PROJECT SUMMARY Mapping the spatial organization of cells and their communication in tissues is essential to understanding the process of development and disease formation. The rapid development of spatial transcriptomic technologies has enabled the profiling of the full transcriptome across thousands of locations in a tissue sample. In addition to transcriptional measurements, this technology also obtains paired histological imaging of the tissue. The spatially resolved profiling of gene expression has the potential to unlock groundbreaking discoveries, however, there are critical barriers in analyzing this data especially in complex tissue samples that involve the mixing of many diverse cell types in capture locations (spots). The low resolution makes it difficult to discern diverse cell types which is essential for downstream analysis of their spatial organization, dynamics, and interactions. Additionally, integrating spatial transcriptomic datasets across multiple tissue samples is not straightforward due to this technical limitation. Existing computational tools for analyzing high-dimensional genomic data were either built for single-cell resolution data or require paired single-cell transcriptomic data to guide the analysis of spatial transcriptomic data. Additionally, current methods do not consider spatial dependencies and information embedded in the histology image. The overarching goal of this proposal is to develop novel machine learning methods for analyzing the new wave of spatial transcriptomic data without the need for paired single-cell data as a reference. These innovative frameworks will enable characterizing diverse cell states and their spatial dynamics through a semi-supervised deconvolution of data (Aim 1) which will also allow integration of data from multiple tissue samples. We will also develop a multi-view framework for the integration of spatial transcriptomic and histological imaging for improved inference of intercellular interactions (Aim 2). By combining image processing algorithms for the alignment of images from replicate samples, we will extend this framework for integrating tissue samples. Our toolbox will be applicable to a broad range of tissue systems and larger clinical cohorts and has the potential to be transformative in understanding spatial dynamics during healthy development and guiding diagnosis and therapeutic strategies based on the spatial organization of cell types specific to the disease microenvironment.

PROJECT NARRATIVE This proposal seeks to address the strong demand for computational tools capable of analyzing emerging spatial transcriptomic datasets in particular in complex tissues. We aim to develop interpretable and semi-supervised machine learning models for characterizing diverse cell types and mapping their spatial organization and interactions from the integration of transcriptomic and imaging datasets. Our tools will be impactful in understanding the role of cell-cell crosstalk in a wide range of normal and disease tissue types.