In traditional computational biology, a tissue sample is often treated as a "bag of cells." Whether through bulk sequencing or standard single-cell RNA-seq (scRNA-seq), the analytical goal has historically been to identify what types of cells are present and in what quantities. However, this approach discards a fundamental biological truth: cells do not exist in isolation. Their identity, function, and disease state are inextricably linked to their physical location and the "social circle" of neighbors surrounding them.
A growing body of research is shifting the paradigm from independent cell analysis to spatially aware representation learning. By treating tissues as structured networks rather than unordered sets, researchers are unlocking the ability to predict gene expression from simple images, discover novel cell states, and better understand disease progression.
The Limits of the Independence Assumption
Most early deep learning models for single-cell analysis, such as scFormer by Haotian Cui and colleagues, focused on the internal gene correlations within a single cell. While powerful, these models suffer from a "spatial blindness." As noted in Deep Learning in Single-Cell Analysis, treating cells as independent observations ignores the complex communication networks and mechanical microenvironments that define biological systems.
When we ignore spatial context, we lose the ability to see how a tumor cell interacts with the surrounding stroma or how immune cells are recruited to specific inflammatory sites. As Vaishnavi Subramanian demonstrates in Integration of Spatial Distribution in Imaging-Genetics, the where is often just as important as the what for determining clinical outcomes.
Modeling Tissues as Graphs and Sentences
To capture these relationships, researchers are borrowing architectures from natural language processing (NLP) and computer vision, reimagined for biological geometry.
- Cells as Tokens: Hongzhi Wen and Wenzhuo Tang introduced CellPLM, which treats individual cells as "tokens" and entire tissue sections as "sentences." By using a transformer architecture, the model learns the "grammar" of tissue organization, allowing it to capture intercellular dependencies that gene-centric models miss.
- Cells as Nodes: Graph Neural Networks (GNNs) have become the standard for modeling physical proximity. In STELLAR, Maria Brbic and Garry P. Nolan use geometric deep learning to annotate cells based on their local neighborhood. Similarly, BANKSY, developed by Li Lin, augments a cell's genetic profile with its "transcriptomic neighborhood," significantly improving the accuracy of cell typing in complex tissues.
Bridging the Gap Between Morphology and Molecules
One of the most impactful applications of this research is "filling in the blanks" of spatial data. High-resolution spatial transcriptomics is expensive and technically challenging. However, standard H&E-stained histology images are cheap and ubiquitous.
Models like ResSAT and ST-GCHB use self-attention and graph contrastive learning to predict spatially resolved gene expression directly from these images. By learning the relationship between tissue morphology (how the tissue looks) and molecular function (which genes are active), these frameworks—such as VISTA by Tianyu Liu and Hongyu Zhao—can impute missing data and provide a more coherent map of biological boundaries.
From Neighborhoods to Clinical Insights
The shift toward spatial awareness isn't just a technical improvement; it changes the biological questions we can answer.
- Disease States: In studies of COVID-19, Arijit Sehanobish and David van Dijk used GNNs to identify how specific cell-to-cell interactions drive severe infection 055d3176.
- Immunotherapy: The COMPASS model by Wan Xiang Shen and Michelle M Li maps gene expression onto immune concepts, providing personalized response maps that account for the tumor microenvironment's architecture.
- Chronic Inflammation: Research into Crohn’s Disease by Si-Chen Zhu uses "spatial signatures" to show how local cellular arrangements shift during inflammation, moving beyond simple cell counts to a more nuanced view of pathology 553e8beb.
By integrating spatial dependencies as a first-class citizen in model architecture—whether through modular plugins like Smoother or joint latent spaces like STACI—the field is moving toward a holistic understanding of how life is organized, one neighborhood at a time.