CSB PhD Thesis Defense: Kamal Maher (Xiao Wang Lab)

Date:

On April 25, 2025 at 10:00 am till 11:00 am

Event Description:

Date: Friday, April 25th, 2025

Time: 10-11 am

Room: 68-181

CSB PhD Candidate: Kamal Maher

Advisor: Xiao Wang (Chemistry, Broad)

TDC Members: Peter Reddien (chair) Alex Shalek, Fabian Theis (external)

Title: Fundamental representations of regions and interactions in spatial transcriptomics

Abstract:

While cells are often considered the fundamental unit of biology, it is their spatial coordination that gives rise to the tissue architectures underlying both health and disease. Spatial transcriptomics technologies offer a unique window into this coordination by simultaneously capturing the spatial and molecular identities of individual cells, providing unprecedented insight into tissue organization. However, the computational landscape for analyzing tissue structure remains fragmented, with a wide array of disparate methods. In this work, we aim to distill these approaches into a unified quantitative framework for analyzing tissue architecture.

Tissue structure can be represented in terms of anatomical regions as well as the cell-cell interactions that occur within them. For regional tissue organization, many existing methods—including those based on probabilistic models and graph neural networks—ultimately perform a form of smoothing, or local averaging of gene expression across neighboring cells. This process emphasizes large-scale spatial variation and enables standard single-cell analysis workflows, such as clustering and trajectory inference, to be applied in spatial contexts. However, we find that naive smoothing introduces artifacts that obscure meaningful spatial features. To address this, we introduce a minimal but powerful modification: subsampling within each neighborhood prior to averaging. This approach generalizes conventional analyses to spatial features: clustering identifies multicellular regions; data integration aligns spatial regions across samples and technologies; and trajectory inference captures spatial gradients. We also show that this subsampling strategy improves the performance of more complex downstream methods.

To further generalize our framework, we formalize the joint analysis of tissue regions and multiscale cell-cell interactions using signal processing over graphs: low-frequency components represent regional gene expression patterns across a tissue mesh; high-frequency components capture fine-scale, cell-cell interactions; and mid-frequency signals correspond to boundaries between regions and diffusive signaling. By interpreting spatial gene expression in this spectral framework, we provide a principled way to bridge conceptual and computational perspectives on tissue structure. Ultimately, this work serves as both a theoretical foundation to understand existing methods and a roadmap for developing future approaches to quantitatively describe molecular tissue architecture.