Machine Learning Methods for Reconstructing Spatially Resolved Transcriptomes
Sparad:
| I publikationen: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Huvudupphov: | |
| Utgiven: |
ProQuest Dissertations & Theses
|
| Ämnen: | |
| Länkar: | Citation/Abstract Full Text - PDF |
| Taggar: |
Inga taggar, Lägg till första taggen!
|
| Abstrakt: | Dissecting the molecular, cellular, and spatial heterogeneity of tissues is critical to characterize how cells of different types function and communicate with each other underlying the spatial context, which eventually advances the understanding of phenotypic variations within tissues under various conditions. Recent advances in spatial transcriptomics technologies have enabled the profiling of spatially resolved maps of the whole transcriptome from intact tissues, revolutionizing transcriptomics studies and providing unique insights into biomedical research, particularly in developmental biology, neuroscience, immunology, and cancer biology. However, missing signals in spatial transcriptomics data due to technical limitations pose unprecedented challenges to their downstream analyses, and prohibitive costs associated with spatial transcriptomics technologies further hinder their practical implementation in large-scale clinical studies. To overcome these issues, my dissertation research seeks to develop computational methods to reconstruct spatially resolved transcriptomics based on either different modalities of spatial transcriptomics data or conventional transcriptomics data to provide high-quality and cost-effective alternatives. Chapter 3 presents our work on reconstructing spatially resolved transcriptomics from existing incomplete spatial transcriptomics data. We propose a graph-guided neural tensor decomposition method coupled with spatial and functional relations in spatial and PPI graphs (GNTD) to model nonlinearity in spatial transcriptomics. Comprehensive benchmarking indicates GNTD clearly outperforms baseline methods for spatially resolved transcriptomics imputation, and extensive comparisons between GNTD and baseline methods show our method excels in several important downstream analyses, such as spatial domain detection, spatially variable gene identification, spatial trajectories inference, and spatially co-expressed gene clustering. Chapter 4 focuses on reconstructing spatially resolved transcriptomics from tissue staining images. We introduce an adaptive spatial graph neural network (asGNN), which utilizes adaptive spatial graphs to better capture spatial relations embedded in tissue morphology through smoothing-based variational optimization, accurately predicting spatial gene expressions. Comparisons with state-of-the-art methods reveal that the predicted spatial expressions of marker genes from asGNN are highly correlated with their ground truth in the matched spatial transcriptomics data. Furthermore, the spatial graphs obtained from asGNN precisely delineate spatial domains in the tissues without the need for additional clustering analyses. Chapter 5 presents our recent work on reconstructing 3D spatially resolved transcriptomics from transcriptome tomography. We propose a collapsed tensor decomposition model (CTFacTomo) with reconstructed expression tensor regularized by spatial and functional relations while the collapsed tensor matching with the expression matrix along given spatial axes. Quantitative comparison on simulations from existing 3D spatial transcriptomics data and qualitative comparison on tomo-seq data demonstrate that CTFacTomo better restores the expression signals and characterizes 3D spatial patterns with focused regions. |
|---|---|
| ISBN: | 9798293866618 |
| Källa: | ProQuest Dissertations & Theses Global |