Investigating Transcriptional Regulation with Single-Cell Genomics and AI System Development 18 views

Cell identity and cell-state transitions are controlled by gene regulatory programs that determine how the genome is used in a specific biological context. Single-cell genomics now makes it possible to measure gene expression and chromatin accessibility across heterogeneous tissues, but these measurements do not directly reveal which transcriptional regulators establish or maintain a cell state. This dissertation investigates transcriptional regulation by combining single-cell genomics, computational regulator inference, biological applications in development and cancer, and AI system development for evidence-grounded interpretation. To examine how chromatin accessibility relates to lineage fate during early organ development, I analyzed the mouse primitive gut tube using single-cell ATAC-seq. Chromatin accessibility patterns predicted future organ identity before clear anatomical boundaries were fully established. These patterns were enriched for lineage-associated transcription factor motifs and closely aligned with transcriptome-defined organ identities. Functional studies further showed that lineage-defining transcription factors can reshape chromatin accessibility and alter organ fate. Together, these results show that chromatin landscapes are closely linked to early regulatory decisions during development. To infer regulators from single-cell data, I developed BARTsc, a computational method for predicting functional transcriptional regulators from scRNA-seq, scATAC-seq, and single-cell multiome data. BARTsc extends the ChIP-seq-informed strategy of BART by connecting cell-state-specific features with public transcriptional regulator binding profiles. This design allows BARTsc to move beyond marker genes and motif enrichment by ranking regulators whose binding profiles best explain cell-state-specific programs. Across benchmark datasets, BARTsc accurately identified known functional regulators and outperformed existing methods across multiple cell types and data modalities. Its application to pancreatic ductal adenocarcinoma further identified NELFA as a candidate regulator of a highly proliferative tumor-cell population, which was supported by experimental knockdown. I then applied single-cell multiome profiling and BARTsc to matched primary and metastatic pancreatic cancer models. This analysis revealed site-associated but overlapping tumor-cell states across two orthotopic patient-derived models. Cross-model comparison identified recurrent states related to adhesion plasticity, stress adaptation, and proliferation. BARTsc linked these states to distinct regulatory programs, while trajectory analysis suggested a recurrent transition from adhesion-associated states toward proliferative states. Clinical analysis further suggested that trajectory-derived genes capture features of metastatic progression and patient outcome. Driven by the need of ground truth knowledge for the evaluation of BARTsc, I also developed GeneKnow, an AI system for source-grounded literature synthesis in gene-context analysis. GeneKnow uses controlled retrieval, evidence selection, hierarchical summarization, self-verification, and deterministic citation construction to synthesize literature evidence for specific genes in defined biological contexts. By linking generated claims to traceable source evidence, GeneKnow helps convert candidate gene and regulator lists into literature-supported summaries. Together, these studies show how single-cell measurement, ChIP-seq-informed regulator inference, and source-grounded literature synthesis can be combined to study gene regulatory programs in development and cancer.

PHD (Doctor of Philosophy)

single-cell genomics; Computational Biology; artificial intelligence; Bioinformatics; transcriptional regulation

English

2026-07-31