Spatial Graph Neural Networks for Tissue Architecture

102 concepts84 papers1 proposal

To understand how a biological tissue functions, it is not enough to know which cells are present; we must also know where they are and who they are talking to. Traditional single-cell RNA sequencing (scRNA-seq) acts like a blender, dissolving tissue architecture and losing the vital "GPS coordinates" of individual cells (Siyuan Huang et al. in Spatial transcriptomics: a new frontier in cancer research). Conversely, standard computer vision models treat histology images as rigid pixel grids or arbitrary image patches. This grid-based assumption fails because biological cells do not organize themselves into neat rows and columns; instead, they form irregular, non-Euclidean communities (Jingwen Wang et al. in Weakly Supervised Prostate TMA Classification via Graph Convolutional Networks).

Spatial graph representation solves this problem by mapping biological tissue as a mathematical network. By treating cells or sequencing capture spots as nodes and their physical interactions as edges, researchers can use Graph Neural Networks (GNNs) to model tissue microenvironments naturally and intuitively.

Constructing the Tissue Graph

The first step in spatial graph modeling is converting a continuous physical tissue slice into a discrete topological graph. This requires defining what constitutes a node and how to draw the edges between them.

Defining Nodes

Depending on the data modality, nodes can represent different biological scales:

Defining Edges

Drawing edges determines how information flows through the network. The authors of Deep Learning in Spatially Resolved Transcriptomics: A Comprehensive Technical View outline several common spatial proximity strategies:

  1. Fixed Distance Thresholds: An edge is drawn if the Euclidean distance between two cells is below a set micron threshold, capturing immediate physical contact (Yanan Wang et al. in Cell graph neural networks enable the precise prediction of patient survival in gastric cancer).
  2. k-Nearest Neighbors (k-NN): Each node connects to its $k$ closest neighbors. This ensures graph connectivity remains stable even in regions with varying cell densities (Jingwen Wang et al. in Weakly Supervised Prostate TMA Classification via Graph Convolutional Networks).
  3. Delaunay Triangulation: This geometric approach forms a planar graph where connections strictly reflect physical adjacency, preventing long-range edges from crossing empty space (Wenqi Lu 0001 et al. in Capturing Cellular Topology in Multi-Gigapixel Pathology Images).

More advanced frameworks use hybrid connectivity. For instance, Seyednami Niyakan in MUSTANG: MUlti-sample Spatial Transcriptomics data ANalysis with cross-sample transcriptional similarity Guidance constructs graphs that link nodes based on physical proximity within a single sample alongside transcriptional similarity across multiple distinct samples.

Message Passing: Blending Intrinsic and Extrinsic Context

Why does this graph representation work so well? The answer lies in the mechanism of message passing inherent to Graph Neural Networks.

In a tissue graph, each node begins with an initial feature vector representing its intrinsic molecular profile (such as gene expression counts or proteomic marker intensities). During a graph convolution layer, nodes send their feature vectors to their neighbors. Each node then aggregates its neighbors' incoming messages to update its own state.

[ Neighbor Cell 1 ] \
[ Neighbor Cell 2 ] -- ( Message Passing & Aggregation ) --> [ Target Cell State Updated ]
[ Neighbor Cell 3 ] /

As Maria Brbic et al. demonstrate in Annotation of Spatially Resolved Single-cell Data with STELLAR, this process generates cell embeddings that seamlessly integrate a cell's internal state with its extrinsic spatial environment. This mirrors the biological reality that a cell's identity and behavior are fundamentally shaped by its local niche (Sebastian Birk et al. in Quantitative characterization of cell niches in spatially resolved omics data).

By utilizing Graph Attention Networks (GATs), models can dynamically weigh these messages. Instead of treating all neighbors equally, the attention mechanism learns which neighboring cell types or signaling pathways are most relevant, smoothing out experimental noise and highlighting functional cellular communication (Arijit Sehanobish et al. in Disease State Prediction From Single-Cell Data Using Graph Attention Networks; Tianyu Liu et al. in spVelo: RNA velocity inference for multi-batch spatial transcriptomics data).

Hierarchical Graph Architectures

The tissue graph concept can also be nested. The authors of HEIST: A Graph Foundation Model for Spatial Transcriptomics and Proteomics Data introduce a hierarchical "graph of graphs." Here, an internal gene co-expression network is modeled inside each individual cell node, which itself sits within a global spatial cell-cell graph. This allows the model to capture intracellular regulatory mechanisms and extracellular tissue architecture simultaneously.

Unlocking Higher-Level Biological Insights

By shifting the analytical focus from isolated cells or arbitrary image pixels to structured spatial graphs, researchers can solve complex downstream tasks that were previously intractable.

Automated Tissue Segmentation

Instead of relying on manual annotations by pathologists, spatial GNNs can automatically discover functional tissue domains—such as tumor cores, healthy margins, or invasive fronts—by clustering nodes that share similar neighborhood topologies and transcriptomic distributions (Siyuan Huang et al. in Spatial transcriptomics: a new frontier in cancer research; Danni Hong et al. in Divergent combinations of enhancers encode spatial gene expression).

Clinical Outcome Prediction

Because the global graph structure preserves long-range spatial relationships, models can identify complex structural signatures that correlate with clinical outcomes. For example, Alexandro E. Trevino et al. in 478 Spatial omics and deep learning define an immunotherapy resistance niche in hepatocellular carcinoma used spatial graphs to identify an "interface niche"—a specific cellular barrier surrounding immune clusters that physically isolates them from tumors, acting as a direct biomarker for immunotherapy resistance.

Tracking and Topology

Representing tissue as a graph also provides robustness against physical deformations. In live-tissue imaging, cells constantly shift, compress, and divide. By focusing on the stable connectivity of a planar graph (which cells remain adjacent to each other) rather than volatile absolute spatial coordinates, tracking algorithms can follow cell lineages accurately across time even through massive tissue structural changes (Jochen Kursawe et al. in Robust cell tracking in epithelial tissues through identification of maximum common subgraphs).

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 84 papers
  2. Version 1 · 2026-10-04 · 170 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_spatial_graph_representation_00cd6f8d,
  title = {spatial graph representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatial-graph-representation/rdr_00cd6f8d50ac4daab7ae80ca16f9141b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_spatial_graph_representation_00cd6f8d_v2,
  title = {spatial graph representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatial-graph-representation/rdr_00cd6f8d50ac4daab7ae80ca16f9141b/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake