Graph Neural Networks for Spatial Transcriptomics: A Review of Graph Design and Downstream Applications

1. Introduction

Bulk RNA sequencing averages gene expression across tissue, whereas dissociated single-cell RNA sequencing resolves cellular heterogeneity but removes the original spatial arrangement of cells. Spatial transcriptomics (ST) measures gene expression while retaining the location of each measurement within tissue, allowing molecular states to be analysed in relation to tissue structure [1,2]. This spatial information enables analyses of tissue layers, tumour microenvironments, developmental gradients, local niches, and candidate cell–cell interactions [2]. Broader reviews have described ST technologies and analytical workflows [3]. In ST data, the biological meaning and spatial resolution of each observation depend on how the tissue is measured and processed. Platform resolution, capture geometry, segmentation, binning, image registration, and reference mapping determine what each observation represents and which spatial relations can be analysed [4].

GNNs are well suited to ST because they operate on entities connected by explicit relations. In an ST-GNN, however, the graph is a modelling decision rather than a direct copy of the tissue. A node may represent a spot, segmented cell, subcellular region, image patch, gene, or reference-derived pseudo-cell. An edge may represent spatial proximity, expression similarity, morphological similarity, a biological prior, or a correspondence across slices or datasets. These alternatives change both the information available during learning and the meaning of the resulting output.

Recent reviews have classified ST-GNN and related deep learning methods by graph architecture, general methodology, data modality, or analytical task [5,6,7]. A broader review of GNNs for single-cell omics provides additional architectural and application context [8]. More recently, graph design has also been examined as an organising principle for deep learning-based multi-omics integration, with attention to node schema, edge semantics, interaction type, and graph context [9]. These classifications identify major model families, tasks, and graph design choices, but they do not by themselves show how measurement units become nodes, how evidence becomes edges, or how a constructed graph affects a specific prediction in spatial transcriptomics. Two methods can use the same GNN backbone while representing different entities, propagating information over different relations, and supporting different biological conclusions.

This review examines ST-GNN methods through the sequence shown in Figure 1: spatial inputs, graph construction, graph-based learning, and downstream applications. We separate five analytical components within this sequence: node definition, edge construction, graph structure, the use of the graph during learning, and the output produced for a downstream task. We use $G=(V,E,X)$ to denote a graph with node set V, edge set E, and node feature matrix X. This organisation makes it possible to compare methods according to the entities they model, the evidence encoded in their relations, the operations performed on those relations, and the resolution and form of their outputs.

The review covers 89 representative ST-GNN studies and uses the main text to explain recurring methodological patterns. Section 2 defines the measurement units and information sources available before graph construction. Section 3 compares node definitions, edge-construction rules, and basic and extended graph structures. Section 4 examines how constructed graphs control propagation, weighting, topology adaptation, reconstruction, and cross-source correspondence. Section 5 compares six downstream application categories and states the output and evaluation requirements of each category. Figure 2 provide method-level comparisons of inputs, graph construction, learning objectives, outputs, tasks, dataset scale, computational resources, and code availability. Section summarise the methodological gaps identified by these comparisons. Non-GNN methods are included only when they clarify a graph-design assumption or an evaluation setting.

2. Spatial Inputs and Information Sources

Before an ST-GNN graph is constructed, two types of inputs are already available: measurement units and information sources. In the spatial data input layer of Figure 1, measurement units refer to the technical units in which gene expression and spatial information are recorded after measurement or preprocessing. Information sources include spatial, molecular, histological, prior, and reference information that can support graph construction. This section introduces these inputs as the data layer that precedes graph construction. As shown in panel (A) of Figure 1, ST data are generated at one of three measurement resolutions—spot, single-cell, or subcellular. These resolutions determine what each graph node can represent in later stages.

2.1. Measurement Settings in Spatial Transcriptomics

Spatial transcriptomics (ST) measures gene expression in situ while preserving tissue context, providing information that is not available from dissociated single-cell profiles [1]. Previous reviews have summarised the development of spatial transcriptomics technologies and their analytical implications [2]. Other surveys have further discussed platform diversity and downstream analysis workflows [3]. ST platforms differ in how molecular signals are captured, localised, and assigned to analysis units [4]. Capture-based platforms typically measure gene expression at spatial locations [1]. High-resolution sequencing platforms can provide spatial bins or molecule-level coordinates that are later aggregated for analysis [10]. Imaging-based platforms, such as MERFISH [11] and seqFISH+ [12], localise transcripts at single-cell or subcellular resolution before segmentation or aggregation. These differences directly influence graph construction because graph models are built from the measurements provided by each platform rather than from an ideal biological representation.

2.2. ST Measurement Units

This subsection describes spots, bins, and segmented cells as the measurement units generated by ST platforms or preprocessing workflows. It focuses on what is measured or assigned before graph construction. How graph nodes are constructed and the biological entities they represent are discussed in Section 3.1.

A capture spot records transcripts collected at a spatial location and may contain signals from multiple cells, depending on the platform resolution, tissue density, and capture geometry [1]. A spatial bin aggregates molecular counts within a defined region when the original measurements are available at a finer spatial resolution than the analysis requires [10]. A segmented cell is obtained by assigning transcript locations to cell boundaries or cell centres, typically in imaging-based or high-resolution workflows [13]. Although it more closely represents an individual biological cell than a spot or bin, it is still produced through computational segmentation. Consequently, misassigned transcripts, inaccurate cell boundaries, or centroid errors may affect both the measured expression and spatial location [13].

2.3. Information Sources for Graph Construction

Edge evidence refers to information that supports an assumed relationship before it is represented in a graph. The relationship may later be encoded as an edge, an edge weight, a graph view, a link across information sources, or a higher-order structure. Five sources of evidence commonly appear in ST-GNN methods: spatial coordinates, molecular profiles, histological or morphological information, biological priors, and reference information. Hybrid and multimodal graphs typically combine several of these sources rather than introducing a separate category of evidence.

2.3.1. Spatial Coordinates and Local Proximity

Spatial coordinates indicate that nearby units may share tissue context, physical neighbourhoods, or local microenvironmental conditions. However, spatial proximity alone does not establish molecular similarity, long-range correspondence, ligand–receptor communication, or functional interactions. Identifying these relationships usually requires additional molecular information, biological priors, or external validation [14]. Spatial coordinates can be converted into graph topology using radius cutoffs, k-nearest neighbours, Delaunay triangulation, or alpha-complex construction. These graph construction choices are discussed in Section .

2.3.2. Molecular Profiles and Expression Similarity

Expression profiles can be used to estimate transcriptomic similarity between spots, bins, or cells, including units that are not immediate spatial neighbours. However, the resulting similarity estimates are sensitive to sparsity, dropout, normalisation, feature selection, batch effects, and platform-specific noise [15]. Similarity computed from the expression profiles of two nodes provides an estimate of their molecular relatedness, whereas spatial coordinates capture positional relatedness. These two roles should remain conceptually distinct until the method specifies how they are used in graph construction and learning.

2.3.3. Histology and Morphology

Paired histology images and morphological features provide information about tissue architecture, cellular morphology, and pathological structure [7]. They can complement spatial coordinates and expression profiles, but morphological similarity is not equivalent to similarity in gene expression. The reliability of histological information depends on staining quality, image resolution, registration, feature extraction, and whether the relevant transcriptional state is visible in tissue morphology [16]. Histology therefore serves as an auxiliary source of information whose relevance depends on the downstream task.

2.3.4. Biological Priors

Biological priors provide external knowledge about relationships among genes, cell types, pathways, and candidate communication events. In current ST-GNN methods, these priors most commonly take the form of ligand–receptor pairs or pathway annotations [17]. Regulatory and gene-level priors are less common. They appear mainly in recent multiview models, which combine multiple sources of information, or heterogeneous graph models, which include different types of nodes or relations, such as stKeep [18] and MAGNET [19]. We therefore regard gene regulatory networks and gene–gene relations as emerging extensions rather than widely used sources of information. Like other external resources, biological priors are curated and context dependent. They are not direct observations from an ST assay.

2.3.5. Reference Information

Reference information links ST measurements to external datasets or atlases. These references are most often single-cell RNA sequencing datasets that contain cell type labels, cell state annotations, or expression profiles [20]. They can support label transfer, deconvolution, alignment, and mapping, but they also introduce assumptions about the correspondence between the reference and target data [20]. Relevant states may be absent from the reference, altered by disease or development, or affected by differences between platforms. In these cases, the model may transfer the closest available annotation rather than the correct biological state. The construction of pseudo-nodes and links across information sources is discussed in Section 3.

The measurement units and information sources described above are inputs to graph construction, but they do not yet form a graph. They become part of the graph only after a method defines the nodes, converts the available information into relations, and represents these relations as edges, weights, views, links across sources, or higher-order structures.

3. Graph Construction

The measurement units and edge evidence defined above provide the inputs for graph construction. Graph construction begins when a method represents these inputs as nodes, edges, and other graph components, as shown in the graph construction block of Figure 1. This section discusses node construction, edge construction, basic and extended graph structures, and reporting practices. How the resulting graph is used in model learning is discussed in Section 4.

3.1. Node Definition and Granularity

A graph node is defined by the modelling process and is not necessarily identical to a measurement unit. In some methods, nodes directly correspond to platform outputs such as spots, bins, or segmented cells. Other methods introduce computational entities, including pseudo-spots, image patches, genes, cell types, morphological regions, and units obtained from reference data. The first question in graph construction is therefore what each node represents and at what biological or technical scale its features are defined.

Panel (B) of Figure 1 illustrates this distinction by showing measurement units as dark dots and the nodes constructed from them as coloured circles. The mapping between the two depends on the graph structure. A measured node may correspond to one measurement unit, a pseudo-node may be constructed from reference data, and a graph with multiple entity types may obtain each node type from a different source. The following subsections discuss these node types in turn.

3.1.1. Spot Nodes

In spot-level graphs, each capture location is represented as a node, usually with a processed expression vector as its features. Such a node represents a local mixture of cellular signals rather than an individually segmented cell. Unless the method introduces additional node types or increases the output resolution, its embeddings, clusters, and predictions remain defined at the spot level. GraphST [21] illustrates this basic setting. STAIG [22] uses the same spot-level nodes but changes how image information modifies the graph between spots.

3.1.2. Cell Nodes

In cell-level graphs, nodes represent segmented cells or cell-level observations, most often in imaging or spatial single-cell data. Because segmentation determines the node set, cell boundaries, centroids, and transcript assignments become part of graph construction rather than merely preprocessing details. CellNEST [23] and DeepTalk [24] illustrate this setting. Both define spatial or ligand–receptor-informed connections between cells, so the resulting neighbourhoods, interactions, and predictions differ fundamentally from spot-level outputs rather than simply providing a denser representation.

3.1.3. Pseudo-nodes and Nodes from Reference Data

Some deconvolution methods construct pseudo-spots from annotated single-cell reference data by mixing several cells and using their cell-type proportions as known labels. DSTG [25] connects these pseudo-spots with measured ST spots through expression similarity. STdGCN [26] further uses cell-type-aware sampling and separates an expression graph containing pseudo-spots and measured spots from a spatial graph containing only measured spots. These design choices determine how reference information is transferred to the spatial data.

3.1.4. Patch, Gene, and Multi-Entity Nodes

Some methods use node types that differ further from the original measurement units. Methods for image integration and super-resolution may introduce image patches, superpixels, or sub-spot units. For example, scstGCN [27] combines information from cells, spots, and image patches to infer expression at single-cell resolution from spot-level ST data.

Other methods introduce genes or several entity types into the graph. GCNG [28] represents genes as nodes. STING [29] constructs a gene co-expression graph for each spot and uses an inner GNN to encode it. The resulting spot representations are then connected in a second graph based on spatial proximity. HEIST [30] connects spatial cell–cell structure with gene or protein co-expression. stKeep [18] represents cells, genes, and histological regions in a heterogeneous graph. For these graphs, each node type and its source should be stated explicitly.

