Decrypting cancer's spatial code: from single cells to tissue niches

Abstract

Spatial transcriptomics (ST) has emerged as a powerful tool to map gene expression patterns to the local tissue structure in cancer, enabling unprecedented insights into cellular heterogeneity and tumour microenvironments. As the technology matures, developing new, spatially informed analytical frameworks will be essential to fully leverage its potential to elucidate the complex organisation and emerging properties of cancer tissues. Here, we highlight key challenges in cancer spatial transcriptomics, focusing on three emerging topics: (a) defining cell states, (b) delineating cellular niches and (c) integrating spatial data with other modalities that can pave the way towards clinical translation. We discuss multiple analytical approaches that are currently implemented or could be adapted in the future in order to tackle these challenges, including classical biostatistics methods as well as methods inherited from geospatial analytics or artificial intelligence. In the rapidly expanding landscape of ST, such methodologies lay the foundation for biological discoveries that conceptualise cancer as an evolving system of interconnected niches.

Keywords: AI; cancer; cell state; cellular niche; digital pathology; geospatial statistics; spatial transcriptomics.

Fig. 1

Schematic representation of a spatial transcriptomics workflow, illustrating key steps of downstream analysis. (A) Sequencing‐based: The workflow begins by mapping demultiplexed raw data to a reference genome, generating spatially barcoded gene expression matrices. Next, spots with low‐quality reads and other artefacts, such as high mitochondrial content or doublets, are filtered. An appropriate normalisation method is then applied based on the requirements of the downstream analysis. If working with multiple tissue sections, data integration is performed to harmonise batch effects and identify spots with similar gene expression profiles. Since each spot contains multiple cells, cellular deconvolution is conducted using a well‐annotated single‐cell RNA‐seq dataset as a reference to estimate the proportion of cell types in each spot. Imaging‐based: RNA molecules are captured using multiplexed hybridisation or sequential imaging. Detected transcript spots are localised and aggregated into a count matrix after cell segmentation. The matrix is then normalised (e.g. per‐cell scaling or z‐scoring) for downstream analysis. (B) Niche or domain identification follows, allowing for the interrogation of tumour ecosystems. (C) Additional downstream analyses, such as pathway enrichment, gene regulatory network reconstruction and ligand‐receptor interaction analysis, provide further insights into the tumour biology.

Fig. 2

Cell type/state and niche identification in spatial transcriptomics data. (A) Identifying cell types is typically independent of spatial context, whereas cell states are modulated by the surrounding niche through gradient‐based reprogramming of gene expression. (B) Niches are defined as spatial microenvironments composed of multiple interacting cell types whose coordinated behaviour leads to specific functions. Within the tumour microenvironment, cells can exploit their plastic capacity to adapt to different niches. The same cell type can be found in different niches (niche 1, 2 and 3), which may affect their cell state (i.e. initial, intermediate and terminal states). Arrows depict spatial trajectories of cell state transitions, which can also be interpreted as temporal transitions. (C) The plasticity of cells and supercellular nature of spots render traditional clustering algorithms (i.e. k‐means, Louvain, Leiden and cNMF) suboptimal in identifying cell types and states. NMF, Non‐negative Matrix Factorisation. (D) Spatially informed approaches, such as graph‐based methods and ecological models, can capture niche structure and reveal how local microenvironments shape cell behaviour.

Fig. 3

Integrating digital pathology with matched gene expression profiles. A summary of methods to integrate spatial transcriptomics and digital pathology, and their applications. cGAN, conditional Generative Adversarial Network; H&E, haematoxylin and eosin; MoE, mixture of experts; ST, spatial transcriptomics; WSI, whole slide images.