Spatial Clonal Architecture of Cancer

QC 2026-09-15

Abstract

Cancer is one of the most common causes of death worldwide and the number of new cases is projected to keep rising, therefore reliably detecting cancers and, in some cases more importantly, reliably predicting which will go on to cause harm is urgently needed. Prostate cancer is a perfect example of that. It is diagnosed far more often than it kills, and what decides the outcome is not whether cancer is present but whether any clone inside it can leave the gland. Treatment is currently decided from a morphological tumor grade, an average taken over a core that usually holds tumor of more than one grade alongside benign glands and stroma. This is a measurement done on a large unit that contains a lot of intra-unit heterogeneity. Spatial transcriptomics measures gene expression in many tiny regions across a tissue section without dissociating it, making the unit of annotation significantly smaller. Every measurement keeps its positional information, and from that expression, copy number can be inferred and cell state resolved.

Leveraging that spatial information, this thesis studies a tumor in two main areas: the cancer cells themselves, and the microenvironment around them. The cancer cells are studied through inferred copy number: we grouped spots into clones across whole sections, ordered those clones into trees, and found copy number events in epithelium that two pathologists had independently called benign, which we refer to as altered benign. Beyond the altered benign, clone boundaries frequently did not follow histological ones, and heterogeneity appeared within areas of uniform Gleason pattern.

Which tool is used for inference can drastically affect the result, and no systematic comparison existed. We benchmarked many copy number callers on simulated tumors with known ground truth and on experimental sections, and no single tool performed consistently on top. Recovering an accurate profile turned out to be a separate problem from grouping them in the correct way, and the benchmark therefore informs us which caller suits which task. The caller that recovered the most accurate profiles was applied to ten men with high-risk disease and nodal metastases. Profiled across the diagnostic biopsy, an axial section of the prostate and the involved lymph nodes, the resulting clone trees spanned all three and identified some of the clones that metastasized. No transcriptional signature consistently separated those clones from the rest across patients, with the differences that did appear being mostly patient-specific. Some of these men, staged as having no detectable spread and treated radically on that basis, nonetheless developed skeletal metastases.

Studying the tumor environment benefits a lot from being able to break down the spatial data into the cell types making up each region. This shows how those cell types are distributed in the tissue as well as helps with selecting the epithelial spots for copy number inference. However, the result depends heavily on the reference used. In our work we generated said reference in two ways. One is a general single-cell and spatial reference for hormone therapy-naïve localized prostate cancer, resolving epithelial states and fibroblast phenotypes among others and placing them in the tissue. The second generates a reference from the very same section the spatial experiment was carried out on, recovering single nuclei, the closest match available. The deconvolution results using this reference gave a better agreement with the underlying tissue compared to when using a public atlas.

In addition to our findings, the generated prostate cohorts will serve as valuable resources moving forward: the matched single-cell and spatial dataset for localized disease, and the diagnostic biopsy, axial section of the prostate and involved lymph nodes of the same ten men, graded spot by spot, for metastatic disease. Where work has come a long way is in deciphering cancer in two dimensions, in sections fixed as a snapshot in time. However, the third dimension is on the horizon and the temporal one hopefully further down the line.

Link to DiVA