3.2. Edge Construction and Graph Structure

After the node set is defined, edge evidence is converted into edges or other graph components. Edges may be binary or weighted, directed or undirected, typed or untyped, and pairwise or higher order. A method may also preserve different information sources as separate graph views rather than merging them into a single adjacency matrix. The next question in graph construction is therefore how each source is represented in the graph.

The information source and the construction rule should be distinguished because the same source can produce different graph structures. Spatial coordinates may define a radius graph, a k-nearest-neighbour graph, a Delaunay graph, an alpha-complex graph, or an adjacency matrix weighted by distance. Expression profiles may define edges between molecularly similar nodes, assign weights to an existing spatial graph, or form a separate expression graph. Histology, biological priors, and reference information may also be represented through scalar weights, candidate edges, typed edges, links between sources, or auxiliary graph components. The construction rule should therefore be reported separately from the information source.

The edge panel in Figure 1(B) illustrates several resulting structures, including a single adjacency matrix, heterogeneous and virtual edges, weights defined by spatial distance or feature similarity, multiple graph views, and hyperedges. The following subsections discuss neighbourhood construction, weighted edges, and structures beyond a single adjacency matrix.

3.2.1. Neighbourhood Construction

Neighbourhood construction defines the candidate edge set using rules such as k-nearest neighbours, radius thresholds, Delaunay triangulation, or alpha-complex construction. Although these rules are often treated as routine preprocessing, each imposes a different assumption about local connectivity. A k-nearest-neighbour graph fixes the number of neighbours for each node. A radius graph instead uses a fixed physical distance. An alpha-complex graph uses spatial geometry to preserve local connectivity.

GraphST [21] and SEDR [31] construct k-nearest-neighbour graphs from spatial coordinates, although they use the resulting graphs in different learning systems. Other methods adopt different geometric rules. Novae [32] constructs a Delaunay graph, whereas SCAN-IT [33] uses an alpha complex. From the perspective of graph construction, these methods differ in which node pairs are connected before learning begins.

Neighbourhood construction is transparent and reproducible, but its parameters remain modelling choices. The radius threshold, number of neighbours, or geometric rule determines which local connections are available to the model. A later GNN may learn attention weights over these connections, but it can only reweight edges already included in the candidate set. Learned edge weighting is discussed in Section 4.2.

3.2.2. Weighted Edges

Weighted edge construction assigns scalar values to connected node pairs. The weights may reflect spatial distance, expression similarity, morphology, biological priors, or fixed combinations of several information sources. The key distinction is whether these weights are specified during graph construction or learned by the model.

SpaGCN [34] provides a clear example of weights specified before learning. It combines spatial distance with information extracted from histology images, so edge strength is determined before message passing. Spatial-ID [35] represents a less clear boundary case because it uses an adaptive weighted spatial graph. Nevertheless, these weights should be distinguished from attention weights or other edge weights learned during model training. Otherwise, fixed graph weights and learned weighting mechanisms may be incorrectly treated as the same design choice.

3.3. Basic Graph Structures

The node and edge choices described above can produce a basic graph structure with one dominant node type and one main adjacency matrix. In ST-GNNs, this is often a graph of spots or cells connected according to spatial proximity, expression similarity, or morphological information. A basic structure is not necessarily biologically simple. The node type, edge construction rule, and weighting scheme still determine which local connections are available to the model. The structure is considered basic because one main edge type connects one dominant set of nodes. It therefore provides a reference point for structures with multiple views, entity types, edge types, or information sources.

3.4. Extended Graph Structures

Some graph structures cannot be represented by a single adjacency matrix over one node type. They may preserve several edge sources as separate graph views, include multiple node or edge types, represent connections among groups of nodes, or link nodes across references, slices, samples, or modalities. These structures do more than add complexity to a basic graph. They allow the graph to represent different forms of biological and technical information.

3.4.1. Multiview Graphs

Multiview graphs preserve spatial proximity, expression similarity, histological similarity, or other edge sources as separate views rather than combining them into one adjacency matrix. Graph construction must therefore specify how each view is built, what information it represents, and whether the views remain separate during learning.

MuCoST [36] uses separate graphs for spatial proximity and expression correlation, whereas stMVC [37] uses separate spatial and histological similarity graphs. Because these relations are represented in different graphs, the model can process each information source separately. This differs from methods that merge spatial, molecular, and histological information into a single weighted adjacency matrix.

3.4.2. Heterogeneous and Hierarchical Graphs

Heterogeneous and hierarchical graphs contain multiple entity types, edge types, or graph levels. stKeep [18] connects cells, genes, and histological regions. STING [29] combines an outer spatial graph with inner gene–gene graphs. HEIST [30] connects spatial cell–cell structure with gene or protein co-expression. Unlike a basic graph of spots, these structures represent connections among different biological entities or across different scales. Each edge type therefore supports a different interpretation.

3.4.3. Hypergraphs

Hypergraphs represent connections among groups of nodes rather than only between node pairs. A hyperedge can connect several units that share a common higher-order pattern. HAST [38], for example, constructs local hypergraphs from gene expression, spatial information, and histology. It then aggregates them to represent many-to-many relationships among spots. In this setting, the basic connection unit changes from a pair of nodes to a group of nodes. This change affects how the resulting embeddings and clusters should be interpreted.

3.4.4. Links Across Sources

Links across sources connect nodes that do not belong to the same original spatial graph. In deconvolution, these links connect pseudo-nodes or reference profiles to measured spatial nodes, as in DSTG [25] and STdGCN [26]. In spatial integration, STAligner [39] uses mutual nearest neighbour anchors across slices. SLAT [40] aligns graphs from different slices through adversarial matching. Graspot [41] constructs probabilistic correspondences using optimal transport, while Tacos [42] combines graph structure with community and anchor information. In these methods, correspondence is explicitly constructed as part of the graph rather than inferred only from a single spatial neighbourhood.

3.5. Summary of Graph Construction Choices

Graph construction determines three properties of an ST-GNN: the entities represented by the model, the relations available for information exchange, and the organisation of those relations within the graph. Node definition sets the analytical unit and constrains the resolution of the resulting output. Edge construction specifies which spatial, molecular, histological, prior, or cross-source relations enter the model. Graph organisation determines whether these relations are merged into one adjacency structure, preserved as separate views or edge types, arranged across hierarchical levels, or represented as higher-order connections.

These construction choices are distinct from the role assigned to the graph during learning. The same constructed graph may define a message-passing neighbourhood, provide candidate relations for scoring, serve as a reconstruction target, support cross-source alignment, or impose a structural constraint. Conversely, similar learning operations may be applied to graphs that represent different entities and relations. Separating graph construction from graph usage therefore allows methods to be compared at both the representation and learning stages.

4. Constructed Graphs in Learning Systems

Panel (C) of Figure 1 shows how a constructed graph enters representation learning. The graph can define message passing, restrict candidate relations, provide structure for correspondence inference, or become a reconstruction target. Its role is determined by how the model uses it, not by the construction rule alone.

4.1. Roles of Constructed Graphs in Learning

In propagation models, the graph specifies which nodes exchange information. GraphST [21] uses a fixed spatial graph for GCN aggregation and reconstructs expression rather than spatial edges. In candidate-relation models, the graph restricts the relations that can be scored. CellNEST [23] builds directed communication candidates from spatial proximity and ligand–receptor expression, then ranks those candidates with attention. GCNG [28] has a different role: it builds a spatial cell graph for each candidate gene pair and performs graph-level binary classification, so the predicted unit is the gene pair rather than an edge in the spatial graph.

Constructed graphs can also provide structure for correspondence inference. SLAT [40] learns graph-informed representations within each slice and then matches cells or spots across slices. Graspot [41] learns graph-aware representations before estimating a soft transport plan with unbalanced optimal transport. Here the within-slice graph informs the matching problem but does not itself encode the cross-slice correspondence.

Some methods treat a graph as a target. DeepLinc [43] generates a cell–cell interaction adjacency from spatial single-cell expression data. The reconstructed graph can be reported directly or used as an auxiliary target that shapes the embedding.

4.2. Graph Structure during Learning

ST-GNNs differ in which parts of graph structure can change during training. SpaGCN [34] fixes a weighted adjacency from spatial distance and histology before optimisation. STAGATE [44] fixes the edge set but learns attention coefficients on those edges; non-neighbouring nodes in the initial graph never enter the attention calculation.

Other methods change the graph more directly. SGAE [45] first removes spatial edges that cross expression-based preclusters, then creates representation-dependent graph views through low-similarity edge removal and graph diffusion. asGNN [46] learns a stochastic subgraph through a variational information bottleneck, retaining or suppressing relations from a candidate spatial graph. STAGUE [47] jointly infers adjacency and node representations, allowing connections absent from the initial spatial graph. These examples separate fixed graph construction, learned edge weighting, candidate-edge selection, and broader topology inference.

Adaptive adjacency should also be separated from view fusion. STMGAMF [48] adapts node relations within spatial and feature graph branches, then fuses the resulting embeddings. The first operation changes propagation within a graph view; the second combines representations across views.

4.3. Graph Operators and Message Passing

Graph operators determine how information moves across constructed relations. Standard graph convolution aggregates over one adjacency matrix with a shared propagation rule. Graph attention changes neighbour contributions within the supplied edge set. Multiview and relational operators preserve information that a single adjacency would merge. Spatial-MGCN [49] encodes spatial and expression graphs separately before combining their representations. RGAST [50] represents communication at several spatial scales and applies relation-specific attention.

Other operators change the scope or geometry of propagation. SiGra [51] uses a graph transformer to integrate transcriptomic and image information. HGNN [52] constructs hyperedges from dense overlapping subgraphs and passes messages between nodes and hyperedges, so propagation occurs through node groups rather than only pairwise edges. MManiST [53] encodes the same graph in Euclidean, Poincaré, and Lorentz spaces before combining the representations.

4.4. Decoders, Prediction Heads, and Outputs

Outputs differ in unit, form, and resolution. SEDR [31] produces a spot embedding and reconstructs expression. Hist2ST [54] and THItoGene [55] predict expression from histology at the measured spot grid, whereas scstGCN [27] predicts expression below the measured resolution.

Pair-level outputs attach predictions to candidate pairs. GCNG [28] classifies candidate gene pairs, CellNEST [23] scores directed ligand–receptor candidates, and DeepTalk [24] predicts communication from subgraphs centred on candidate cell pairs. Graph-level outputs reconstruct relations instead of scoring a fixed candidate list. DeepLinc [43] reconstructs a cell–cell interaction network, and MAGNET [19] reconstructs several biological interaction graphs.

Composition outputs remain node-indexed but have a different interpretation. DSTG [25] and STdGCN [26] estimate a vector of cell-type proportions for each spot from reference-derived pseudo-spots. NicheCompass [56] reports cell- or spot-indexed niche representations and gene-programme activities that describe local microenvironments. These output types support different levels of biological interpretation, from spatial embeddings and expression estimates to candidate interactions, reconstructed networks, and reference-dependent composition estimates.

4.5. Training Objectives

Training objectives specify what the representation must recover, distinguish, match, regularise, or predict. Reconstruction losses target different objects: SEDR [31] reconstructs expression and graph structure, DeepLinc [43] reconstructs a cell–cell interaction network, and MAGNET [19] reconstructs several biological relation graphs.

Contrastive objectives depend on how related units or views are defined. CCST [57] contrasts nodes with context from a spatial neighbour graph. MuCoST [36] defines relatedness through spatial and expression views. conST [58], STAIG [22], Tacos [42], and SpaMask [59] use multimodal, augmented, community-aware, or masked views. HAST [38] uses hyperedges for group-wise message passing, but its contrastive positives are not defined by hyperedge membership alone.

Alignment and transfer objectives connect datasets or slices. STAligner [39] forms triplets from mutual nearest neighbours across slices. stGuide [60] uses reference annotations to organise the representation and transfer labels. Graspot [41] uses unbalanced optimal transport to estimate partial correspondence. Adversarial learning appears in stAA [61], where a discriminator separates graph-autoencoder codes from samples drawn from a prescribed prior and the encoder is trained to match the two distributions. Supervised objectives include reference-based label transfer in Spatial-ID [35] and reference-derived proportion prediction in DSTG [25].

