This thesis presents a novel computational framework for the analysis of single-cell and spatial transcriptomics data, driven by the challenge of characterising the heterogeneous tumour microenvironment of Glioblastoma (GBM). Existing clustering methods often struggle with the scale, sparsity, and spatial dependency inherent in modern transcriptomic datasets. To address these limitations, we propose a scalable Bayesian non-parametric framework. We first introduce the Dirichlet Process Negative Binomial Variational Inference (DPNBVI) model, which couples a generative Negative Binomial likelihood with a Dirichlet Process prior. Unlike standard clustering algorithms, this generative approach models the full posterior distribution of expression parameters, enabling direct gene-level interpretation of identified cellular states.

To resolve tissue architecture, we extend this framework to a Spatially-Aware model (SDDPVI) and subsequently a hierarchical Nested Dirichlet Process with Dual Spatial Regularisation (NDPNBVI_{Dual-MRF+Marker}). This modular architecture demonstrates that the model is easily extensible with sophisticated structures, allowing for the seamless integration of biological priors (such as marker guidance) and complex spatial constraints without compromising computational efficiency. To ensure scalability to atlas-level datasets (>600,000 cells), we implement Stochastic Variational Inference accelerated by Just-In-Time hybrid compilation.

We apply this framework to a novel GBM spatial transcriptomics cohort to investigate the synergistic effects of Arginine Deprivation therapy (ADI-PEG20) and Radiotherapy. The model uncovers a hierarchical regulatory architecture and identifies a critical mechanism of treatment response: while radiation monotherapy depletes the myeloid compartment, combination therapy rescues antigen-presenting myeloid cells and enhances CD8+ T-cell infiltration. Furthermore, we infer cell-type-specific Gene Regulatory Networks, identifying TFEB, HIF1A, RELA and E2F3 as key molecular drivers of these therapeutic shifts. This work bridges the gap between statistical modelling and mechanistic biology, providing a robust engine for hypothesis generation in cancer research.