Spatial Transcriptomics Integration: Methods for Combining Data Across Platforms

Spatial transcriptomics integration refers to the computational process of combining gene expression data that retains spatial context from multiple tissue sections, technology platforms, or experimental batches into a unified analytical framework. For students, researchers, analysts, and life-science professionals, the practical goal is to identify cell types, spatial domains, and molecular interactions that are reproducible across samples while removing technical variation that can obscure biological signals. This article covers the core principles of data integration, the main methodological families, workflow decisions, quality controls, and common failure patterns, with an emphasis on concrete choices that affect downstream interpretation.

The Integration Problem in Spatial Transcriptomics

Spatial transcriptomics technologies generate gene expression profiles that retain the physical location of measurements within a tissue section. These platforms differ substantially in resolution, throughput, and the type of information they capture. Some methods measure transcripts at the level of individual cells, while others capture expression in small spatial bins that may contain multiple cells. The field has advanced rapidly, with new platforms enabling high-throughput analysis of transcript localization in diverse biological systems, as described in a review of spatial transcriptomics data and analytical methods published in Drug Discovery Today [9].

The integration problem arises because researchers rarely analyze a single tissue section in isolation. Typical studies combine data from multiple donors, disease states, time points, or technology platforms. Each dataset carries its own technical artifacts, including differences in capture efficiency, sequencing depth, tissue processing, and platform-specific biases. Without correction, these technical differences can dominate the biological signal and lead to incorrect conclusions about cell types, spatial domains, and gene expression patterns.

A second integration challenge involves combining spatial transcriptomics with single-cell RNA sequencing (scRNA-seq) data. Single-cell approaches provide deeper transcriptome coverage per cell but lose spatial context. Spatial approaches retain location but often capture fewer genes per measurement unit. The complementary nature of these data types makes their integration a central task in spatial transcriptomics analysis. A review in Briefings in Functional Genomics compiled 19 integration methods designed to combine scRNA-seq data with spatial transcriptomics data, classifying them into two main groups and emphasizing the use of highly variable genes for annotating different technologies [10].

The integration problem also extends beyond transcriptomics. Spatial multi-omics technologies can now capture transcriptomic, proteomic, and epigenomic information from the same tissue section. These modalities differ in distributional properties, data dimensionality, and noise levels, and the inherent sparsity and incompleteness of spatial information pose major challenges for data integration and modeling [11]. Deep learning-based algorithms have emerged as a prominent approach for constructing unified latent representations and achieving cross-modal fusion in this context [11].

Core Principles of Data Integration

Technical Variation Versus Biological Signal

The central principle of integration is distinguishing technical variation from biological signal. Technical variation includes batch effects introduced by sample processing dates, reagent lots, sequencing runs, and platform differences. Biological signal includes genuine differences in cell type composition, spatial organization, and gene expression programs between samples. Integration methods aim to remove the former while preserving the latter.

The challenge is that technical and biological variation are often confounded. For example, if all diseased samples are processed in one batch and all healthy samples in another, an integration method cannot fully separate batch effects from disease effects. This confounding is a fundamental limitation that no computational method can fully resolve. Researchers must design experiments with balanced batch assignments whenever possible.

Shared Latent Space Construction

Most integration methods work by projecting data from multiple sources into a shared latent space, a lower-dimensional representation where corresponding biological states from different datasets align. The latent space is constructed so that technical differences are minimized while biological differences are preserved. After integration, downstream analyses such as clustering, differential expression, and spatial domain identification are performed in this shared representation.

The choice of latent space construction method affects the resolution at which biological structures can be recovered. Methods that preserve local structure may excel at identifying fine-grained cell states, while methods that preserve global structure may be better at aligning large-scale tissue organization. The tradeoff between local and global structure preservation is a key consideration in method selection.

Spatial Context as an Integration Cue

Unlike scRNA-seq integration, spatial transcriptomics integration can use spatial context as an additional cue. Spatially adjacent measurements are more likely to belong to the same tissue domain or cell type. Methods such as GraphST exploit this by using graph neural networks with self-supervised contrastive learning, minimizing the embedding distance between spatially adjacent spots while maximizing distance between non-adjacent spots [19]. This approach learns informative and discriminative spot representations that incorporate both gene expression and spatial location.

The use of spatial context is particularly valuable for integrating multiple tissue slices. GraphST is described as the only method among those benchmarked that can jointly analyze multiple tissue slices in vertical or horizontal integration while correcting batch effects [19]. Vertical integration refers to aligning sections from the same tissue region across different samples, while horizontal integration refers to aligning sections from different regions or conditions.

