective techniques for dimension reduction and ordering cells along a biological process. However, they largely failed to offer meaningful mechanistic insights toward the underlying gene expression states along the reconstructed trajectories. More recently, drift-diffusion equations were demonstrated as an appropriate framework to model single-cell dynamics from first principles. Existing solutions leverage this framework to describe the deterministic drift term-associated dynamics. However, these solutions model diffusion as invariant to cell state, limiting our ability to study the stochastic dynamics that underpin cell decision-making. To address these shortcomings, we describe scDiffEq, a generative modeling framework. scDiffEq seeks to capture and study the balance between deterministic and stochastic biology by learning a neural stochastic differential equation. We use lineage-traced single-cell data to demonstrate the prediction of cell fates from multipotent progenitors during hematopoiesis. The fate prediction capability of scDiffEq is competitive with leading methods. We also describe an improved ability to reconstruct unseen cell distributions along a trajectory. We study the model’s ability to predict the effects of in silico perturbations on multipotent progenitor cells. These perturbation experiments show that scDiffEq can partially capture the perturbed dynamics imparted by CRISPR-based perturbations in early hematopoiesis. We also generalize this approach to data lacking temporal annotation. In the third chapter, we build on the context of introducing functional perturbations to study gene regulation, shifting our focus towards the role of Lysine-specific demethylase 1 (LSD1) in acute myeloid leukemia (AML). LSD1 demethylates Histone H3 Lysine 4 (H3K4), an epigenetic modification generally associated with transcriptional activation. AML proliferation is dependent on LSD1. Inhibitors of LSD1 that act on its enzyme activity, irreversibly binding to its FAD cofactor, are in clinical trials at the time of this writing. LSD1 is not somatically mutated in AML, and the mechanism by which AML is dependent on LSD1 was ambiguous. It was widely presumed that inhibition of LSD1 rewired gene expression in AML, such that the LSD1 demethylase activity was critically important. To more precisely understand LSD1 function – and its inhibition – in AML, we employed a CRISPR-based tiling mutagenesis strategy to highlight functionally critical domains of the LSD1 protein. This method, CRISPR-suppressor scanning, combines massively parallel mutagenesis with chemical suppression to asymmetrically deplete cells harboring mutations that hinder endogenous function while at the same time enriching cells with mutations that confer resistance to the chemical suppressor. This approach, paired with functional genomic and biochemical follow-up studies, demonstrates that the LSD1 enzyme activity is dispensable in AML. Our work goes on to establish LSD1 as a critical chaperone for the transcription factor GFI1B. Maintenance of the LSD1-GFI1B interaction on chromatin is essential to AML survival. We definitively demonstrate that disruption of the LSD1-GFI1B interaction is the primary role of mechanistic inhibitors of LSD1 in AML. This study also advances the development of CRISPR-SAR, multiplexing CRISPR suppressor scanning with medicinal chemistry to map the structure-activity relationships of LSD1 inhibitors and protein variants of LSD1. Finally, our CRISPR-suppressor scan of LSD1 led to a surprising role for the intrinsically disordered terminal tail of LSD1. Despite its previously underestimated significance, several mutations in the N-terminal conferred resistance to inhibition of LSD1. Others subsequently investigated this observation to find that the intrinsically disordered N-terminal tail of LSD1 serves as a conformational switch that dynamically tunes the repression of active cis-regulatory elements in AML. The fourth chapter intersects the increasing utility of scATAC-seq (single-cell Assay for Transposase Accessible Chromatin using sequencing) to profile the chromatin states of individual cells. scATAC-seq directly samples DNA, of which there are only two copies in a normal diploid human cell, creating an inherent sparsity to the recovered reads per cell. Considering such challenges, an optimal strategy for feature selection, analytical transformation, and dimension reduction required clarification. These methodological uncertainties motivated an effort to benchmark ten computational methods scATAC-seq available at the time of publication against 13 synthetic and real datasets. Ultimately, methods were judged on their ability to discriminate cell types through clustering based on a cell-by-feature matrix. Clustering performance was cross-referenced against practical considerations like computational complexity. In general, we found that discretization of the genome into binned regions performed well against various compositions of cell populations, noise, and read coverage. Since this study (2019) and at the time of this writing, improved methods have been developed. However, the citation record of this benchmark study indicates its utility to the field as a reference point in the maturation of methods development for scATAC-seq. The fifth chapter of this thesis further develops and formalizes our intuition for representing single-cell data using knowledge graphs. Generating datasets with multi-omic assays, such as joint scRNA-seq and scATAC-seq, has become increasingly common. Most computational methods for single-cell analysis are built for a specific task (e.g., batch integration). Inherently, these methods cannot infer relationships across feature types. Additionally, clustering-based feature discovery is pervasively used in single-cell analysis workflows, though it remains prone to user bias and error. Our group proposed SIMBA (single-cell embedding along with features) as a unifying and versatile single-cell embedding method to overcome these limitations. SIMBA represents cells and their comprising features as nodes in a graph, encoding relationships between entities as graph edges. SIMBA then uses a graph embedding algorithm from social networking research to embed all nodes into a common latent space. This space then serves as a unified representation from which biological queries and single-cell analysis tasks may be performed. The final chapter of this thesis presents a forward-looking perspective on how future generations of deep learning models might be employed to study biology. It envisions a paradigm shift towards foundation models that leverage the vast collection of data generated by the global scientific community, moving beyond isolated measurements and analyses. Such models might lead to a more cohesive and comprehensive understanding of biological systems. In addition to addressing contemporary challenges, the methodologies and insights developed in this thesis serve as precursors along the route of achieving such a vision.