4.6. Summary of Graphs in Learning Systems

The value of graph learning in spatial transcriptomics is not greater flexibility by default. It is the placement of graph flexibility where the biological uncertainty lies. When spatial proximity, histology, or ligand–receptor knowledge provides a reliable prior, the graph can constrain representation learning. When these priors are only proxies for communication, correspondence, or predictive dependence, the model needs mechanisms that revise edge weights, select candidate relations, or infer structure.

This distinction matters for edge-indexed outputs. Attention weights, interaction scores, reconstructed adjacencies, and adaptive predictive graphs all appear as edges, but they encode different quantities. The relevant question is where the graph departs from construction-time priors and whether that freedom matches the biological uncertainty of the ST task.

5. Downstream Applications

The preceding sections traced the process from ST inputs to graph construction and graph learning. This section turns to the application layer in Figure 1. It examines the outputs produced by ST-GNNs, the graph structures commonly used for different tasks, and the biological interpretations supported by these outputs.

The discussion is organised around six task categories. The broader categories of representation, relation inference, and mapping are used only to help orient the reader. Representative methods are included to show how downstream tasks are connected to choices in graph construction and learning.

The six task categories reviewed here focus on the primary outputs of ST-GNN methods. These outputs may also support subsequent analyses, including spatially variable gene detection, pathway enrichment, trajectory analysis, and the identification of disease-associated regions. Because these analyses are usually performed after embeddings, spatial domains, predicted expression, or aligned representations have been obtained, we do not treat them as separate task categories in this review.

5.1. Downstream Tasks, Outputs, and Graph Construction Patterns

This subsection groups ST-GNN applications according to their main downstream outputs. For each task category, we examine what the model produces, which graph structures are commonly used, and what biological interpretation the output can support. The categories are organised by analytical purpose rather than by GNN architecture, because similar operators may be used for different tasks, while the same task may be addressed using different graph structures and learning objectives.

5.1.1. Spatial Representation and Domain Identification

Methods for spatial representation and domain identification produce node embeddings, spatial domain labels, or both. Most assume that local spatial context is informative, but they incorporate this context in different ways. A spatial graph may define the message-passing neighbourhood or the candidate edges for attention. It may also provide a reconstruction target, define positive pairs for contrastive learning, be refined during training, or support the integration of multiple modalities. The task output alone therefore does not reveal the graph assumptions used to produce it. Methods for spatial domain identification illustrate this distinction. GraphST [21] uses a spatial graph in a self-supervised representation learning system. STAGATE [44] begins with a similar assumption about local neighbourhoods but learns different weights for candidate neighbours. Its outputs are therefore shaped by attention rather than uniform aggregation. SEDR [31] learns spatial representations through expression and graph reconstruction, whereas CCST [57] uses contrastive learning between nodes and their graph context. These methods may produce similar embeddings or domain labels, but they differ in what information the local graph is expected to preserve. These outputs mainly support the identification of spatially coherent groups, regional patterns, and node representations that incorporate local context. They do not by themselves establish biological tissue boundaries or cell-level mechanisms. Such claims require additional evidence from marker genes, histology, explicit boundary evaluation, perturbation experiments, or other biological validation. The comparison of recent methods in Section 5.2 develops this point further by showing that similar task labels and benchmark scores can conceal important differences in graph construction and graph usage.

5.1.2. Integration, Alignment, and Mapping

Methods for integration, alignment, and mapping produce aligned embeddings, correspondences across slices or samples, transferred labels, or mapped spatial structures. Unlike methods for domain identification, they do not focus only on whether spatial context is coherent within a single sample. They also identify which units across slices, samples, modalities, or references should correspond to one another. Graph construction must therefore preserve spatial organisation within each source while introducing links that support comparison across sources.

STAligner [39] combines within-slice spatial structure with anchors across slices to align comparable spatial contexts. SLAT [40] matches graph representations from different slices. Graspot [41] uses optimal transport to infer probabilistic correspondences rather than fixed one-to-one matches. Tacos [42] combines local graph representations with supervision from cross-slice anchors. The resulting correspondences are supported by the model rather than directly observed. Their interpretation depends on how anchors are defined, how well spatial structure is preserved within each source, and whether differences specific to individual samples remain detectable.

5.1.3. Cell–cell Communication and Niche Inference

Methods for cell–cell communication and niche inference produce two main types of output. Communication methods usually prioritise candidate interactions between cells or ligand–receptor pairs. Niche methods instead identify the microenvironmental state or programme associated with a cell or tissue region.

GCNG [28] and CellNEST [23] mainly score candidate interactions. DeepLinc [43] reconstructs an inferred interaction network, whereas NicheCompass [56] learns niche representations informed by biological priors. These outputs can support hypotheses about cellular interactions or descriptions of local microenvironments. However, they do not by themselves establish active signalling or downstream biological responses. Such claims require additional molecular, spatial, or perturbation evidence.

5.1.4. Imputation and Denoising

Methods for imputation and denoising produce reconstructed, denoised, or imputed expression values. These tasks are often grouped together because they all generate expression matrices. However, the meaning of a recovered value depends on the information available to the model and the assumptions used during recovery. The estimate may be informed by neighbouring nodes, generated by a probabilistic model, predicted from multiple modalities, or produced through a combination of these mechanisms.

Impeller [62] estimates missing or noisy expression using spatial context and similarities between expression profiles. stMCDI [63] instead uses masked conditional diffusion. Its imputed values are generated under spatial and gene correlation constraints rather than obtained only by averaging nearby observations. SiGra [51] combines transcriptomic and histological information through multimodal graph learning, so its recovered expression may reflect both molecular neighbourhoods and image context.

These methods may produce similar denoised expression matrices despite relying on different recovery assumptions. Imputed or denoised expression should therefore be interpreted as an estimate produced by a specific graph and learning system, not as directly recovered ground truth. Stronger validation requires held-out genes or locations, independent measurements, biological replicates, or estimates of prediction uncertainty.

5.1.5. Deconvolution

Deconvolution methods estimate cell type proportions or related measures of cellular composition for spatial units. Here, the graph is used mainly to transfer information from a labelled reference dataset to spatial measurements rather than to identify local tissue domains. The resulting estimates therefore depend on the reference data, including their cell type coverage, similarity to the spatial sample, and the procedure used to connect the two sources. Spatial context may further influence the estimated composition.

DSTG [25] makes this transfer process explicit. It constructs pseudo-spots from single-cell reference data and uses them to infer cell type composition in measured spatial spots. STdGCN [26] distinguishes expression links between reference pseudo-spots and spatial spots from spatial links among measured spots. Its estimates therefore reflect both similarity to the reference and local spatial context. GTAD [64] also transfers information between single-cell reference data and spatial measurements through a graph.

Across these methods, the graph helps incorporate reference information into spatial analysis, but it does not remove dependence on the reference. Claims about tissue composition should therefore be supported by checks of reference coverage and comparability between the reference and target samples. Tests with alternative references and independent validation of cell types or spatial distributions can provide further support.

5.1.6. Expression Prediction from Histology and Super-Resolution

Methods for expression prediction from histology and super-resolution produce predicted expression maps, estimates at finer spatial resolution, or values for sub-spot locations. Unlike conventional denoising, these methods often predict beyond the locations or resolution directly measured by the transcriptomic assay. They combine morphology, spatial context, and observed expression, but their outputs remain predictions for unmeasured locations or spatial scales.

Hist2ST [54] uses image patch representations and local spatial dependencies to predict gene expression from histology. Its output is therefore an expression estimate conditioned on tissue morphology. THItoGene [55] follows a similar approach by learning representations from images and graphs to predict gene expression. scstGCN [27] extends this setting by combining several data sources to infer expression at single-cell resolution from spot-level measurements.

These methods can generate informative expression maps, but predicted expression should remain distinct from directly observed expression. This distinction is especially important when predictions are made below the native resolution of the assay. Stronger evidence may come from validation on held-out spatial locations, comparison with higher-resolution measurements, analysis of errors for individual genes, and uncertainty estimates for unmeasured locations.

5.2. Spatial Domain Identification as a Comparison Case

Spatial domain identification shows why benchmark scores should be read together with graph construction and learning choices. SpaMask, STAIG, STMGAMF, DACN, and stGCL are all evaluated on DLPFC benchmarks and report ARI or NMI against annotated cortical layers. Their outputs and metrics look comparable, but their graph designs differ. SpaMask masks nodes and edges during self-supervised learning. DACN combines a graph branch with an autoencoder. stGCL propagates expression and histology through graph attention and combines reconstruction with contrastive learning. STAIG perturbs edges using histology images. STMGAMF uses separate spatial and expression graphs and learns how to combine them.

ARI and NMI measure agreement with annotated domains, but they do not identify which model component produced the improvement. The reported DLPFC results also use different aggregation rules. SpaMask reports a median ARI of 0.596 across 12 slices. STAIG reports a median ARI of 0.69 and a median NMI of 0.71. STMGAMF reports a median ARI of 0.61 and a median NMI of 0.68. stGCL reports a mean ARI of 0.56 and a mean NMI of 0.65. DACN is not directly comparable because its main text presents boxplots and significance tests across 12 DLPFC sections, while reporting a slice-level ARI of 0.67 for section 151674. These values are useful evidence, but not a unified leaderboard.

Table 1. Graph construction and graph usage in methods for spatial domain identification evaluated on DLPFC benchmarks. Methods with similar task outputs and clustering metrics may use different graph structures, learning roles, and model components.

Method	Graph construction	Graph usage during learning	Component not reflected by the score
SpaMask [59]	Spatial graph constructed from spot coordinates	Masked nodes are used for feature reconstruction, while masked edges are used for contrastive learning within a shared encoder	Node and edge masking
STAIG [22]	Spatial graph whose edges are perturbed using histology images	Representations from perturbed graphs are compared through contrastive learning between neighbouring spots	Edge perturbation using histology
STMGAMF [48]	Separate spatial and expression graphs, with edge weights learned during training	The two graphs are encoded separately and then combined under reconstruction and spatial regularisation objectives	Individual graph views and loss terms
DACN [65]	Spatial graph constructed from spot coordinates	The graph encodes expression representations produced by an autoencoder, and the two branches are combined for reconstruction	Autoencoder and graph branches
stGCL [66]	Spatial graph containing expression and histology features	The graph propagates both modalities, reconstructs each modality, and supports contrastive learning	Contribution of histology

Table 1. Graph construction and graph usage in methods for spatial domain identification evaluated on DLPFC benchmarks. Methods with similar task outputs and clustering metrics may use different graph structures, learning roles, and model components.

Method	Graph construction	Graph usage during learning	Component not reflected by the score
SpaMask [59]	Spatial graph constructed from spot coordinates	Masked nodes are used for feature reconstruction, while masked edges are used for contrastive learning within a shared encoder	Node and edge masking
STAIG [22]	Spatial graph whose edges are perturbed using histology images	Representations from perturbed graphs are compared through contrastive learning between neighbouring spots	Edge perturbation using histology
STMGAMF [48]	Separate spatial and expression graphs, with edge weights learned during training	The two graphs are encoded separately and then combined under reconstruction and spatial regularisation objectives	Individual graph views and loss terms
DACN [65]	Spatial graph constructed from spot coordinates	The graph encodes expression representations produced by an autoencoder, and the two branches are combined for reconstruction	Autoencoder and graph branches
stGCL [66]	Spatial graph containing expression and histology features	The graph propagates both modalities, reconstructs each modality, and supports contrastive learning	Contribution of histology

Transfer to other datasets depends on the available modalities and modelling assumptions. A method that benefits from histology on DLPFC may lose that advantage when images are unavailable or differ in quality. STAIG uses histology to guide edge perturbation, and stGCL reports lower performance when its histology encoder is removed. Component-level comparisons are therefore needed to separate graph construction, graph usage, training objectives, modality availability, and clustering procedures.