At a Glance: Integration Method Families

Practical Workflow for Integration

Step 1: Define the Integration Goal

Before selecting a method, define the specific analytical question. Integration goals fall into several categories. Cell type annotation involves mapping spatial spots to known cell types from a single-cell reference. Spatial domain identification involves finding regions of the tissue with coherent gene expression programs. Multi-sample comparison involves identifying reproducible differences between conditions across multiple tissue sections. Cross-modal integration involves combining spatial transcriptomics with other molecular layers such as chromatin accessibility or protein abundance.

Each goal favors different methods. For cell type annotation, methods that anchor spatial data to scRNA-seq references are appropriate. For spatial domain identification, graph-based methods that use spatial adjacency are often superior. For multi-sample comparison, batch correction methods that preserve biological variation between conditions are essential.

Step 2: Assess Data Compatibility

Evaluate whether the datasets to be integrated are compatible. Key considerations include the gene overlap between platforms, the resolution of spatial measurements, and the number of samples per condition. Platforms with limited gene overlap require methods that can work with a reduced feature set. Resolution differences between platforms may require aggregation or deconvolution steps before integration.

The choice of highly variable genes is important for integration. A review of integration methods emphasized the incorporation of high variance genes in annotating various technologies, aiming to obtain biologically relevant information aligned with the intended purpose [10]. Selecting genes that are variable across cell types instead of across batches improves integration performance.

Step 3: Preprocess and Normalize

Standard preprocessing includes quality control to remove low-quality spots or cells, normalization to account for differences in sequencing depth, and feature selection to identify informative genes. Normalization choices affect integration outcomes. Library size normalization is common, but more sophisticated approaches that account for platform-specific biases may be needed when integrating across technologies.

For spatial data, additional quality metrics include the number of genes detected per spot, the fraction of mitochondrial reads, and the spatial distribution of these metrics. Spots with low gene counts or high mitochondrial fractions may represent damaged tissue or empty capture locations and should be filtered before integration.

Step 4: Select and Apply Integration Method

Method selection depends on the integration goal, data characteristics, and available computational resources. For scRNA-seq anchoring approaches, the reference dataset must be of high quality and contain the cell types expected in the spatial data. For graph-based methods, the spatial adjacency graph must be constructed appropriately, with parameters such as the number of neighbors and the kernel width affecting results.

For multi-modal integration, methods such as SIMO use probabilistic alignment to map diverse single-cell modalities, including RNA, chromatin accessibility, and DNA methylation, onto spatial tissues [16]. SIMO is described as expanding beyond spatial transcriptomics integration with scRNA-seq to enable integration across multiple single-cell modalities that have not been co-profiled spatially before [16]. This capability is relevant for studies that generate complementary molecular measurements from the same tissue.

Step 5: Validate Integration Results

Validation is a critical step that is often underemphasized. Integration results should be validated using multiple criteria. Cluster marker genes should be expressed in the expected cell types and spatial locations. Spatial domains should correspond to known tissue anatomy. Batch effects should be reduced, as measured by the mixing of samples from different batches within clusters. Biological differences between conditions should be preserved, as measured by the separation of conditions within expected cell types.

Benchmarking studies provide guidance on expected performance. A benchmarking study of spatial and single-cell transcriptomics integration methods for transcript distribution prediction and cell type deconvolution was published in Nature Methods [21]. Such benchmarks help researchers understand the strengths and limitations of different methods across tissue types and technology platforms.

Options and Tradeoffs in Integration Methods

scRNA-seq Anchoring Methods

These methods use a single-cell reference dataset to annotate spatial spots. The approach typically involves identifying shared gene expression patterns between spatial spots and reference cells, then mapping each spot to the most similar reference cell type or cluster. The result is a cell type composition estimate for each spatial location.

The main advantage of anchoring methods is interpretability. The output is directly tied to known cell types, making results easy to communicate. The main limitation is dependence on reference quality. If the reference lacks a cell type present in the spatial data, or if the reference contains batch effects of its own, the annotation will be inaccurate. Additionally, anchoring methods may miss spatial-specific cell states that are not represented in the single-cell reference.

Graph-Based Spatial Learning Methods

These methods construct a graph where nodes are spatial spots and edges connect spatially adjacent spots. Graph neural networks learn representations that incorporate both gene expression and spatial context. The learned representations can be used for clustering, integration, and deconvolution.

Graph-based methods excel at identifying spatial domains and integrating multiple tissue sections. GraphST demonstrated higher clustering accuracy and better delineation of fine-grained tissue structures in brain and embryo tissues compared to existing methods [19]. The method also demonstrated superior cell-type deconvolution for capturing spatial niches such as lymph node germinal centers and exhausted tumor infiltrating T cells in breast tumor tissue [19].

The main limitation of graph-based methods is computational cost. Training graph neural networks requires substantial memory and processing time, particularly for large datasets. Parameter tuning is also required, and the choice of graph construction parameters can substantially affect results.

