Exploring Neurodevelopmental Origins of Pediatric Brain Cancer using Transferable Topic Modeling

Abstract

Brain cancers are the leading cause of cancer-related death in children and adolescents, with many childhood central nervous system (CNS) tumors originating during prenatal development. Understanding the developmental origins of pediatric brain cancers requires identifying transcriptional programs that persist from normal development into disease states. Despite remarkable advances in single-cell genomics, fundamental challenges remain in comparing cellular heterogeneity across datasets, particularly the inability of conventional clustering approaches to capture transitional cell states and the difficulty of integrating data across different sequencing technologies, timepoints, and biological contexts.Objectives. This dissertation develops and validates a transferable topic modeling framework to identify conserved transcriptional programs across developmental and disease contexts. The central hypotheses are that (1) transcriptional signatures present in neurodevelopment persist in pediatric brain cancers, and (2) these signatures can be uncovered using topic modeling while enabling systematic comparison across diverse datasets. Methods. We use Latent Dirichlet Allocation (LDA), a probabilistic framework from natural language processing, to single-cell transcriptomics by treating cells as "documents" and genes as "words." This approach represents each cell as a probabilistic mixture of transcriptional programs (topics) rather than assigning cells to discrete clusters, naturally capturing the continuous nature of developmental transitions. We combined LDA with Gene Set Variation Analysis (GSVA) to enable topic transfer across datasets, allowing topics learned from one context to be scored in independent datasets. The framework was applied across three biological contexts: (1) mouse embryogenesis to establish validity against comprehensive ground truth, (2) human cerebellar development and medulloblastoma to demonstrate developmental-cancer correspondence, and (3) transcript-level analysis of TARGET pediatric cancers to reveal isoform-specific disease associations. Results. In mouse organogenesis, topic modeling successfully recovered 50 transcriptional programs organized into 11 lineage-specific groups that recapitulated known developmental hierarchies while revealing transitional states invisible to conventional clustering. Topics subdivided major cell lineages (hematopoietic, neural, epithelial) and captured developmental gradients from progenitor to mature cell identities. In human cerebellar development and medulloblastoma, developmental topics transferred successfully across timepoints and sequencing technologies. Specific developmental programs showed enrichment in molecularly defined medulloblastoma subgroups, with developmental topic enrichment correlating with clinical outcomes. This established topic transfer as a viable approach for linking normal development to cancer biology. Extension to transcript-level resolution revealed that alternative isoforms organize according to developmental context rather than protein structure or function. Analysis of the NTRK receptor tyrosine kinase family identified three focus topics with distinct disease associations: an NTRK3-containing topic capturing sympathoadrenal differentiation signatures enriched in neuroblastoma, and two NTRK2-containing topics representing distinct renal developmental programs differentially enriched in Wilms tumor, clear cell sarcoma of the kidney, and rhabdoid tumor. Critically, full-length and truncated isoforms of the same gene showed distinct topic associations and disease enrichment patterns, demonstrating that isoform-level analysis reveals regulatory complexity invisible to gene-level approaches. Conclusions. Topic modeling provides a generalizable computational framework for identifying conserved transcriptional programs across developmental and disease contexts. The approach successfully scales from mouse to human systems, from single-cell to bulk sequencing, and from gene-level to isoform-level resolution. Key contributions include: (1) demonstration that topics capture transitional developmental states and hierarchical organization within lineages, (2) validation that developmental programs persist in pediatric cancers and correlate with clinical outcomes, (3) revelation that alternative splicing creates context-dependent expression programs organized by developmental lineage, and (4) establishment of a publicly available, technology-agnostic framework for cross-dataset comparison. The findings support developmental context as a fundamental determinant of both normal and pathological gene expression programs. Future directions include functional validation of identified programs, spatial transcriptomics integration, long-read sequencing validation of isoform predictions, and extension to additional pediatric cancer types and adult cancers with developmental origins. Significance. This work establishes topic modeling as a powerful analytical framework for understanding how developmental programs organize at multiple scales of resolution and how these programs contribute to disease. The transferable nature of the approach enables systematic comparison across the growing landscape of single-cell atlases without requiring computationally intensive integration methods. The framework is publicly available and broadly applicable to diverse biological contexts where identification of conserved expression programs is relevant, from comparative developmental biology to disease modeling and therapeutic target identification.