Biological tissues are not just bags of cells; they are highly organized architectures where a cell’s function is often defined by its neighbors. For years, researchers faced a trade-off: they could either use single-cell RNA sequencing (scRNA-seq) to get a high-resolution "parts list" of every cell type in a sample (but lose their physical location) or use spatial transcriptomics (ST) to see where genes are expressed (but often at a resolution too coarse to see individual cells).
The emerging research direction across these papers is the computational integration of these two modalities. By using scRNA-seq as a high-resolution "reference dictionary" and ST as a physical "scaffold," researchers can now project molecular identities back into their original anatomical context.
The Resolution-Context Trade-off
To understand why this integration is necessary, consider the limitations of current hardware. Technologies like Visium often capture "spots" that contain 10–50 cells. Without help, it is impossible to tell if a signal comes from one hyperactive cell or many quiet ones. Conversely, scRNA-seq requires dissociating tissue into a "soup," which destroys the very spatial niches—like the invasive front of a tumor or the layers of the brain—that drive disease.
Researchers like Joseph Bergenstråhle and Joakim Lundeberg addressed this in Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography by developing stereoscope. This probabilistic framework models gene expression in spatial spots as a mixture of profiles derived from single-cell data, allowing researchers to "deconvolve" a single spot into its constituent cell types.
Aligning Modalities with Deep Learning
Beyond simple deconvolution, the field has moved toward "alignment"—treating the problem as a translation task between two different data spaces. Gabriele Scalia, Tommaso Biancalani, and colleagues introduced Tangram in Deep learning and alignment of spatially-resolved whole transcriptomes of single cells in the mouse brain with Tangram.
Tangram works by looking at the genes both methods measured and finding the best "fit" to map single-cell profiles onto spatial coordinates. This allows for the spatial projection of features that were never actually measured in the spatial experiment, such as chromatin accessibility or rare protein markers. This "Google Maps" for the body allows researchers to zoom from the organ level down to the specific genetic programming of a single neuron.
Identifying Cellular Niches and Hubs
The power of this integration is most evident in clinical discovery. By mapping high-resolution profiles onto physical architecture, researchers can identify "cellular hubs"—localized neighborhoods where specific cell types congregate to drive pathology.
For example, Jia Li and Alan J. Simmons used this dual-track approach in Identification and multimodal characterization of a specialized epithelial cell type associated with Crohn’s disease to discover LND cells. These are rare epithelial cells that expand during inflammation and act as communication hubs to recruit immune cells. Without the spatial context, these cells might have been dismissed as a minor subpopulation; with it, they were revealed as central players in the disease's physical architecture.
Similarly, in cancer research, Alexandro E. Trevino and Zhenqin Wu identified an "interface niche" in liver cancer—a physical barrier of cells that isolates immune clusters from the tumor, explaining why some patients resist immunotherapy (478 Spatial omics and deep learning define an immunotherapy resistance niche in hepatocellular carcinoma).
Scaling to 3D and Whole Organisms
The frontier of this research is moving from 2D slices to 3D reconstructions of entire organs. Frameworks like STitch3D, developed by Gefei Wang and Can Yang, use deep learning to "stitch" multiple 2D spatial slices together while simultaneously deconvolving them with scRNA-seq references (Construction of a 3D whole organism spatial atlas by joint modeling of multiple slices).
This scaling is supported by massive data efforts like Stereo-seq, which Ao Chen and Yidi Sun used to create the first single-cell resolution spatial map of the macaque brain (Global Spatial Transcriptome of Macaque Brain at Single-Cell Resolution). To manage this complexity, the community is gravitating toward unified frameworks like SpatialData, which provides standardized structures for aligning multi-modal, multi-scale data (SpatialData: an open and universal data framework for spatial omics).
Summary of the Integrated Approach
The shift in the field is away from viewing spatial and single-cell data as competing technologies and toward viewing them as a single, integrated computational problem. By combining the transcriptomic depth of dissociated cells with the physical coordinates of tissue slices, researchers can now:
- Deconvolve multi-cell spots into precise cell-type proportions.
- Predict the location of unmeasured genes or epigenetic states.
- Identify localized cellular "niches" that drive disease progression and treatment resistance.