Transferable Transcriptional Topic Modeling Traces Medulloblastoma Subtypes to Distinct Cerebellar Developmental States

Single-cell transcriptomics transformed our understanding of cellular heterogeneity, yet cross-dataset comparison remains fundamentally limited by batch effects, technology differences, and inconsistent annotation frameworks. These challenges have impeded efforts to connect developmental programs with disease states which creates a critical gap for understanding pediatric cancers with suspected developmental origins. Here, we demonstrate that topic modeling, an unsupervised natural language processing technique, overcomes these barriers by learning transferable transcriptional genetic signatures that generalize across datasets, technologies, and biological contexts without requiring data integration. Applying this framework to over one million fetal cerebellar nuclei, we identify seven distinct topics that capture the developmental spectrum of rhombic lip progenitors through external granule layer (EGL) differentiation, including transitional states missed by conventional clustering. These topics transfer successfully across sequencing technologies (SPLiT-seq to sci-RNA-seq3), developmental timepoints, species (human to mouse), and from single-cell to bulk RNA sequencing of 876 medulloblastoma tumors. We reveal that Sonic hedgehog (SHH) medulloblastoma subtypes retain these topics, corresponding to distinct stages of EGL development. Our findings highlight that developmental timing at tumor initiation fundamentally shapes tumor biology, with immediate implications for subtype-specific therapeutic strategies in pediatric brain cancer. Our workflow establishes topic modeling as a scalable solution for mining expanding genomic atlases with broad applications across different datasets, technologies, and biological contexts.