5.3. Summary of Downstream Outputs and Validation

Across the six downstream task categories, the reviewed methods fall into three broader analytical modes according to what their outputs represent: representation and recovery, relational inference, and mapping and transfer across sources. This grouping does not replace the task taxonomy above. It connects each output type to the biological conclusion it can support and to the evidence relevant to evaluating that conclusion. Representation and recovery methods produce embeddings, labels, or expression estimates. Relational-inference methods produce candidate interactions, inferred networks, or summaries of local microenvironments. Mapping and transfer methods produce correspondences, transferred annotations, or composition estimates across slices, samples, modalities, or reference datasets. Table 2 states the role of the graph, the model output, the supported conclusion, and the corresponding validation evidence for each analytical mode.

Benchmark metrics assess specific properties of model outputs, while broader biological conclusions benefit from additional forms of validation. DLPFC spatial-domain benchmarks test clustering agreement and spatial representation, but they do not evaluate communication inference, deconvolution, cross-dataset integration, imputation, or histology-based expression prediction [67]. Benchmark design and platform characteristics can change reported method rankings [68,69]. Boundary and marker-gene evidence strengthen spatial-domain results. Held-out genes or locations test imputation results. Signalling evidence not used to construct candidate interactions strengthens communication results. Sensitivity analyses across references assess deconvolution results. Matched measurements at the target resolution provide direct validation for super-resolution results.

6. Future Directions

Future work can build on four priorities identified in this review. These are testing graph assumptions, developing evaluation for specific tasks, improving scalability, and representing a broader range of biological connections.

6.1. Testing Graph Assumptions

Graph construction should be evaluated as part of the model rather than treated as fixed preprocessing. Useful controls include removing individual edge sources, shuffling biological priors, changing neighbourhood rules, subsampling reference data, and testing alternative links across datasets. For example, methods for cell–cell communication can be tested with shuffled or held-out ligand–receptor pairs, while integration methods can be evaluated under different anchor rules. Representation methods can be tested across different neighbourhood sizes, distance thresholds, and spatial topologies.

These experiments can reveal whether performance is robust or depends strongly on one graph assumption. Future studies should also report the node type, edge information, construction rule, weighting scheme, retained graph views, links across sources, and the role of the graph during learning. This would make it easier to separate the effects of graph construction, GNN operations, output modules, and training objectives.

6.2. Task-specific Benchmarks and Validation

Evaluation should reflect the output and interpretation of each task. Spatial representation requires more than clustering agreement and can include boundary quality, marker support, agreement with histology, and sensitivity to graph parameters. Communication and niche inference require validation independent of the priors used to define candidate interactions. Integration should assess both correspondence across datasets and preservation of biological structure. Deconvolution and expression prediction should test reference robustness, resolution limits, and uncertainty.

DLPFC and similar benchmarks are useful for spatial domain identification, but they should not serve as general benchmarks for all ST-GNN applications. Broader benchmark suites should use separate evaluation criteria for integration, communication and niche inference, imputation, deconvolution, and expression prediction from histology. They should also distinguish the quality of the primary model output from the reliability of any later biological analysis built on that output.

6.3. Scaling Graph Systems

The methods summarised in our comparison table range from a few hundred spots to millions of cells, but computational information is reported unevenly. Future work should report graph size, hardware, runtime, memory use, and any sampling or approximation used during training and inference.

These choices can also change the biological information available to the model. Subgraph training, sparsification, approximate neighbour search, and partitioning may remove rare cell states, sharp boundaries, weak interactions, or distant correspondences. Their effects should therefore be evaluated rather than treated only as implementation details. Novae [32], for example, uses dynamically generated local subgraphs and lazy loading to operate on millions of cells. Such strategies improve efficiency, but they also determine which local structures are presented to the model.

Systems with several entity types create an additional challenge. HEIST [30] jointly models spatial connections among cells and co-expression among genes or proteins. Similar systems may combine spots, cells, genes, image patches, reference units, or tissue regions. Each node type, edge type, and output scale should be stated explicitly. The contribution of additional modalities should also be tested separately to determine whether they add independent information or mainly repeat spatial structure already present in other inputs.

Graph learning may also enter earlier stages of the pipeline. Segger [70], for example, uses graph learning during segmentation and therefore affects which cells later become graph nodes. This dependency should be made explicit because errors in node definition can propagate through all later stages.

6.4. Expanding Biological Connections

Future models may include regulatory links, pathway dependencies, cell state transitions, physical contacts, extracellular matrix interactions, paracrine gradients, and causal links supported by perturbation experiments. These connections should not be treated as interchangeable edges because they represent different biological mechanisms and support different claims.

Regulatory and gene-level priors are still less common than ligand–receptor pairs and pathway annotations. Future methods may incorporate gene regulatory networks, protein interactions, perturbation data, literature knowledge, or clinical information. Each source should be assigned a clear role in the graph. It may define candidate edges, assign edge weights, provide node features, guide supervision, or serve as validation evidence.

The main challenge is not simply to build larger or more complex graphs. It is to represent richer biological connections while keeping their meaning, scale, source, and validation requirements explicit.

7. Conclusions

This review organised ST-GNN methods as a sequence from spatial inputs and graph construction to graph-based learning and downstream applications. This view shows that methods with similar neural architectures can encode different analytical assumptions because they define different nodes, relations, graph structures, learning roles, and outputs. The graph should therefore be examined as part of the model specification rather than treated as a neutral preprocessing step.

Across application areas, the suitability of an ST-GNN method depends on whether its graph representation matches the biological scale, relation, and prediction required by the task. Comparisons based only on model architecture or aggregate performance obscure these differences. A component-level description of the input, graph, learning process, output, and evaluation provides a clearer basis for comparing methods and interpreting their results.

Table 3. Method-level catalogue of GNN methods for spatial transcriptomics. Each entry summarises the main task, data type, additional inputs, graph construction, GNN backbone, approximate data scale, computational scale, and code availability. Scale records quantities stated in the original study, including numbers of samples, slices, spots, cells, or genes. Computational scale records explicitly stated hardware, runtime, memory use, or strategies for processing larger graphs. Only information reported in the original study is shown.

