Cell type deconvolution of prostate cancer spatial transcriptomics

Abstract

Prostate cancer is one of the most common cancer types worldwide, and its heterogeneity poses challenges to the diagnosis and treatment of the disease. Single-cell RNA sequencing has shed light on the different cell subpopulations within tumor tissue, but the method lacks spatial information of these cells. Recent advances in spatial transcriptomics have made it possible to study transcriptome-wide gene expression in a spatial context. This is important because cells interact actively with their surroundings, and the spatial location of cells has a huge impact on their function. Although the throughput of spatial methods has increased, the resolution of methods that yield enough reads to be able to confidently conduct gene expression analysis is still not on a single-cell level. This means that a certain transcript cannot be traced back to its originating cell although its location on the tissue can be determined in a resolution of tens of micrometers. To tackle this problem, spatial transcriptomics data can be deconvolved by using reference cell type signatures obtained from single-cell RNA sequencing data. The purpose of the deconvolution analysis is to find out what type of cells produced the transcripts that were captured in a certain location. Several algorithms have been developed for this purpose. In this thesis, I used four such algorithms to investigate the cell composition of four prostate cancer samples from three different patients. The used algorithms were Seurat, RCTD, stereoscope, and cell2location. The aim was to compare the results of the different analysis tools and evaluate the spatial organization of cells in our prostate cancer samples. There were a lot of differences between the analysis tools. In addition to differences in usage, like the freedom of choosing different parameters for the models, the results turned out to be quite different as well. Especially the results from Seurat integration analysis devi-ated from the results of the other algorithms. The mast cell proportion estimates were especially divergent between the tools, most likely due to problems in defining the reference cell type gene signatures from the single-cell data. RCTD, stereoscope, and cell2location gave quite similar predictions of cell type propor-tions and the results seemed in line with tissue morphology. For future deconvolution analysis, I would use cell2location due to its flexibility and transparency. The results of this thesis show that deconvolution algorithms offer a good way of studying the spatial organization of cells in tissue.