For years, genomic research faced a fundamental trade-off between depth and geography. Single-cell RNA sequencing (scRNA-seq) provides an unparalleled look at a cell's internal state but requires dissociating the tissue, destroying the spatial architecture. Conversely, early spatial transcriptomics (ST) methods preserved tissue structure by capturing mRNA on barcoded grids, but these grids were often too coarse, capturing a mixture of multiple cells within a single "spot."
To resolve this, modern computational biology treats spatial mapping not as a static imaging problem, but as a multi-modal alignment problem. By combining the deep cellular dictionaries of single-cell sequencing with the physical coordinates of spatial arrays, researchers can deconvolve mixed signals, map cellular niches, and reconstruct tissue architectures at single-cell resolution.
The Mechanics of Spot Deconvolution
The core challenge of spot-based spatial transcriptomics is that a single observation spot is a composite signal. If a spot contains a mixture of macrophages, T cells, and epithelial cells, its measured gene expression vector is the sum of the transcripts contributed by each cell type.
Deconvolution works because gene expression profiles are highly structured and redundant. Instead of treating every gene independently, computational frameworks model the observed expression vector of a spatial spot, $Y$, as a probabilistic mixture or linear combination of reference profiles derived from scRNA-seq data:
$$Y \approx \sum_{k} w_k X_k$$
where $X_k$ represents the characteristic expression signature of cell type $k$, and $w_k$ represents the proportion (or weight) of that cell type within the spot.
As demonstrated by Joseph Bergenstråhle et al. in their framework stereoscope, this can be framed as a probabilistic inference problem (Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography). By modeling single-cell data using a negative binomial distribution, the framework estimates the parameters of each cell type's expression profile. It then uses these parameters to find the maximum likelihood composition of cell types for each spatial coordinate. Because the reference profiles are highly distinct, the optimization can accurately untangle the mixed signals, mapping cell type topography even when the underlying spatial technology lacks single-cell resolution.
Bridging Modalities with Multimodal Alignment
While transcriptomic data is highly informative, it is often noisy and subject to technical dropouts. Tissues, however, possess inherent structural constraints: cells do not arrange themselves randomly, and cellular morphology strongly correlates with functional states. To exploit this, recent frameworks integrate high-resolution histology images (such as standard H&E staining) directly into the deconvolution pipeline.
This multi-modal integration relies on the assumption that visual features—like cell density, nuclear shape, and tissue texture—provide a structural prior that can smooth and correct transcriptomic predictions. For instance, Siyu He et al. developed Starfysh, a deep generative framework that utilizes a product-of-experts architecture to combine transcriptomic data with histological image features (Starfysh integrates spatial transcriptomic and histologic data to reveal heterogeneous tumor–immune hubs).
By processing the histology image through a convolutional neural network, the model extracts spatial density and architectural priors. If the transcriptomic data for a spot is ambiguous, the histological expert guides the model based on visual context (e.g., recognizing the characteristic density of an immune hub). As noted in a technical review by Roxana Zahedi Nasab et al., leveraging this inherent correlation between morphology and molecular profiles allows deep learning models to outperform traditional statistical methods, effectively denoising spatial expression maps (Deep Learning in Spatially Resolved Transcriptomics: A Comprehensive Technical View).
Reconstructing Space via Optimal Transport
What happens if you have a rich single-cell dataset but no paired spatial transcriptomic grid? Can you map dissociated cells back into an abstract or physical space using gene expression alone?
The mathematical answer lies in optimal transport theory. Physical proximity in a tissue correlates with gene expression similarity; cells residing in the same microenvironment share signaling pathways and metabolic constraints. Wenbin Zhou et al. demonstrated this principle by implementing distance-preserving generative models that map high-dimensional genetic data directly to spatial coordinates based on these expression gradients (Distance-Preserving Generative Modeling of Spatial Transcriptomics).
Taking this further, Qing Nie developed SpaOTsc, a framework that uses structured optimal transport to map dissociated single-cell data onto a spatial reference geometry (Inferring spatial and signaling relationships between cells from single cell transcriptomic data). The algorithm treats the single-cell dataset and the spatial coordinates as two probability distributions. It then calculates a "transport plan" that minimizes the cost of moving cells into spatial positions, where the cost is determined by the mismatch between a cell's expression profile and the spatial landmarks.
[Dissociated scRNA-seq] \
--> [Optimal Transport / Probabilistic Alignment] --> [Spatially Resolved Single-Cell Map]
[Spatial Reference/Grid] /
This mapping allows researchers to classify genes based on their shared spatial activity patterns rather than traditional functional annotations, uncovering localized cell-to-cell signaling networks that would be completely invisible in dissociated data.
From Coordinates to Niches
The ultimate goal of spatial transcriptomics is to move beyond individual coordinates and understand "cellular neighborhoods" or niches. By aligning multiple modalities—such as transcriptomics, proteomics, and pathology—into a single physical coordinate system, researchers can observe how cells cooperate or clash in situ.
For example, Alexandro E. Trevino et al. utilized graph deep learning to align spatial transcriptomics and multiplexed imaging into a unified coordinate system (1415 Multimodal spatial signatures of immunotherapy response in hepatocellular carcinoma). By treating cells as nodes in a spatial graph and physical distances as edges, graph neural networks can learn the molecular signatures of specific tissue interfaces. This approach reveals how localized environments—such as the boundary between a tumor core and its inflammatory periphery—dictate disease progression and treatment resistance, proving that a cell's behavior is fundamentally defined by where it lives.