Computational methods for learning, evaluating, and interpreting spatial organization in transcriptomic data

Abstract

Spatial transcriptomics (ST) enables the study of gene expression within intact tissue architecture, creating new opportunities to investigate cellular organization, tissue heterogeneity, and intercellular communication. This dissertation develops and evaluates computational methods to address key challenges in ST analysis, including spatial domain identification, benchmarking, data integration, cell-cell communication inference, and mechanistic interpretability of single-cell foundation models. It first presents a comprehensive review of ST methodologies and computational frameworks to contextualize current advances and limitations in the field. It then introduces ADEPT, a method for spatial domain identification that improves clustering accuracy through the use of differentially expressed genes and imputation. A systematic benchmarking study evaluates existing tools for clustering, integration, and alignment in ST data. Building on this, the dissertation presents MaskGraphene, a framework for integrating multi-slice and multi-condition ST datasets to improve cross-sample comparability and downstream analysis. The dissertation also combines a benchmark of cell-cell communication inference methods with the construction of a mouse spinal cord spatial atlas, using this application to assess method performance in a biologically relevant setting and to characterize spatially organized intercellular signaling. In addition, it investigates the interpretability of current ST foundation models, focusing on multi-level analyses of model representations and the biological meaning captured by them. Overall, this dissertation contributes new methods, evaluation frameworks, and interpretability strategies for integrative spatial transcriptomics analysis.