Method	Year	Main task	Data	Additional input	Graph construction	GNN backbone	Scale	Computational scale	Code
GCNG	2020	cell-cell communication inference	seqFISH+; MERFISH	-	Spatial radius	GCN	913 cells; 2,050 cells across 7 fields of view; 1,368 cells; 10,000 genes		url [28]
SpaGCN	2021	spatial domain identification; spatially variable gene detection	ST; 10x Visium; Slide-seqV2; STARmap; MERFISH	optional histology	Histology-weighted spatial	GCN	224 spots; 12 tissue slices; 3,353 spots; 16,448 genes	single Intel Core i5-8259U CPU @ 2.30 GHz; 16 GB memory; For mouse posterior brain data, SpaGCN completed spatial domain and SVG detection in less than 1 min,required 1.3 GB memory	url [34]
DSTG	2021	deconvolution	10x Visium; Slide-seqV2; ST	scRNA-seq reference; pseudo-ST	Canonical correlation analysis/mutual nearest neighbor similarity	GCN	2–8 cells per pseudo-spot; 500–4,000 spots; 5,000–50,000 reads per cell; 2,000 variable genes		url [25]
SCAN-IT	2021	spatial domain identification	osmFISH; seqFISH; seqFISH+; Visium; Slide-seq	-	Alpha-complex spatial	GCN	5,328 cells; 1,597 cells; 523 cells; 33 genes		url [33]
STAGATE	2022	spatial domain identification	10x Visium; Slide-seq; Slide-seqV2; Stereo-seq	-	Spatial neighbor; optional pruning	GAT	12 sections; 3498–4789 spots per section; 19,109 spots; 3000 highly variable genes	Largest real dataset with >50k spots, approx. 40 min; simulated dataset with 50k spots, less than 40 min, approx. 4 GB GPU memory; local subgraphs or mini-batches	url [44]
CCST	2022	spatial domain identification; cell clustering	MERFISH; seqFISH+; 10x Visium	-	Hybrid spatial adjacency	GCN			url [57]
SpaceFlow	2022	Representation; spatial domain identification; pseudo-space map	10x Visium; Stereo-seq; Slide-seqV2; seqFISH	-	Spatial-expression graph	GCN	12 sections; 18,197 cells; 12 tissue sections; 28,243 genes	GeForce RTX 2080 Ti GPU; ST data with fewer than 10,000 cells, less than 5 min on GPU; 3,000–50,000 cells/spots, 30 s to 3 min on GeForce RTX 2080 Ti GPU	url [71]
conST	2022	Representation; spatial domain identification	MERFISH; seqFISH; 10x Visium; Stereo-seq	optional histology	Spatial kNN	GCN	12 slices		url [58]
DeepLinc	2022	cell-cell communication inference	seqFISH; MERFISH; HDST	-	Spatial kNN	GCN	1,597 cells; 4,975 cells; 125 genes		url [43]
stMVC	2022	spatial domain identification; tumor heterogeneity	10x Visium; STARmap	Histology; region segmentation	Multi-view histology-similarity graph; spatial-location graph	GAT	12 slices; 3460–4789 spots per slice; 3844 spots; 2000 highly variable genes	20K spots, 38 min; local subgraphs or mini-batches, sampling	url [37]
Hist2ST	2022	histology-to-expression prediction	ST; 10x Visium; Space-TREX; Stereo-seq	Histology patches	Spatial kNN	GraphSAGE	32 sections from 7 patients; 9612 spots; 12 sections from 4 patients; 785 genes	Ubuntu 18.04.7 LTS, Intel Core i7-8700K CPU @ 3.70 GHz, 256 GB memory	url [54]
SD2	2022	deconvolution	seqFISH+; MERFISH; 10x Visium; ST	scRNA-seq reference	Transcriptional similarity; spatial	GCN	523 cells; 6 cell; 2, 4.5 and 15.7 cells per spot; 135 genes		url [72]
DeepST	2022	spatial domain identification; integration	10x Visium; Slide-seqV2; Stereo-seq; MERFISH; 4i; MIBI-TOF	optional histology; molecular features	Spatial kNN; expression/morphology features	GNN	12 slides; 10 cell; 3 consecutive imaging-based molecular slides; 33,538 genes	approx. 4000 spots and 30,000 genes, approx. 7 min on GPU, approx. 6G memory	url [73]
Spatial-ID	2022	cell type annotation; label transfer	MERFISH; Slide-seq; CosMx SMI; Stereo-seq	scRNA-seq reference labels	Distance-weighted Spatial neighbor	GCN	12 samples; 280,186 cells; 159,738 cells; 254 genes	Workstation with 40 GB RAM, 10 cores of 2.5 GHz Intel Xeon Platinum 8255C CPU, Nvidia Tesla T4 GPU with 8 GB memory; Mouse spermatogenesis, average time cost per sample, 5.0 min	url [35]
NCEM	2022	cell-cell communication inference	MERFISH; CODEX; MIBI-TOF; MELC; chip cytometry; 10x Visium	cell-type labels; spot deconvolution	Radius-based spatial cell or spot graph	GNN	40,864 WT cells; 284,098 cells; 11,321 cells; 132 genes		url [74]
GraphST	2023	spatial domain identification; integration; deconvolution	10x Visium; Stereo-seq; Slide-seqV2	optional scRNA-seq reference	Spatial neighborhood	GCN	12 slices; 3639 spots; 78,886 cells; 33,538 genes	Intel Core i7-8665U CPU, NVIDIA RTX A6000 GPU; E14.5 mouse embryo, approx. 100,000 spots, 30 min wall-clock time	url [21]
HoloNet	2023	cell-cell communication inference; functional cell effect decoding	10x Visium; Slide-seqV2	cell-type labels; ligand–receptor prior	Directed ligand–receptor multi-view	Attention GNN	2, 5000 spots; 586 highly variable genes		url [75]
STAligner	2023	integration; spatial domain identification	10x Visium; Stereo-seq; Slide-seqV2; Slide-seq	multiple slices	Spatial neighbor; triplet-based cross-slice alignment	GAT	4 adjacent DLPFC slices; 12 DLPFC slices; 4 mouse embryo slices	Runs on GPU or CPU; GPU is recommended	url [39]
SLAT	2023	integration	10x Visium; MERFISH; Stereo-seq; seqFISH; Xenium; spatial-ATAC-seq	multi-omics; multiple slices	Per-slice spatial kNN	GCN	3000 spots;  6500 cells;  100,000 cells; 20,000 genes	16 cores of Intel Xeon Platinum 8358 CPU, 128 GB RAM, NVIDIA A100 GPU with 80 GB VRAM; Slices with over 100,000 cells each, 3 min	url [40]
STGNNks	2023	spatial domain identification; cell clustering	10x Visium	-	Hybrid spatial adjacency	GCN			url [76]
SCGDL	2023	spatial domain identification	10x Visium; Slide-seqV2	-	Spatial neighbor	Gated GCN	12 samples; 3798 spots; 150,392 reads per spot; 3000 highly variable genes	Ubuntu 20.04.5 LTS, Intel Core i9-12900F CPU @ 2.40 GHz, 64 GB memory, GeForce RTX 3090Ti GPU	url [77]
SiGra	2023	spatial domain identification; expression enhancement	CosMx SMI; MERSCOPE; 10x Visium	multichannel images	Spatial proximity cell/spot	Graph Transformer	83,621 cells; 395,215 cells; 12 slices; 982 genes		url [51]
Spatial-MGCN	2023	spatial domain identification; gene imputation	10x Visium; Stereo-seq	-	Expression kNN; spatial radius	GCN	3460–4789 spots; 12 slices; 19,109 spots; 33,538 genes		url [49]
ConSpaS	2023	spatial domain identification; denoising	10x Visium; Stereo-seq; Slide-seqV2	-	Spatial neighbor; radius/kNN	GCN	12 slices		url [78]
spaCI	2023	cell-cell communication inference	seqFISH+; CosMx SMI; MERSCOPE	candidate ligand–receptor pairs	Spatial kNN; adaptive edge attention	Attention GNN	523 cells; 81,236 cells; 18 cell; 10,000 genes		url [79]
CGCom	2023	cell-cell communication inference	seqFISH	cell-type labels; ligand–receptor/downstream prior	Directed spatial threshold	GAT	17,806 cells; 14,185 cells; 20,577 cells; 29,452 genes		url [80]
SEDR	2024	Representation; spatial domain identification; integration	10x Visium; Stereo-seq; Slide-seqV2	-	Spatial kNN over spots/cells	GCN	12 sections; 3460–4789 spots per section; 3493 spots; 27,106 genes		url [31]
SSGCN	2024	spatial domain identification	10x Visium; Stereo-seq; Slide-seqV2	-	Spatial kNN with expression similarity edge reweighting	GCN	2,695 spots; 3,798 spots; 21,724 cells; 3,000 highly variable genes	Intel Xeon Gold 5220 CPU at 2.20 GHz; approx. 1 min for each Visium dataset and approx. 8 min for each high-resolution mouse olfactory bulb dataset	-  [81]
SGCAST	2024	spatial domain identification	10x Visium; Stereo-seq; Seq-scope	-	Mini-batch expression; spatial proximity	GCN	12 slides; 28,399 spots; 51,335 spots	Datasets with 3k–121k spots,  0.25 GB GPU memory; all datasets, less than 20 min; local subgraphs or mini-batches	url [82]
SpaGCAC	2024	spatial domain identification	10x Visium; STARmap; Stereo-seq	-	Spatial kNN with adaptive self-loop	GCN			url [83]
SGAE	2024	spatial domain identification	10x Visium; seqFISH; MERFISH; Slide-seqV2; Stereo-seq	-	Spatial kNN; type-aware pruning; distortion	GNN	12 continuous slides; 19,416 cells; 3,106 cells; 351 genes		url [45]
stAA	2024	spatial domain identification	10x Visium; Slide-seqV2; Stereo-seq; STARmap	-	Spatial neighbor	GCN	12 slices; 3460–4789 spots per slice; 3798 spots; 10,725–21,897 genes		url [61]
stMMR	2024	spatial domain identification	10x Visium; ST; NanoString CosMx SMI	Histology	Weighted spatial	GCN	12 sections; 8 major cell; 4 slices		url [84]
stImpute	2024	imputation	STARmap; osmFISH; MERFISH; Xenium; paired scRNA-seq references	scRNA-seq reference; ESM-2 embeddings	ESM-2 protein-similarity gene graph	GraphSAGE	1549 spatial cells; 14,249 cells; 3405 spatial cells; 1020 genes	Ubuntu 18.04.7 LTS, Intel Core i7-8700K CPU @ 3.70 GHz, 256 GB RAM, 2 NVIDIA GeForce RTX 4090 GPUs; 1.3 million cells, 961 min runtime, 13.8 GB memory	url [85]
MuCoST	2024	spatial domain identification	10x Visium; Stereo-seq	-	Spatial adjacency; co-expression; shuffled	GCN	12 slices; 3000 highly variable genes		url [36]
stKeep	2024	tumor microenvironment; cell-cell communication inference	10x Visium; NanoString CosMx SMI	Histology/region labels; GRN/PPI/LRP priors	Heterogeneous cell-gene-region; semantic cell; LRP CCC graphs	GNN	12 slices; 3460–4789 spots per slice; 4727 spots; 3000 highly variable genes	17K cells, 24 min, 13 GB memory	url [18]
STdGCN	2024	deconvolution	seqFISH; seqFISH+; MERFISH; Slide-seq; 10x Visium	scRNA-seq reference	Expression mutual nearest neighbor real/pseudo graph; spatial distance	GCN	109 spots; 6 cell; 3111 spots	Two Intel Xeon CPU E5-2650 v4 @ 2.20 GHz, 24 cores total, 516 GB memory, Tesla V100S-PCIE-32GB GPU, CentOS Linux 7	url [26]
STGAT	2024	deconvolution	Simulated ST; STARmap-derived data; 10x Visium	scRNA-seq reference; pseudo-ST	Canonical correlation analysis-mutual nearest neighbor combined pseudo-real; real-real; pseudo-pseudo graphs	GAT		sampling	-  [86]
EGGN	2024	gene expression prediction	STNet breast cancer; 10x Visium spatial proteomics	Histology windows; measured exemplars	Window-exemplar heterogeneous graph	GraphSAGE		NVIDIA Tesla P100 GPUs; sampling	url [87]
SpaGIC	2024	spatial domain identification	10x Visium; Stereo-seq; Slide-seqV2; STARmap; osmFISH	-	Spatial kNN with edge/local structural MI	GCN	12 slices; 3498–4789 spots per slice; 3798 cells; 33,538 genes		url [88]
Graspot	2024	integration	10x Visium DLPFC; 10x Visium mouse brain; HER2 breast cancer ST; human heart developmental ST	multiple slices	Within-slice spatial; cross-slice UOT	GAT	12 slices; 3 samples; 4 slices per sample		url [41]
asGNN	2024	gene expression prediction	ST1K breast cancer cohorts	H&E patches	Spatial kNN; adaptive pruning	GTN	68 tissue sections; 23 patients; 1007 capturing spots per section; 26,949 genes		url [46]
