Untangling heterogeneity: From complexity to prediction

Prostate cancer is a high-incidence male cancer that progresses to metastatic disease in a subset of diagnosed cases. Early detection of high-risk disease is a key translational challenge and if successful will reduce overtreatment while potentially improving the outcomes for those patients most in need of early interventions. There are many challenges that need to be surmounted to achieve this translational outcome, not least the fact that prostate cancer is a multi-focal disease. Multi-focality is observed in the often-diverse histopathology of the disease in a single patient, as well as in the mutational and transcriptional heterogeneity in sequencing data from selected multi-site cores.¹ In this study, Salachan et al. use spatial transcriptomics on samples from patients with castrate-resistant (CRPC) and neuroendocrine prostate cancer (NEPC), poor-prognosis treatment-resistant prostate cancers.² By deriving a signature associated with an immune-dysregulated malignant niches or ‘ecosystem’, they are able to prognosticate larger publicly available cohorts for which bulk sequencing data are available.

The chosen platform for this study was Visium (10x Genomics) and the work was undertaken on fresh-frozen tissue samples. The technology as applied here provides a spot-level resolution of 10 to 15 cells and coverage of a fraction of the cellular transcriptome per spot. The cellular and transcriptomic resolution of spatial platforms is improving all the time and this study represents an early window into these niches. Undoubtedly, our ability to discriminate between the contributions of distinct cells and cell types to spatial transcriptomic profiles will improve. New computational methods to de-convolute these data according to cell type and spatial location are also being rapidly developed. In this study, Spatial Cellular Estimator for Tumours (SpaCET) was used to infer cellular identities within the ST data.³ As well as determining local cell densities and the fractions of cells of different types within the spot-level data, this method also uses ligand-receptor co-expression analysis to infer intercellular interactions between different cell types.³ This represents one approach amongst many being developed to make spatial data biologically interpretable, and particularly in deciphering cross-talk/cellular co-dependencies that may permit niches to emerge and survive within tissue samples. In deriving signatures, the authors have focused on histo-pathologically malignant regions of tissue within the three samples selected. SpaCET distinguishes between malignant cells and other cell types by comparing spatial transcriptomic data initially to curated cancer-type specific dictionary gene patterns reflecting copy-number alteration or malignant gene expression signatures. The intent is to be able to identify malignant cell populations that have both high and low levels of copy-number instability. This dictionary has been derived from the Cancer Genome Atlas (TCGA)⁴ datasets. Public single-cell RNA-seq datasets are then used as reference datasets from which to identify other non-malignant cell types with immune and other lineage characteristics. This approach has been most comprehensively described in a recent paper by Ru et al.,³ and the approach taken here follows this very closely through to assessing ligand-receptor interactions and the relationship between cancer-associated fibroblasts (CAFs) and polarised M2 macrophages. Applying this approach, the authors are able to begin to interpret the tumour immune microenvironment. In one CRPC case (CRPC1), they identified cells with luminal-epithelial characteristics expressing prostate-specific antigen (kallikrein 3:KLK3). They were also able to identify a regulatory T-cell cluster surrounded by these luminal epithelial cells and located within a stromal region expressing high levels of immune-suppressive Galectin-1. High Galectin-1 expression along within a stromal region was also a feature of the second CRPC (CRPC2) case. In addition, there was high expression of an osteonectin marker gene previously reported to correlate with increased risk of biochemical relapse in TCGA data. Pathway enrichment analysis revealed enrichment in the CAF cluster of genes associated with focal adhesions and extracellular matrix-receptor interactions. Spatially they found that the CAF cluster isolated the luminal epithelial KLK3 cluster from the immune-related cluster. Collectively their analysis suggests that CAFs contribute to an immune-suppressive tumour microenvironment shielding cancer cells in CRPC.

In the NEPC case, they were also able to identify KLK3-positive cells. The sample-lacked expression of canonical immune checkpoint regulators but the non-canonical immune checkpoint regulator CD276 was widely expressed. Two distinct CAF clusters were identified within the NEPC data out of a total of 12 clusters. Genes associated with immune activation were also highly expressed in one of the two CAF clusters. Here too, the authors were able to conclude that an immune modulatory CAF population was a feature of this NEPC and might perhaps contribute to metastatic progression. To extend this further they applied SpaCET revealing a strong positive correlation in the spatial coexpression of genes associated with CAFs and M2 macrophages surrounding KLK3-positive cell cluster. SpaCET clustered the malignant cells in the NEPC case into three subtypes in total. A low expressing KLK3 subtype clustered distinct from the two others (subtypes 2 and 3). Differential gene expression analysis of these two spatially proximal malignant cell subtypes revealed that enrichment of pathways associated with IL-17 signalling, Th17 cell signalling and other immune-related biologies were hallmark discriminators of these subtypes with decreased enrichment for several immune cell types predicted for subtype 3. Projecting impaired immune regulation as a feature of this predominant malignant subtype, they then derived an immune signature consisting of predominantly downregulated genes in subtype 3 vs 2. This signature was then applied to two publicly available bulk datasets, TCGA⁴ and the metastatic Stand-Up 2 Cancer (SU2C) dataset.⁵ In the TCGA dataset, this signature was associated with a greater probability of disease recurrence and in the SU2C dataset with greater genomic instability.

This study indicates that immune-dysregulated niches sustained by cancer-associated fibroblasts may support disease progression. The cost of spatial genomics and the scale of the data currently preclude large-scale cohort studies to determine how widespread these features might be in patients and whether they are reflected in metastatic samples or even emerge in histo-pathologically benign tissue/in an early-onset and precancerous state. The latter point is important in light of a recent paper that identified copy-number instability in areas of prostate tissue that are histo-pathologically benign.⁶ The costs of spatial genomics may fall for example through the development of novel bar-coding strategies, as exemplified by a recent study that used such a method to create a three-dimensional view of gene expression by profiling consecutive organoid sections.⁷ The field will also need to expand analysis beyond areas of defined as cancerous by histopathology. For example, recent work identified copy-number instability in areas of prostate tissue that are histo-pathologically benign.⁶ In the future, integrating spatial proteomics and metabolomics with genomic data may lead to a greater understanding of functional cross-talk if equivalent resolutions and same-sample data generation are feasible. Since this study was undertaken higher resolution spatial genomic methods now permit signature testing at true single-cell resolution. Significant progress is to be expected in spatial de-convolution, niche characterisation and our understanding of clonal selection in prostate and other cancer types.