(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Cell2Spatial is a computational framework that maps single cells to spatial transcriptomic spots to reconstruct tissue architecture [1]

['Huamei Li', 'Department Of Rheumatology', 'Immunology', 'Nanjing Drum Tower Hospital', 'Affiliated Hospital Of Medical School', 'Nanjing University', 'Nanjing', 'Jingchao Liu', 'Tcrx', 'Keshihua']

Date: 2025-11 

Single-cell (SC) sequencing enables detailed characterization of transcriptional heterogeneity but lacks spatial context, while spatial transcriptomics (ST) preserves tissue organization yet is limited by resolution and incomplete gene capture. To bridge these gaps, we developed Cell2Spatial, a computational framework that segments spatial spots at single-cell resolution, even when SC and ST datasets are not fully matched in cell types. The method integrates information-theoretic gene selection, spatially weighted likelihood modeling, and spatial hotspot detection to improve signal fidelity. A corrected saturation model calibrates library size against gene complexity, ensuring accurate cell count estimation in low-resolution ST. To enhance scalability and spatial coherence, Cell2Spatial incorporates neural-network-guided clustering and a cost-minimizing assignment algorithm that balances transcriptional similarity with spatial proximity. Evaluations on synthetic data demonstrated that Cell2Spatial consistently outperforms existing tools in reconstructing tissue architectures and cellular compositions, with particular strength in handling unmatched datasets. Applications to 10× Visium data across mouse brain, human thymus, mouse kidney, and human dorsolateral prefrontal cortex revealed detailed anatomical structures and developmental trajectories. Moreover, for high-resolution platforms including Xenium In Situ, Visium HD, and Slide-seqV2, Cell2Spatial remained robust despite reduced transcript capture, effectively delineating fine-scale spatial patterns in complex tissues. Collectively, these results highlight Cell2Spatial as a versatile framework that expands the analytical scope of ST and provides a powerful tool for uncovering the spatial organization of cellular function and tissue architecture.

Funding: This work was supported by grants from Jiangsu Funding Program for Excellent Postdoctoral Talent (2022ZB699 to H.M.L.), the National Key Research and Development Program of China (2020YFA0710800 to L.Y.S.), the Key Program of the National Natural Science Foundation of China (82330055 to L.Y.S.), the National Natural Science Foundation of China (12132004 and 32471367 to Y.Y.L.), the Alzheimer’s Association and the National Alzheimer’s Coordinating Center New Investigator Award (NIAP24_1273876 to L.M.), the Alzheimer’s Association Research Grant—New to the Field (AARG-NTF-22-971669 to L.M.), and the NIH Award (P30AG072975 and K01AG084813 to L.M.). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data Availability: All spatial transcriptomics datasets and their corresponding single-cell data were obtained from publicly available repositories, with detailed sources provided in S1 Table. Cell2Spatial, an R package designed to reconstruct tissue architecture at single-cell resolution by assigning single cells to spatial transcriptomics spots, is available on GitHub ( https://github.com/lihuamei/Cell2Spatial ) and Zenodo (DOI: https://doi.org/10.5281/zenodo.17217934 ). The full list of R packages used is specified in the Materials and methods section. Reproducible analysis code and the source data underlying the figures are openly accessible via GitHub ( https://github.com/lihuamei/Cell2Spatial.Reproduce ) and Zenodo (DOI: https://doi.org/10.5281/zenodo.17212677 ). The structured spatial and corresponding single-cell data can be accessed on Zenodo (DOI: https://doi.org/10.5281/zenodo.17430116 ).

Herein, we present Cell2Spatial, a unified framework for reconstructing tissue architecture at SC resolution. Cell2Spatial integrates a spatially weighted maximum likelihood model with spatial hotspot detection to robustly quantify single-cell spot similarity and reduce mismatches between SC and ST datasets. For low-resolution platforms, it employs a corrected saturation model to estimate spot-level cell counts by modeling the relationship between library size and gene counts. To enable efficient large-scale mapping, a feedforward neural network (FNN) is combined with a linear sum assignment algorithm, ensuring accurate and scalable cell-to-spot allocation. Compared with existing approaches, Cell2Spatial offers a practical improvement in reconstructing tissue structure and cellular composition, seamlessly integrates unpaired SC–ST data, and reliably captures the spatiotemporal dynamics of cell states.