GTAD	2024	deconvolution	10x Visium; Slide-seqV2; ST	scRNA-seq	Weighted pseudo/real ST graph	GAT	2–8 cells per spot; 2500–12,500 pseudo-spots		url [64]
SpatialcoGCN	2024	deconvolution; imputation	MERFISH; Stereo-seq; ST; ISS	scRNA-seq	Cell-type/spot mutual nearest neighbor co-embedding	GCN	5–15 cells per spot; 4–15 cells per spot; 8, 953 spots; 250 genes		url [89]
stMCDI	2024	imputation	ST; 10x Visium; MERFISH; seqFISH	masked values	Spatial kNN from coordinates	GCN	121891, 6225 spots; 176078, 4785 spots; 142489, 2278 spots; 14,192–28,601, 1000 gene	1 NVIDIA RTX 4090 GPU, 24 GB	url [63]
Impeller	2024	imputation	10x Visium; Stereo-seq; Slide-seqV2	-	Heterogeneous spatial; expression paths	GNN	12 samples; 3460–4789 cells; 3449–4700 cells; 33,538 genes		url [62]
THItoGene	2024	histology-to-expression prediction	10x Visium	H&E patches	Spatial kNN	GAT	32 sections from 8 patients; 9612 spots; 12 tissue sections; 785 genes		url [55]
DeepTalk	2024	cell-cell communication inference	MERFISH; 10x Visium; ST	sc/snRNA-seq; ligand–receptor prior	Spatial cell; subgraph construction	GAT	1549 cells; 15 cell; 189 spots; 268 genes		url [24]
STAGUE	2024	spatial domain identification; cell-cell interaction inference	10x Visium; STARmap; MERFISH	-	Learned adjacency via graph structure learning	GNN	1207 cells; 3 sections; 1050 cells per section; 1020 genes	Ubuntu 20.04.6 server, 64 GB RAM, single NVIDIA RTX 4090 GPU with 24 GB memory; Large-scale MERFISH benchmark, 87,305 cells, less than 2 min on 64 GB CPU memory and 24 GB GPU memory	url [47]
stHGC	2025	spatial domain identification	10x Visium; Stereo-seq; Slide-seqV2	-	Spatial proximity; expression similarity	GAT	12 slices; 3460–3789 spots; 1 slice; 33,538 genes		url [90]
HGNN	2025	spatial domain identification	Mouse brain ST	Histology patches	Hypergraph; overlapping dense subgraphs	HGNN			url [52]
SpaGRA	2025	spatial domain identification	10x Visium; MERFISH; BaristaSeq; Stereo-seq; Visium HD	-	spatial distance with GAT augmentation	GAT	3798 spots; 5 adjacent slices; 5913 spots; 33,601 genes		url [91]
SpaInGNN	2025	spatial domain identification; integration	10x Visium; Stereo-seq; Slide-seqV2	Histology	Weighted spatial; histology-aware pruning	GCN	12 DLPFC sections; 3,498–4,789 spots per section; 3,000 highly variable genes		-  [92]
CellNEST	2025	cell-cell communication inference	Visium; MERFISH; Visium HD	Ligand–receptor database	Proximity cell/spot; ligand–receptor coexpression	GAT	4035 spots; 4095 spots; 24,068 cells		url [23]
MGGNN	2025	spatial domain identification	10x Visium DLPFC; coronal mouse brain ST	marker-guided labels	Spatial kNN	GCN	12 samples; 7 samples; 151673, sample		url [93]
STING	2025	spatial domain identification	10x Visium; STARmap; BaristaSeq; osmFISH; MERFISH	-	Nested spot; gene-gene coexpression	GATv2; GCN	45 samples; 4 samples; 12 slices; 5 spatial transcript	24 GB GPU memory mentioned; One 10x Visium human DLPFC sample, 4789 spots, 3000 genes, approx. 23 GB memory, approx. 6 min for 600 epochs	url [29]
MManiST	2025	spatial domain identification	10x Visium; osmFISH; STARmap; MERFISH; BaristaSeq	-	Spatial kNN	GCN	12 sections	NVIDIA RTX 3090 GPU; 3500 epochs, approx. 5 min, approx. 8 GB GPU memory	url [53]
STMIGCL	2025	spatial domain identification	10x Visium; Stereo-seq; STARmap	-	Spatial radius; expression cosine kNN	GCN	12 slices; 5913 cells; 23,015 genes	High-performance CPU/CUDA server, AMD Radeon Graphics CPU, NVIDIA-SMI 510.73.05 GPU environment	url [94]
STMGAMF	2025	spatial domain identification	10x Visium; Stereo-seq	-	Spatial radius; expression kNN	GCN	12 slices; 3000 highly variable genes		url [48]
stGRL	2025	spatial domain identification; imputation	10x Visium; Slide-seqV2; Stereo-seq	-	Spatial kNN	GCN	12 sections; 3460–4789 spots per section; 1 section; 33,538 genes	Tesla A100 GPU with 80 GB memory	url [95]
SpaMask	2025	spatial domain identification; integration	10x Visium; ST; Stereo-seq; osmFISH; MERFISH	-	Spatial kNN with node/edge masking	GNN	12 sections; 5 slices; 8 domains per slice; 33 genes	NVIDIA GeForce RTX 3090	url [59]
3D-spaGNN-E	2025	cell-cell communication inference	3D seqFISH-HCR; 3D MERFISH	-	Subcellular patch kNN	GNN	up to 100 optical slices; 844 CD4+ T cells; 636 CD8+ T cells; 156 genes		url [96]
STAIG	2025	spatial domain identification; integration	10x Visium; Stereo-seq; Slide-seqV2; STARmap; MERFISH	Histology	Spatial; image-guided augmentation	GNN	12 slices; 2 sections		url [22]
GRASS	2025	integration; alignment	multi-slice 10x Visium; 10x Xenium; Stereo-seq; Slide-seqV2; ST	-	Intra-slice Spatial graph; inter-slice correspondence links	GNN	12 slices; 4992 spots	NVIDIA GeForce RTX 4090 GPU, 24 GB; Intel Core i7-12700K CPU @ 3.60 GHz	url [97]
SpaICL	2025	spatial domain identification; imputation	10x Visium; Visium-like	Histology	Spatial kNN	GCN	12 slices; 3460–4789 spots; 1 slice; 33,538 genes		url [98]
Tacos	2025	integration; imputation	multi-slice 10x Visium; Slide-seqV2; Stereo-seq; seqFISH; Xenium	-	Alpha-complex; community augmentation; mutual nearest neighbor anchors	GCN	3400–4800 spots; 20,139 spots; 19,109 spots	NVIDIA RTX 3090 GPU	url [42]
SAINT	2025	spatial domain identification	10x Visium; sequence-augmented ST	sequence features	Spatial; expression feature; combined	GCN	12 sections; approx. 3600–4000 barcoded spots per section; 2695 spots; over 33,000 genes	Intel Core i9-9900K CPU, 64 GB RAM, NVIDIA RTX 3090 Ti GPU	-  [99]
SPICEiST	2025	cell clustering	Xenium; CosMx SMI	-	Per-cell subcellular grid	GCN	268,072 cells; 275,556 cells; 630,998 cells; 32,073,729 transcripts		-  [100]
STCase	2025	cell-cell communication inference; niche analysis	10x Visium; Slide-seq; Stereo-seq	Ligand–receptor prior	Spatial neighbor; ligand–receptor communication views	Attention GNN	3,000 highly variable genes	Supports GPU and CPU execution; Public example reports approx. 1 hour with GPU and approx. 20 hours without GPU	url [101]
RGAST	2025	cell-cell communication inference; spatial domain identification	MERFISH; HDST; 10x Visium; Stereo-seq; seqFISH+	-	Heterogeneous spatial; expression relations	Relational GAT	approx. 1 million cells; 6 consecutive slices; 4787–5926 spots per section; 155 genes		url [50]
OrgaCCC	2025	cell-cell communication inference	STARmap; MERFISH; seqFISH+; 10x Visium	Ligand–receptor prior	Cell/spot spatial; gene ligand–receptor graph	GNN	7224 cells; 4975 cells; 523 cells; 903 genes		url [102]
MAGNET	2025	cell-cell communication inference	seqFISH; MERFISH; STARmap; 10x Visium	Ligand–receptor prior; regulatory network	Multi-view cell graphs; gene regulatory graph	Attention GNN	100–2000 cells; 20–45 genes per cell		url [19]
STMGraph	2025	spatial domain identification; integration	10x Visium; Slide-seqV2; Stereo-seq; STARmap	-	Spatial neighbor; 2D/3D spatial neighbor graph	GAT	12 slices; 151673–151676, 4 slices; 3000 highly variable genes		url [103]
STCGAN	2025	deconvolution	seqFISH+; MERFISH; 10x Visium	scRNA-seq	Spatial kNN	GATv2	71 spots; 3067 spots; 10,000 genes		url [104]
CLPLS	2025	deconvolution	10x Visium; Slide-seqV2; Stereo-seq	scRNA-seq; optional scATAC	Alpha-complex ST; shared nearest neighbor single-cell	GCN	490 spots; 10,260 cells; 3639 spots		url [105]
stGuide	2025	label transfer; annotation transfer	10x Visium; MERFISH; STARmap	reference labels	Intra-slice spatial; inter-slice mutual nearest neighbor	GAT	12 slices		-  [60]
STAIR	2025	integration; alignment; 3D reconstruction	10x Visium; MERFISH; ST; Stereo-seq; Slide-seqV2; seqFISH	-	Intra-slice spatial; inter-slice expression	GAT	12 slices; 24 slices; 40 coronal half-brain slices; 342 genes		url [106]
HiSTaR	2025	spatial domain identification; integration	10x Visium; Stereo-seq; Slide-seqV2; STARmap	-	Spatial kNN	GCN	12 slices; 3798 spots; 191,091 spots		url [107]
HEIST	2025	foundation model; multi-task	Xenium; MERFISH; CODEX; MIBI	protein counts; cell-type labels; Leiden groups	Hierarchical cell-cell; cell-type gene/protein	Graph Transformer	124 tissues; 1 slice; 100,000 cells	40 GB NVIDIA L40 GPU used for block-size setting	url [30]
MERGE	2025	histology-to-expression prediction	ST; 10x Visium	Whole-slide image patches	Hierarchical patch graph; spatial/feature/shortcut edges	GAT	68 samples from 23 patients; approx. 450 spots per sample; 36 samples from 8 patients; 250 genes	NVIDIA RTX A6000 GPU	url [108]
scstGCN	2025	super-resolution; enhancement	Xenium; Visium HD; 10x Visium; ST	H&E image	Superpixel-level Spatial neighbor	GCN	1 section; 3 sections; 32 sections; 17,797 genes		url [27]
NicheCompass	2025	cell-cell communication inference; integration	cell/spot-level spatial omics; spatial multi-omics	communication prior	Spatial neighborhood	GNN	8.4 million cells; 3 embryo tissues; 11 cell	NVIDIA A100-PCIE-40 GB GPU	url [56]
Novae	2025	foundation model; spatial domain identification	Xenium; MERSCOPE; CosMx	-	Delaunay spatial graph	GAT	78 slides; approx. 29–30 million cells; 18 tissues	NVIDIA A100 GPU used for time comparisons; large-scale training demonstrated with 40 GB GPU memory; Uses lazy loading and dynamically generated local graph mini-…	url [32]
SpaGT	2025	spatial domain identification; expression denoising	10x Visium; Slide-seqV2; Stereo-seq; osmFISH; Seq-Scope; ST	-	Spatial graph; structure-reinforced self-attention (evolving topology)	Graph Transformer	12 sections	local subgraphs or mini-batches	url [109]
HAST	2026	spatial domain identification	10x Visium; ST; Visium HD	Histology	Local hypergraphs; gene; morphology; spatial	Hypergraph CNN	12 slices; 3460–4789 spots; 2 sections; 33,538 genes	GPU devices with CUDA support; local subgraphs or mini-batches, graph sparsification	url [38]
stGCL	2026	spatial domain identification; integration	10x Visium; CosMx SMI; MERFISH; Stereo-seq; Slide-seqV2; STARmap; ST; 10x Xenium	Histology	Spatial neighborhood; coordinate alignment	GAT	12 slices; 151674 with 3673 spots; 4 slices; 19,856 genes	Intel Xeon Gold 6258R CPU @ 2.70 GHz, NVIDIA Quadro GV100 GPU; Breast cancer 10x Xenium dataset, over 167k cells, approx. 14 min runtime, 26.9 GB GPU memory	url [66]
DACN	2026	spatial domain identification	10x Visium; BaristaSeq; osmFISH	-	Spatial graph from coordinates	GCN	12 slices; 3 tissue sections		url [65]
STransfer	2026	spatial domain identification; domain adaptation	10x Visium; MERFISH; STARmap; CosMx; Xenium; Stereo-seq; Visium HD	source labels	Spatial kNN; positive pointwise mutual information graph	GCN	12 slices; 3460–4789 spots per slice; 5 slices; 33,538 genes		url [110]
SPICE	2026	cell-cell communication inference	MERFISH; Xenium	optional cell-type labels	Radius-based spatial cell	GMMConv; GCN	approx. 1 million cells; 181 unique tissues; 161 genes		url [111]

