Spatially resolved transcriptomics (SRT) is a technology that measures gene activity at thousands of physical locations within tissue sections, preserving the spatial context that standard single-cell sequencing destroys. This spatial information is essential for understanding how different cell types are organized within organs and tumors.
A fundamental task in SRT analysis is spatial domain identification -- partitioning tissue spots into groups that share both similar gene expression profiles and similar spatial locations. This is analogous to identifying anatomically and functionally distinct regions within the tissue.
Existing computational approaches for spatial domain identification fall into two camps: methods based purely on gene expression (which ignore spatial relationships) and methods that incorporate spatial structure. Neither approach alone fully captures the complex interplay between molecular identity and physical location that defines tissue organization.
The authors developed AE-GCN (Autoencoder-assisted Graph Convolutional Neural Network), a deep learning framework that simultaneously learns from both gene expression patterns and spatial coordinates to identify tissue domains more accurately than existing methods.
AE-GCN has three core components working in concert. The first is an autoencoder (AE) -- a neural network that compresses high-dimensional gene expression data into a compact lower-dimensional representation called a latent embedding. This compression filters out noise while preserving the most informative molecular signatures of each tissue spot.
The second component is a graph convolutional neural network (GCN), which models the spatial relationships between tissue spots. Each spot is treated as a node in a graph, with edges connecting it to its nearest spatial neighbors. The GCN propagates information between connected nodes, allowing the model to incorporate the molecular context of surrounding tissue regions.
The third and most distinctive component is a clustering-aware contrastive module, which bridges the autoencoder and GCN outputs. This module uses a self-supervised learning strategy that brings together representations of spots likely to belong to the same tissue domain while pushing apart those from different domains -- aligning the two information streams into a unified spatial clustering solution.
The final spatial domains are identified by applying K-means clustering to the joint embeddings produced by the model. The number of clusters is set by the user based on expected tissue complexity, and the resulting assignments represent biologically and spatially coherent tissue regions.
AE-GCN was first evaluated on a widely used benchmark dataset: 12 tissue sections of the human dorsolateral prefrontal cortex (DLPFC), a brain region with well-characterized layered organization. Ground-truth annotations from neuropathologists provided a gold standard for comparison.
Performance was measured using the Adjusted Rand Index (ARI), which quantifies agreement between predicted and true tissue domain assignments on a scale from 0 (random) to 1 (perfect). AE-GCN achieved a mean ARI of 0.561 across all 12 DLPFC slices -- the highest among all compared methods, which included BayesSpace, STAGATE, stLearn, and Seurat.
The improvement was particularly notable in identifying the thin and structurally subtle Layer 1 of the cortex, which many existing methods collapse with adjacent layers. AE-GCN's spatial graph modeling allows it to detect sharp transitions between layers that expression-only methods miss.
The authors also tested the model on mouse hippocampus data generated with Slide-seqV2, an ultra-high-resolution SRT platform. AE-GCN successfully identified fine-grained anatomical sub-regions including the CA2 subfield, the lateral ventricle, and the habenula -- structures that no other compared method resolved correctly.
AE-GCN was applied to two cancer tissue datasets to test its ability to identify biologically meaningful regions within tumors. The first was pancreatic ductal adenocarcinoma (PDAC) -- a highly aggressive cancer notorious for its complex mixture of cancer cells and non-cancer support cells (stroma).
In PDAC tissue, AE-GCN identified spatial domains that showed strong alignment with known cell-type distributions. The cluster purity score -- measuring how well spatial clusters correspond to independent cell-type reference data -- reached 0.756, outperforming all five comparison methods on this dataset.
The second cancer dataset was invasive ductal carcinoma (IDC) of the breast. AE-GCN achieved a cluster purity of 0.865, again the highest of all methods, successfully separating tumor regions from normal ducts, stromal tissue, and adipose areas -- distinctions that are critical for understanding how different microenvironments within a tumor influence patient outcomes.
These results demonstrate that AE-GCN's combined expression-and-spatial learning approach captures the biologically relevant organization of cancer tissues, not just statistical patterns in gene expression data.
Beyond tissue domain identification, the autoencoder component of AE-GCN performs a secondary function: gene expression denoising. SRT data contains high levels of technical noise from dropout events (genes that fail to be detected despite being expressed). The autoencoder's reconstruction fills in these missing values, producing a cleaner expression matrix.
Using these denoised expression profiles, the authors identified differentially expressed genes (DEGs) that distinguish cancer-enriched domains from stromal regions in both cancer datasets. In PDAC tissue, standout genes included S100P and TNS4 -- both of which have established associations with pancreatic cancer aggressiveness and poor prognosis in the scientific literature.
In the IDC breast cancer dataset, top spatially enriched genes in tumor domains included SLC7A5 (a transporter linked to cancer cell metabolism and resistance to therapy) and RDH16 (associated with cancer cell differentiation). These findings were consistent with known cancer biology, validating the biological relevance of the spatial domains identified.
This demonstrates that AE-GCN's spatial domain analysis is not merely a geometric exercise -- it extracts biologically meaningful tissue regions within which clinically relevant molecular signatures can be discovered, including potential therapeutic targets and prognostic biomarkers.
A key practical advantage of AE-GCN is that it requires no manually annotated training data. The clustering-aware contrastive module operates through self-supervised learning -- the model creates its own training signal by generating augmented versions of each spot's embedding and learning to recognize that augmented versions of the same spot should be similar to each other and different from other spots.
This self-supervised design means AE-GCN can be applied to any SRT dataset without requiring expert annotators to label tissue regions in advance. Given the diversity of tissue types, species, and pathological conditions studied with SRT, this label-free design is essential for broad applicability.
The model's training combines three loss functions: a reconstruction loss (for the autoencoder), a graph reconstruction loss (for the GCN), and a contrastive loss (for the clustering-aware module). Balancing these three objectives allows the model to optimize gene expression fidelity, spatial structure preservation, and clustering alignment simultaneously.
Hyperparameter sensitivity was evaluated through ablation experiments. The model showed consistent performance across a range of values for key parameters including the number of graph neighbors, the dimensionality of the latent embedding, and the weighting of the contrastive loss -- indicating robustness rather than sensitivity to exact parameter choices.
The ability to automatically identify functionally distinct tissue domains has direct applications in cancer pathology. Understanding which physical regions of a tumor contain cancer cells versus stroma versus immune infiltrates provides a spatial map of the tumor microenvironment -- the complex ecosystem that determines how tumors respond to treatment.
Spatial domain analysis enables more precise tumor heterogeneity characterization -- identifying whether cancer cells in different parts of a tumor have different molecular states. This spatial heterogeneity is directly relevant to treatment resistance, as drugs may effectively target cells in one region while cells in an adjacent region survive and drive recurrence.
The denoising function of AE-GCN is particularly valuable for current-generation SRT platforms, which still have substantial dropout rates. Improved expression estimates within each spatial domain make downstream analyses -- such as cell type deconvolution, pathway activity analysis, and ligand-receptor interaction inference -- more accurate and reliable.
As SRT technology continues to improve in resolution and scalability, tools like AE-GCN will become central to building molecular atlases of disease -- comprehensive spatial maps of how gene activity varies across healthy and diseased tissue that can guide diagnosis, drug target identification, and treatment monitoring.