Traditional genomics faces a frustrating trade-off: you can either have a high-resolution "photograph" of tissue structure (histology) or a detailed "parts list" of gene activity (bulk sequencing), but rarely both at the same scale. Spatial Transcriptomics (ST) was designed to bridge this gap, yet most current platforms suffer from a "blurry pixel" problem.
In technologies like 10x Visium, gene expression is captured in discrete "spots" that are often 55 micrometers in diameter. As Emelie Berglund and Andrew Erickson observe in their study of prostate cancer, these spots act as "mini-bulk" averages, aggregating signals from 10 to 50 different cells. This lack of granularity obscures the very thing researchers want to see: the precise "conversations" happening between a tumor cell and an adjacent immune cell at the tissue's leading edge.
Recent research is converging on a shared goal: moving beyond these coarse spots to achieve true single-cell and even sub-cellular resolution through a combination of computational "unmixing" and high-density hardware.
The Computational Bridge: Unmixing the Spot
If a spatial transcriptomics spot is a blurry average of multiple cells, how do we recover the individual signals? The most prominent strategy involves using high-resolution histology (H&E) images as a guide.
While ST data is sparse and coarse, H&E images are dense and capture the physical boundaries of every cell. Frameworks like St2cell, developed by Kuan Tian, use deep learning to segment individual nuclei in the image and then apply convex quadratic programming to "unmix" the spot-level gene expression data. By assigning the right transcripts to the right cell boundaries, these models turn a blurry superpixel into a biologically accurate cellular map.
Similarly, Kalin Nonchev and Gunnar Ratsch introduced DeepSpot2Cell, which uses spot-level supervision to predict "virtual" single-cell profiles. This approach allows researchers to recover sharp spatial boundaries that are otherwise lost in the averaging effect of the capture technology.
Modeling Tissue as a Continuous Field
A second major shift is moving away from treating tissue as a grid of discrete points. In reality, biology is continuous. STINR, a framework by Yisi Luo and Xi-Le Zhao, applies Implicit Neural Representations (INR) to this problem.
Instead of storing gene expression in a table, STINR learns a continuous function that maps any 3D coordinate $(x, y, z)$ directly to a gene expression vector. This allows the model to "interpolate" between spots, effectively zooming in to infer expression at any arbitrary resolution. This continuity is essential for identifying "niches"—communities of interacting cells—which Sebastian Birk and Adib Miraki Feriz model using graph deep learning in NicheCompass.
Hardware Innovation and the Subcellular Frontier
While computational methods sharpen existing data, new hardware is pushing the physical limits of capture resolution. HiFi-Slides, developed by Tianyang Xu and colleagues, repurposes used DNA sequencing flow cells to create capture slides with sub-micron resolution (0.4 µm). This is fine enough to see not just individual cells, but the distribution of RNA within a single cell.
This move to the subcellular level requires a new analytical toolkit. Clarence K. Mah and the team behind Bento argue that treating a cell as a "homogeneous unit" is a mistake. RNA molecules are often localized to specific organelles or the cell membrane to perform their functions. By mapping the exact coordinates of individual molecules, researchers can finally see how the internal architecture of a cell changes in response to its environment.
Why Resolution Matters: The "Battlefront" of Disease
The push for single-cell resolution isn't just a technical exercise; it reveals biological structures that are otherwise invisible.
In colorectal cancer, Feiyu Tang used high-resolution mapping to identify a specific "battlefront" where stem-like tumor cells remodel the surrounding stroma to favor invasion. Without single-cell resolution, the signaling "conversations" between these tumor architects and their neighbors would be drowned out by the background noise of the surrounding tissue.
Similarly, in brain research, Abby Spangler and Kelsey D. Montgomery used high-resolution maps of the human prefrontal cortex to identify molecular sublayers that correlate with schizophrenia and autism—layers that the classic six-layer histological model simply cannot see.
By bridging the gap between the visual structure of tissue and the molecular identity of its cells, this research direction is turning "blurry" spatial data into a high-definition window into development and disease.