Table 3. Method-level catalogue of GNN methods for spatial transcriptomics. Each entry summarises the main task, data type, additional inputs, graph construction, GNN backbone, approximate data scale, computational scale, and code availability. Scale records quantities stated in the original study, including numbers of samples, slices, spots, cells, or genes. Computational scale records explicitly stated hardware, runtime, memory use, or strategies for processing larger graphs. Only information reported in the original study is shown.

Method	Year	Main task	Data	Additional input	Graph construction	GNN backbone	Scale	Computational scale	Code
GCNG	2020	cell-cell communication inference	seqFISH+; MERFISH	-	Spatial radius	GCN	913 cells; 2,050 cells across 7 fields of view; 1,368 cells; 10,000 genes		url [28]
SpaGCN	2021	spatial domain identification; spatially variable gene detection	ST; 10x Visium; Slide-seqV2; STARmap; MERFISH	optional histology	Histology-weighted spatial	GCN	224 spots; 12 tissue slices; 3,353 spots; 16,448 genes	single Intel Core i5-8259U CPU @ 2.30 GHz; 16 GB memory; For mouse posterior brain data, SpaGCN completed spatial domain and SVG detection in less than 1 min,required 1.3 GB memory	url [34]
DSTG	2021	deconvolution	10x Visium; Slide-seqV2; ST	scRNA-seq reference; pseudo-ST	Canonical correlation analysis/mutual nearest neighbor similarity	GCN	2–8 cells per pseudo-spot; 500–4,000 spots; 5,000–50,000 reads per cell; 2,000 variable genes		url [25]
SCAN-IT	2021	spatial domain identification	osmFISH; seqFISH; seqFISH+; Visium; Slide-seq	-	Alpha-complex spatial	GCN	5,328 cells; 1,597 cells; 523 cells; 33 genes		url [33]
STAGATE	2022	spatial domain identification	10x Visium; Slide-seq; Slide-seqV2; Stereo-seq	-	Spatial neighbor; optional pruning	GAT	12 sections; 3498–4789 spots per section; 19,109 spots; 3000 highly variable genes	Largest real dataset with >50k spots, approx. 40 min; simulated dataset with 50k spots, less than 40 min, approx. 4 GB GPU memory; local subgraphs or mini-batches	url [44]
CCST	2022	spatial domain identification; cell clustering	MERFISH; seqFISH+; 10x Visium	-	Hybrid spatial adjacency	GCN			url [57]
SpaceFlow	2022	Representation; spatial domain identification; pseudo-space map	10x Visium; Stereo-seq; Slide-seqV2; seqFISH	-	Spatial-expression graph	GCN	12 sections; 18,197 cells; 12 tissue sections; 28,243 genes	GeForce RTX 2080 Ti GPU; ST data with fewer than 10,000 cells, less than 5 min on GPU; 3,000–50,000 cells/spots, 30 s to 3 min on GeForce RTX 2080 Ti GPU	url [71]
conST	2022	Representation; spatial domain identification	MERFISH; seqFISH; 10x Visium; Stereo-seq	optional histology	Spatial kNN	GCN	12 slices		url [58]
DeepLinc	2022	cell-cell communication inference	seqFISH; MERFISH; HDST	-	Spatial kNN	GCN	1,597 cells; 4,975 cells; 125 genes		url [43]
stMVC	2022	spatial domain identification; tumor heterogeneity	10x Visium; STARmap	Histology; region segmentation	Multi-view histology-similarity graph; spatial-location graph	GAT	12 slices; 3460–4789 spots per slice; 3844 spots; 2000 highly variable genes	20K spots, 38 min; local subgraphs or mini-batches, sampling	url [37]
Hist2ST	2022	histology-to-expression prediction	ST; 10x Visium; Space-TREX; Stereo-seq	Histology patches	Spatial kNN	GraphSAGE	32 sections from 7 patients; 9612 spots; 12 sections from 4 patients; 785 genes	Ubuntu 18.04.7 LTS, Intel Core i7-8700K CPU @ 3.70 GHz, 256 GB memory	url [54]
SD2	2022	deconvolution	seqFISH+; MERFISH; 10x Visium; ST	scRNA-seq reference	Transcriptional similarity; spatial	GCN	523 cells; 6 cell; 2, 4.5 and 15.7 cells per spot; 135 genes		url [72]
DeepST	2022	spatial domain identification; integration	10x Visium; Slide-seqV2; Stereo-seq; MERFISH; 4i; MIBI-TOF	optional histology; molecular features	Spatial kNN; expression/morphology features	GNN	12 slides; 10 cell; 3 consecutive imaging-based molecular slides; 33,538 genes	approx. 4000 spots and 30,000 genes, approx. 7 min on GPU, approx. 6G memory	url [73]
Spatial-ID	2022	cell type annotation; label transfer	MERFISH; Slide-seq; CosMx SMI; Stereo-seq	scRNA-seq reference labels	Distance-weighted Spatial neighbor	GCN	12 samples; 280,186 cells; 159,738 cells; 254 genes	Workstation with 40 GB RAM, 10 cores of 2.5 GHz Intel Xeon Platinum 8255C CPU, Nvidia Tesla T4 GPU with 8 GB memory; Mouse spermatogenesis, average time cost per sample, 5.0 min	url [35]
NCEM	2022	cell-cell communication inference	MERFISH; CODEX; MIBI-TOF; MELC; chip cytometry; 10x Visium	cell-type labels; spot deconvolution	Radius-based spatial cell or spot graph	GNN	40,864 WT cells; 284,098 cells; 11,321 cells; 132 genes		url [74]
GraphST	2023	spatial domain identification; integration; deconvolution	10x Visium; Stereo-seq; Slide-seqV2	optional scRNA-seq reference	Spatial neighborhood	GCN	12 slices; 3639 spots; 78,886 cells; 33,538 genes	Intel Core i7-8665U CPU, NVIDIA RTX A6000 GPU; E14.5 mouse embryo, approx. 100,000 spots, 30 min wall-clock time	url [21]
HoloNet	2023	cell-cell communication inference; functional cell effect decoding	10x Visium; Slide-seqV2	cell-type labels; ligand–receptor prior	Directed ligand–receptor multi-view	Attention GNN	2, 5000 spots; 586 highly variable genes		url [75]
STAligner	2023	integration; spatial domain identification	10x Visium; Stereo-seq; Slide-seqV2; Slide-seq	multiple slices	Spatial neighbor; triplet-based cross-slice alignment	GAT	4 adjacent DLPFC slices; 12 DLPFC slices; 4 mouse embryo slices	Runs on GPU or CPU; GPU is recommended	url [39]
SLAT	2023	integration	10x Visium; MERFISH; Stereo-seq; seqFISH; Xenium; spatial-ATAC-seq	multi-omics; multiple slices	Per-slice spatial kNN	GCN	3000 spots;  6500 cells;  100,000 cells; 20,000 genes	16 cores of Intel Xeon Platinum 8358 CPU, 128 GB RAM, NVIDIA A100 GPU with 80 GB VRAM; Slices with over 100,000 cells each, 3 min	url [40]
STGNNks	2023	spatial domain identification; cell clustering	10x Visium	-	Hybrid spatial adjacency	GCN			url [76]
SCGDL	2023	spatial domain identification	10x Visium; Slide-seqV2	-	Spatial neighbor	Gated GCN	12 samples; 3798 spots; 150,392 reads per spot; 3000 highly variable genes	Ubuntu 20.04.5 LTS, Intel Core i9-12900F CPU @ 2.40 GHz, 64 GB memory, GeForce RTX 3090Ti GPU	url [77]
SiGra	2023	spatial domain identification; expression enhancement	CosMx SMI; MERSCOPE; 10x Visium	multichannel images	Spatial proximity cell/spot	Graph Transformer	83,621 cells; 395,215 cells; 12 slices; 982 genes		url [51]
Spatial-MGCN	2023	spatial domain identification; gene imputation	10x Visium; Stereo-seq	-	Expression kNN; spatial radius	GCN	3460–4789 spots; 12 slices; 19,109 spots; 33,538 genes		url [49]
ConSpaS	2023	spatial domain identification; denoising	10x Visium; Stereo-seq; Slide-seqV2	-	Spatial neighbor; radius/kNN	GCN	12 slices		url [78]
spaCI	2023	cell-cell communication inference	seqFISH+; CosMx SMI; MERSCOPE	candidate ligand–receptor pairs	Spatial kNN; adaptive edge attention	Attention GNN	523 cells; 81,236 cells; 18 cell; 10,000 genes		url [79]
CGCom	2023	cell-cell communication inference	seqFISH	cell-type labels; ligand–receptor/downstream prior	Directed spatial threshold	GAT	17,806 cells; 14,185 cells; 20,577 cells; 29,452 genes		url [80]
SEDR	2024	Representation; spatial domain identification; integration	10x Visium; Stereo-seq; Slide-seqV2	-	Spatial kNN over spots/cells	GCN	12 sections; 3460–4789 spots per section; 3493 spots; 27,106 genes		url [31]
SSGCN	2024	spatial domain identification	10x Visium; Stereo-seq; Slide-seqV2	-	Spatial kNN with expression similarity edge reweighting	GCN	2,695 spots; 3,798 spots; 21,724 cells; 3,000 highly variable genes	Intel Xeon Gold 5220 CPU at 2.20 GHz; approx. 1 min for each Visium dataset and approx. 8 min for each high-resolution mouse olfactory bulb dataset	-  [81]
SGCAST	2024	spatial domain identification	10x Visium; Stereo-seq; Seq-scope	-	Mini-batch expression; spatial proximity	GCN	12 slides; 28,399 spots; 51,335 spots	Datasets with 3k–121k spots,  0.25 GB GPU memory; all datasets, less than 20 min; local subgraphs or mini-batches	url [82]
SpaGCAC	2024	spatial domain identification	10x Visium; STARmap; Stereo-seq	-	Spatial kNN with adaptive self-loop	GCN			url [83]
SGAE	2024	spatial domain identification	10x Visium; seqFISH; MERFISH; Slide-seqV2; Stereo-seq	-	Spatial kNN; type-aware pruning; distortion	GNN	12 continuous slides; 19,416 cells; 3,106 cells; 351 genes		url [45]
stAA	2024	spatial domain identification	10x Visium; Slide-seqV2; Stereo-seq; STARmap	-	Spatial neighbor	GCN	12 slices; 3460–4789 spots per slice; 3798 spots; 10,725–21,897 genes		url [61]
stMMR	2024	spatial domain identification	10x Visium; ST; NanoString CosMx SMI	Histology	Weighted spatial	GCN	12 sections; 8 major cell; 4 slices		url [84]
stImpute	2024	imputation	STARmap; osmFISH; MERFISH; Xenium; paired scRNA-seq references	scRNA-seq reference; ESM-2 embeddings	ESM-2 protein-similarity gene graph	GraphSAGE	1549 spatial cells; 14,249 cells; 3405 spatial cells; 1020 genes	Ubuntu 18.04.7 LTS, Intel Core i7-8700K CPU @ 3.70 GHz, 256 GB RAM, 2 NVIDIA GeForce RTX 4090 GPUs; 1.3 million cells, 961 min runtime, 13.8 GB memory	url [85]
MuCoST	2024	spatial domain identification	10x Visium; Stereo-seq	-	Spatial adjacency; co-expression; shuffled	GCN	12 slices; 3000 highly variable genes		url [36]
stKeep	2024	tumor microenvironment; cell-cell communication inference	10x Visium; NanoString CosMx SMI	Histology/region labels; GRN/PPI/LRP priors	Heterogeneous cell-gene-region; semantic cell; LRP CCC graphs	GNN	12 slices; 3460–4789 spots per slice; 4727 spots; 3000 highly variable genes	17K cells, 24 min, 13 GB memory	url [18]
STdGCN	2024	deconvolution	seqFISH; seqFISH+; MERFISH; Slide-seq; 10x Visium	scRNA-seq reference	Expression mutual nearest neighbor real/pseudo graph; spatial distance	GCN	109 spots; 6 cell; 3111 spots	Two Intel Xeon CPU E5-2650 v4 @ 2.20 GHz, 24 cores total, 516 GB memory, Tesla V100S-PCIE-32GB GPU, CentOS Linux 7	url [26]
STGAT	2024	deconvolution	Simulated ST; STARmap-derived data; 10x Visium	scRNA-seq reference; pseudo-ST	Canonical correlation analysis-mutual nearest neighbor combined pseudo-real; real-real; pseudo-pseudo graphs	GAT		sampling	-  [86]
EGGN	2024	gene expression prediction	STNet breast cancer; 10x Visium spatial proteomics	Histology windows; measured exemplars	Window-exemplar heterogeneous graph	GraphSAGE		NVIDIA Tesla P100 GPUs; sampling	url [87]
SpaGIC	2024	spatial domain identification	10x Visium; Stereo-seq; Slide-seqV2; STARmap; osmFISH	-	Spatial kNN with edge/local structural MI	GCN	12 slices; 3498–4789 spots per slice; 3798 cells; 33,538 genes		url [88]
Graspot	2024	integration	10x Visium DLPFC; 10x Visium mouse brain; HER2 breast cancer ST; human heart developmental ST	multiple slices	Within-slice spatial; cross-slice UOT	GAT	12 slices; 3 samples; 4 slices per sample		url [41]
asGNN	2024	gene expression prediction	ST1K breast cancer cohorts	H&E patches	Spatial kNN; adaptive pruning	GTN	68 tissue sections; 23 patients; 1007 capturing spots per section; 26,949 genes		url [46]
GTAD	2024	deconvolution	10x Visium; Slide-seqV2; ST	scRNA-seq	Weighted pseudo/real ST graph	GAT	2–8 cells per spot; 2500–12,500 pseudo-spots		url [64]
SpatialcoGCN	2024	deconvolution; imputation	MERFISH; Stereo-seq; ST; ISS	scRNA-seq	Cell-type/spot mutual nearest neighbor co-embedding	GCN	5–15 cells per spot; 4–15 cells per spot; 8, 953 spots; 250 genes		url [89]