Despite these advances, current mapping tools face notable limitations. Label transfer-based strategies, such as Seurat’s, struggle under certain conditions. For high-resolution (SC or subcellular resolution; e.g., image-based platforms) ST data, they can annotate cell types per spot while preserving the original gene expression profiles, which benefits mechanistic analyses. However, in platforms such as Xenium, the low transcript capture rate leads to extremely sparse expression matrices, limiting transcriptome coverage and reducing analytical depth. For low-resolution (multiple cells per spot; e.g., spatial barcoding-based platforms) ST data, these tools typically assume each spot originates from a single cell type, overlooking potential cell mixtures within spots and thereby introducing bias. Beyond label transfer, SC mapping strategies also encounter two major challenges. First, aligning unpaired SC and ST datasets is difficult, particularly in minimizing mismatches and missing mappings. Second, in low-resolution ST data, spots frequently contain multiple cells, complicating estimates of cell counts and the inference of spatial relationships within heterogeneous neighborhoods. Different tools address these challenges with varying trade-offs. CellTrek [ 3 ] performs well in mapping unpaired SC and ST datasets, assigning cell types to corresponding ST regions, but it has limited ability to resolve cellular composition or fine-grained spatial relationships within spots. Conversely, CytoSPACE [ 20 ], by integrating the RCTD framework [ 12 ], provides more accurate estimates of cellular composition and spatial distribution, yet it assigns cells at the tissue-wide level and does not resolve the issue of unpaired SC and ST mapping. Together, these limitations underscore the need for new methods that can simultaneously handle unpaired SC–ST data integration and accurately infer spatial relationships within heterogeneous spots—an essential step toward advancing integrative SC and ST research.

To address this challenge, the widespread adoption of spatial deconvolution strategies has emerged. These strategies leverage SC reference data to deconvolute ST data and estimate the cellular composition within each spatial spot, as demonstrated by methods like Cell2location [ 10 ], CARD [ 11 ], RCTD [ 12 ], SpatialDWLS [ 13 ], STRIDE [ 14 ], DestVI [ 15 ], and Stereoscope [ 16 ]. However, these approaches are still constrained by resolution and typically provide information on the relative proportions of cell types, lacking detailed insights at the SC level, which hinders the in-depth exploration of cell states, interaction patterns, and neighboring cell populations. Establishing a channel that maps individual cells to specific spatial spots, enabling segmentation at SC granularity and potentially inferring unmeasured gene expressions, offers a broader perspective for the integrated analysis of ST and SC data [ 17 ]. Several methods have been publicly disclosed. One common approach, like Seurat [ 18 ], employs a label transfer method that maps SC labels onto spots. This method is based on Seurat’s anchor-based data integration framework, which identifies shared features between the reference SC dataset and the query ST dataset, thereby enabling the transfer of cell type labels from SC to ST spots. Tangram [ 19 ] employs a deep learning framework to reveal intricate spatial structures at the SC level, aiding in the detection of complex cell distributions and relationships. On the other hand, CellTrek [ 3 ] utilizes a random forest model trained on ST data to predict spatial coordinates, benefiting from shared dimension reduction features with SC data, enhancing its efficiency with large-scale ST datasets. CytoSPACE [ 20 ] employs a correlation-driven cost function based on convex optimization to allocate individual cells to precise spatial spots, ensuring effective and reliable cell positioning.

Single-cell (SC) sequencing has evolved into an indispensable tool for conducting comprehensive investigations into gene expression and functional attributes within individual cells, thereby significantly advancing our understanding of the intrinsic heterogeneity among cell types and their pivotal roles in developmental and disease processes [ 1 , 2 ]. However, the process of SC sequencing inherently sacrifices spatial information during sample processing, a critical component essential for gaining insights into the dynamic cellular microenvironment [ 3 ]. Spatial transcriptomics (ST) technologies, like 10× Visium, preserve cell locations within tissue sections [ 4 , 5 ]. Nonetheless, ST technologies often grapple with challenges related to spatial resolution and the provision of comprehensive SC gene expression data. The integration of ST with SC data represents a highly effective approach for expediting the analysis of cell distribution within tissues, revealing intercellular relationships, and gaining a profound understanding of the structural and functional intricacies of tissues [ 6 – 9 ]. Regrettably, this integration encounters complexities stemming from the high-dimensional nature of the data.

Results

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3003477

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/