Nuclear-Free Zoning

Spatial proteomics and transcriptomics are mainstream technologies that molecularly characterize individual cells or groups of cells at spatial locations throughout the tissue. As a result, these methods produce new understandings of organ and organism development and of disease progression, including elucidating the role of immune cells in carcinogenesis. The steps in the execution of such imaging-based technologies are to cut a thin tissue section (≈ 5 μm thickness), uniquely label the specific protein or RNA molecules of interest, acquire images of the labeled section, and analyze the images. In most cases, the labels are fluorescent, and some methods cyclically iterate between labeling and acquisition to build up a profile of scores of proteins or RNA transcripts across the tissue.

State of the art acquisition methods produce images of sufficient spatial resolution to facilitate localization of the labeled species in the individual cells. Consequently, the canonical image analysis methods first detect each cell by segmenting its counter-stained nucleus followed by quantifying each labeled species in the nucleus or in the surrounding cytoplasm of the nucleus. Such methods work well for cells that have small, ring-shaped cytoplasm surrounding their nuclei (e.g., T cells) and cells that adhere to each other into a cobble stone arrangement (e.g., epithelial cells). However, some cell types take on elongated morphology with their cytoplasm extending tens of microns from their nucleus (e.g., neurons, myocytes, and fibroblasts), and for these types, the nucleus is not representative of the overall cell extent and shape, leading to failed estimation of the cytoplasmic zone for these cells. Directly finding the borders of such cells by explicitly labeling the plasma membranes has shown promise, but universal plasma membrane markers have proven elusive. Moreover, some proteins of interest are inherently extracellular, such as matrix-metalloproteinases that can play a key role in tumor cell invasion.

Recently, several studies reported cell detection methods that circumvent nucleus segmentation and instead rely on certain molecular markers, or combinations thereof, being present in the cytoplasm of one cell with different levels of the markers in neighboring cells. In one early study [1], cell type signatures were calculated by clustering from combinations of gene expression markers in osmFISH and MERFISH images. Examples of subsequent works were an approach that optimizes cell boundary locations by considering the joint likelihood of transcriptional expression along with cell morphology [2]. In 2023, Liu et al. [3] used unsupervised clustering of pixel-level features for capturing relevant objects, such as the extracellular matrix, outside of the cell, and in addition, clustering of pixels that lie within the cells improved cell segmentation over standard methods.

Methods to date to investigate tissue architecture without explicit cell segmentation have been limited in the spatial scale of detected zones and may execute slowly, although the latter depends on available hardware. Building on their previous work investigating this task, Wählby et al. [4, 5] describe the computational tool, “Points2Regions,” that identifies image zones of similar RNA composition existing across a diverse range of spatial scales. The article starts with a brief and excellent review of standard practices for image analysis including methods for cell- and nucleus-free segmentation. The methodology applied in this article [5], combining hierarchical and k-means clustering of multiscale counts for each RNA class, proved to be not only accurate but also computationally efficient. Importantly, it compared favorably with the results obtained using the methods requiring explicit cell segmentation across a range of simulated and real spatial transcriptomics datasets. Their software is open source and uniquely the authors provide an interactive web-based version. Figure 1A is an example of application of the method applied to a colorectal cancer liver metastasis tissue sample [6] labeled using a direct RNA-targeted in situ sequencing method [8] (Figure 1B). Points2Regions automatically identified 10 zones, revealing a progression in zone type from healthy liver on the left to tumor on the right (Figure 1C).

The article by Andersson et al. [5] is clearly a significant advance in the analysis of multiomics spatial images and furthermore is readily usable by a wide range of researchers in the spatial-omics fields. We look forward to the further developments from the Wählby Team and others. There would be promise in merging cell segmentation-free methods with those that start from nuclei segmentation, using the latter for those markers expected to be only in or proximal to the nuclei and the former for markers that do not localize to nuclei. Multiplexed consensus cell segmentation is an example of this approach where non-nucleated fibroblasts were segmented based on positive αSMA staining and merged with segmentation masks of nucleated cells [9]. Prospecting for marker combinations that consistently identify cell boundaries [10, 11], even if application dependent, would be worthwhile to continue, however. Looking forward, overcoming the major limitation in spatial-omics methodologies of using thin sections (≈ 5 μm thickness) is ripe for progress. In such sections only a fraction of each cell exists, thus limiting quantitative accuracy of marker and cell morphology measurements, as well as losing the full 3D context of each cell. Others, however, are progressing in this area [12-15] and in one example are reporting significantly new interpretation of tertiary lymph node structures [15]. Furthermore, high-throughput volume electron microscopy, which although currently lacking molecularly specific markers, does afford two orders of magnitude in linear spatial resolution, and combined with artificial intelligence analysis of subcellular features [16] will provide a powerful and orthogonal information stream to existing methods.

Stephen Lockett: writing – original draft, writing – review and editing. Andrew Weisman: writing – review and editing.

The authors declare no conflicts of interest.