stMCDI	2024	imputation	ST; 10x Visium; MERFISH; seqFISH	masked values	Spatial kNN from coordinates	GCN	121891, 6225 spots; 176078, 4785 spots; 142489, 2278 spots; 14,192–28,601, 1000 gene	1 NVIDIA RTX 4090 GPU, 24 GB	url [63]
Impeller	2024	imputation	10x Visium; Stereo-seq; Slide-seqV2	-	Heterogeneous spatial; expression paths	GNN	12 samples; 3460–4789 cells; 3449–4700 cells; 33,538 genes		url [62]
THItoGene	2024	histology-to-expression prediction	10x Visium	H&E patches	Spatial kNN	GAT	32 sections from 8 patients; 9612 spots; 12 tissue sections; 785 genes		url [55]
DeepTalk	2024	cell-cell communication inference	MERFISH; 10x Visium; ST	sc/snRNA-seq; ligand–receptor prior	Spatial cell; subgraph construction	GAT	1549 cells; 15 cell; 189 spots; 268 genes		url [24]
STAGUE	2024	spatial domain identification; cell-cell interaction inference	10x Visium; STARmap; MERFISH	-	Learned adjacency via graph structure learning	GNN	1207 cells; 3 sections; 1050 cells per section; 1020 genes	Ubuntu 20.04.6 server, 64 GB RAM, single NVIDIA RTX 4090 GPU with 24 GB memory; Large-scale MERFISH benchmark, 87,305 cells, less than 2 min on 64 GB CPU memory and 24 GB GPU memory	url [47]
stHGC	2025	spatial domain identification	10x Visium; Stereo-seq; Slide-seqV2	-	Spatial proximity; expression similarity	GAT	12 slices; 3460–3789 spots; 1 slice; 33,538 genes		url [90]
HGNN	2025	spatial domain identification	Mouse brain ST	Histology patches	Hypergraph; overlapping dense subgraphs	HGNN			url [52]
SpaGRA	2025	spatial domain identification	10x Visium; MERFISH; BaristaSeq; Stereo-seq; Visium HD	-	spatial distance with GAT augmentation	GAT	3798 spots; 5 adjacent slices; 5913 spots; 33,601 genes		url [91]
SpaInGNN	2025	spatial domain identification; integration	10x Visium; Stereo-seq; Slide-seqV2	Histology	Weighted spatial; histology-aware pruning	GCN	12 DLPFC sections; 3,498–4,789 spots per section; 3,000 highly variable genes		-  [92]
CellNEST	2025	cell-cell communication inference	Visium; MERFISH; Visium HD	Ligand–receptor database	Proximity cell/spot; ligand–receptor coexpression	GAT	4035 spots; 4095 spots; 24,068 cells		url [23]
MGGNN	2025	spatial domain identification	10x Visium DLPFC; coronal mouse brain ST	marker-guided labels	Spatial kNN	GCN	12 samples; 7 samples; 151673, sample		url [93]
STING	2025	spatial domain identification	10x Visium; STARmap; BaristaSeq; osmFISH; MERFISH	-	Nested spot; gene-gene coexpression	GATv2; GCN	45 samples; 4 samples; 12 slices; 5 spatial transcript	24 GB GPU memory mentioned; One 10x Visium human DLPFC sample, 4789 spots, 3000 genes, approx. 23 GB memory, approx. 6 min for 600 epochs	url [29]
MManiST	2025	spatial domain identification	10x Visium; osmFISH; STARmap; MERFISH; BaristaSeq	-	Spatial kNN	GCN	12 sections	NVIDIA RTX 3090 GPU; 3500 epochs, approx. 5 min, approx. 8 GB GPU memory	url [53]
STMIGCL	2025	spatial domain identification	10x Visium; Stereo-seq; STARmap	-	Spatial radius; expression cosine kNN	GCN	12 slices; 5913 cells; 23,015 genes	High-performance CPU/CUDA server, AMD Radeon Graphics CPU, NVIDIA-SMI 510.73.05 GPU environment	url [94]
STMGAMF	2025	spatial domain identification	10x Visium; Stereo-seq	-	Spatial radius; expression kNN	GCN	12 slices; 3000 highly variable genes		url [48]
stGRL	2025	spatial domain identification; imputation	10x Visium; Slide-seqV2; Stereo-seq	-	Spatial kNN	GCN	12 sections; 3460–4789 spots per section; 1 section; 33,538 genes	Tesla A100 GPU with 80 GB memory	url [95]
SpaMask	2025	spatial domain identification; integration	10x Visium; ST; Stereo-seq; osmFISH; MERFISH	-	Spatial kNN with node/edge masking	GNN	12 sections; 5 slices; 8 domains per slice; 33 genes	NVIDIA GeForce RTX 3090	url [59]
3D-spaGNN-E	2025	cell-cell communication inference	3D seqFISH-HCR; 3D MERFISH	-	Subcellular patch kNN	GNN	up to 100 optical slices; 844 CD4+ T cells; 636 CD8+ T cells; 156 genes		url [96]
STAIG	2025	spatial domain identification; integration	10x Visium; Stereo-seq; Slide-seqV2; STARmap; MERFISH	Histology	Spatial; image-guided augmentation	GNN	12 slices; 2 sections		url [22]
GRASS	2025	integration; alignment	multi-slice 10x Visium; 10x Xenium; Stereo-seq; Slide-seqV2; ST	-	Intra-slice Spatial graph; inter-slice correspondence links	GNN	12 slices; 4992 spots	NVIDIA GeForce RTX 4090 GPU, 24 GB; Intel Core i7-12700K CPU @ 3.60 GHz	url [97]
SpaICL	2025	spatial domain identification; imputation	10x Visium; Visium-like	Histology	Spatial kNN	GCN	12 slices; 3460–4789 spots; 1 slice; 33,538 genes		url [98]
Tacos	2025	integration; imputation	multi-slice 10x Visium; Slide-seqV2; Stereo-seq; seqFISH; Xenium	-	Alpha-complex; community augmentation; mutual nearest neighbor anchors	GCN	3400–4800 spots; 20,139 spots; 19,109 spots	NVIDIA RTX 3090 GPU	url [42]
SAINT	2025	spatial domain identification	10x Visium; sequence-augmented ST	sequence features	Spatial; expression feature; combined	GCN	12 sections; approx. 3600–4000 barcoded spots per section; 2695 spots; over 33,000 genes	Intel Core i9-9900K CPU, 64 GB RAM, NVIDIA RTX 3090 Ti GPU	-  [99]
SPICEiST	2025	cell clustering	Xenium; CosMx SMI	-	Per-cell subcellular grid	GCN	268,072 cells; 275,556 cells; 630,998 cells; 32,073,729 transcripts		-  [100]
STCase	2025	cell-cell communication inference; niche analysis	10x Visium; Slide-seq; Stereo-seq	Ligand–receptor prior	Spatial neighbor; ligand–receptor communication views	Attention GNN	3,000 highly variable genes	Supports GPU and CPU execution; Public example reports approx. 1 hour with GPU and approx. 20 hours without GPU	url [101]
RGAST	2025	cell-cell communication inference; spatial domain identification	MERFISH; HDST; 10x Visium; Stereo-seq; seqFISH+	-	Heterogeneous spatial; expression relations	Relational GAT	approx. 1 million cells; 6 consecutive slices; 4787–5926 spots per section; 155 genes		url [50]
OrgaCCC	2025	cell-cell communication inference	STARmap; MERFISH; seqFISH+; 10x Visium	Ligand–receptor prior	Cell/spot spatial; gene ligand–receptor graph	GNN	7224 cells; 4975 cells; 523 cells; 903 genes		url [102]
MAGNET	2025	cell-cell communication inference	seqFISH; MERFISH; STARmap; 10x Visium	Ligand–receptor prior; regulatory network	Multi-view cell graphs; gene regulatory graph	Attention GNN	100–2000 cells; 20–45 genes per cell		url [19]
STMGraph	2025	spatial domain identification; integration	10x Visium; Slide-seqV2; Stereo-seq; STARmap	-	Spatial neighbor; 2D/3D spatial neighbor graph	GAT	12 slices; 151673–151676, 4 slices; 3000 highly variable genes		url [103]
STCGAN	2025	deconvolution	seqFISH+; MERFISH; 10x Visium	scRNA-seq	Spatial kNN	GATv2	71 spots; 3067 spots; 10,000 genes		url [104]
CLPLS	2025	deconvolution	10x Visium; Slide-seqV2; Stereo-seq	scRNA-seq; optional scATAC	Alpha-complex ST; shared nearest neighbor single-cell	GCN	490 spots; 10,260 cells; 3639 spots		url [105]
stGuide	2025	label transfer; annotation transfer	10x Visium; MERFISH; STARmap	reference labels	Intra-slice spatial; inter-slice mutual nearest neighbor	GAT	12 slices		-  [60]
STAIR	2025	integration; alignment; 3D reconstruction	10x Visium; MERFISH; ST; Stereo-seq; Slide-seqV2; seqFISH	-	Intra-slice spatial; inter-slice expression	GAT	12 slices; 24 slices; 40 coronal half-brain slices; 342 genes		url [106]
HiSTaR	2025	spatial domain identification; integration	10x Visium; Stereo-seq; Slide-seqV2; STARmap	-	Spatial kNN	GCN	12 slices; 3798 spots; 191,091 spots		url [107]
HEIST	2025	foundation model; multi-task	Xenium; MERFISH; CODEX; MIBI	protein counts; cell-type labels; Leiden groups	Hierarchical cell-cell; cell-type gene/protein	Graph Transformer	124 tissues; 1 slice; 100,000 cells	40 GB NVIDIA L40 GPU used for block-size setting	url [30]
MERGE	2025	histology-to-expression prediction	ST; 10x Visium	Whole-slide image patches	Hierarchical patch graph; spatial/feature/shortcut edges	GAT	68 samples from 23 patients; approx. 450 spots per sample; 36 samples from 8 patients; 250 genes	NVIDIA RTX A6000 GPU	url [108]
scstGCN	2025	super-resolution; enhancement	Xenium; Visium HD; 10x Visium; ST	H&E image	Superpixel-level Spatial neighbor	GCN	1 section; 3 sections; 32 sections; 17,797 genes		url [27]
NicheCompass	2025	cell-cell communication inference; integration	cell/spot-level spatial omics; spatial multi-omics	communication prior	Spatial neighborhood	GNN	8.4 million cells; 3 embryo tissues; 11 cell	NVIDIA A100-PCIE-40 GB GPU	url [56]
Novae	2025	foundation model; spatial domain identification	Xenium; MERSCOPE; CosMx	-	Delaunay spatial graph	GAT	78 slides; approx. 29–30 million cells; 18 tissues	NVIDIA A100 GPU used for time comparisons; large-scale training demonstrated with 40 GB GPU memory; Uses lazy loading and dynamically generated local graph mini-…	url [32]
SpaGT	2025	spatial domain identification; expression denoising	10x Visium; Slide-seqV2; Stereo-seq; osmFISH; Seq-Scope; ST	-	Spatial graph; structure-reinforced self-attention (evolving topology)	Graph Transformer	12 sections	local subgraphs or mini-batches	url [109]
HAST	2026	spatial domain identification	10x Visium; ST; Visium HD	Histology	Local hypergraphs; gene; morphology; spatial	Hypergraph CNN	12 slices; 3460–4789 spots; 2 sections; 33,538 genes	GPU devices with CUDA support; local subgraphs or mini-batches, graph sparsification	url [38]
stGCL	2026	spatial domain identification; integration	10x Visium; CosMx SMI; MERFISH; Stereo-seq; Slide-seqV2; STARmap; ST; 10x Xenium	Histology	Spatial neighborhood; coordinate alignment	GAT	12 slices; 151674 with 3673 spots; 4 slices; 19,856 genes	Intel Xeon Gold 6258R CPU @ 2.70 GHz, NVIDIA Quadro GV100 GPU; Breast cancer 10x Xenium dataset, over 167k cells, approx. 14 min runtime, 26.9 GB GPU memory	url [66]
DACN	2026	spatial domain identification	10x Visium; BaristaSeq; osmFISH	-	Spatial graph from coordinates	GCN	12 slices; 3 tissue sections		url [65]
STransfer	2026	spatial domain identification; domain adaptation	10x Visium; MERFISH; STARmap; CosMx; Xenium; Stereo-seq; Visium HD	source labels	Spatial kNN; positive pointwise mutual information graph	GCN	12 slices; 3460–4789 spots per slice; 5 slices; 33,538 genes		url [110]
SPICE	2026	cell-cell communication inference	MERFISH; Xenium	optional cell-type labels	Radius-based spatial cell	GMMConv; GCN	approx. 1 million cells; 181 unique tissues; 161 genes		url [111]

Future progress will depend on graph designs that are better aligned with the structure of ST measurements, more explicit about the assumptions encoded by different information sources, and evaluated against task-specific biological claims. These priorities apply across spatial domain analysis, integration, communication inference, expression modelling, deconvolution, and resolution enhancement, despite the distinct outputs of these tasks.

Key Points 

• We provide a graph-centred survey of GNN-based methods for spatial transcriptomics, covering 89 representative studies across graph design, learning, and downstream applications.
• We organise ST-GNN methods by graph construction, graph learning, message passing, model outputs, and training strategies.
• We cover ST-GNN applications from tissue structure analysis to communication inference, alignment, expression modelling, and cell-type deconvolution.
• We compare representative methods by data inputs, graph construction, learning objectives, model outputs, and ST tasks.