Probabilistic Alignment Methods

These methods model the data generation process and align datasets through probabilistic matching. SIMO uses probabilistic alignment for the spatial integration of multi-omics datasets [16]. The method is designed to integrate spatial transcriptomics with single-cell RNA-seq and to expand beyond this to integrate multiple single-cell modalities such as chromatin accessibility and DNA methylation [16].

Probabilistic methods offer the advantage of principled uncertainty quantification. The alignment produces probability distributions instead of point estimates, which can be useful for downstream analyses that require confidence measures. The main limitation is that the assumed distributional forms may not hold for all platforms, particularly for sparse or zero-inflated data.

Deep Learning Fusion Methods

Deep learning-based spatial multi-omics integration algorithms have emerged rapidly, offering new approaches for constructing unified latent representations and achieving cross-modal fusion [11]. These methods use neural network architectures to learn joint representations from multiple data modalities.

The main advantage of deep learning methods is flexibility. They can accommodate diverse data types and learn complex nonlinear relationships between modalities. The main limitations are the requirement for large training datasets and the difficulty of interpreting the learned representations. Validation is particularly important for deep learning methods, as they can overfit to technical artifacts if not properly regularized.

Observations and Measurements for Integration Quality

Metrics for Batch Effect Removal

Several metrics can quantify the degree of batch effect removal. The entropy of batch labels within clusters measures how well samples from different batches are mixed. Higher entropy indicates better mixing and thus more effective batch correction. The silhouette score measures how well clusters are separated, with higher values indicating more distinct clusters. The conservation of biological variation can be measured by the ability to distinguish known biological conditions within expected cell types.

These metrics should be computed before and after integration to quantify the improvement. A useful practice is to report these metrics alongside integration results, as they provide evidence that the integration achieved its intended effect.

Metrics for Biological Signal Preservation

Batch effect removal must be balanced against biological signal preservation. Metrics for biological signal preservation include the ability to recover known marker gene expression patterns, the concordance of spatial domains with anatomical annotations, and the reproducibility of cell type proportions across replicate samples.

A particularly important check is whether integration preserves rare cell populations. A study on specificity-driven cell-gene graph learning noted that batch correction can dilute rare-cell-specific signals in multi-sample analyses [13]. The study introduced scFormer, a heterogeneous graph transformer framework that constructs a Z-score-guided cell-gene heterogeneous graph in which highly specific marker genes serve as informational bridges [13]. This design provides a biological rationale for rare-cell recovery, as low-abundance cells can remain connected through shared high-specificity genes even when local cell-cell neighborhoods are sparse [13].

Spatial Domain Concordance

After integration, spatial domains should be evaluated for concordance with known tissue anatomy. This evaluation can be qualitative, through visual inspection of domain maps, or quantitative, through comparison with anatomical annotations or marker gene expression patterns. Discordance between integrated domains and known anatomy may indicate over-correction or under-correction of batch effects.

Records and Documentation for Reproducibility

Version Control and Parameter Logging

Integration analyses involve many parameters, including preprocessing thresholds, feature selection criteria, graph construction parameters, and model hyperparameters. All parameters should be logged and version controlled. The software versions of all tools used should be recorded, as different versions of the same tool can produce different results.

Data Provenance

Record the source of each dataset, including the platform, tissue type, sample identifier, and any preprocessing applied before integration. This information is essential for reproducing the analysis and for interpreting results in the context of data quality. Public data repositories such as those maintained by the National Center for Biotechnology Information provide standardized access to many spatial transcriptomics datasets [2].

Output Artifacts

Save all intermediate and final outputs, including normalized expression matrices, integrated embeddings, cluster assignments, and spatial domain maps. These artifacts allow downstream analyses to be repeated without rerunning the full integration pipeline. They also facilitate comparison with alternative integration methods.

Common Failure Patterns in Integration

Over-Correction of Biological Variation

The most common failure pattern is over-correction, where the integration method removes genuine biological differences along with technical variation. This failure is particularly likely when biological conditions are confounded with batches. Over-correction can be detected by the loss of known biological differences between conditions or by the merging of distinct cell types into single clusters.

Under-Correction of Batch Effects

The opposite failure is under-correction, where batch effects persist after integration. This failure is detected by the clustering of samples by batch instead of by biological condition, or by the presence of batch-specific gene expression programs. Under-correction is more common when batch effects are strong and the integration method is not sufficiently flexible.

Rare Cell Type Loss

Integration methods that rely on global similarity can merge rare cell populations into dominant populations, masking their subtle transcriptional signatures [13]. This failure is particularly problematic for studies of development, differentiation, and disease-associated transformation, where rare cell populations often drive the biological process of interest. Methods that preserve local structure or use marker gene bridges may be more effective at retaining rare cell types.

Spatial Domain Fragmentation

Integration can fragment spatial domains, producing a patchwork of small clusters instead of coherent tissue regions. This failure is often caused by over-correction or by integration methods that do not use spatial context. Graph-based methods that incorporate spatial adjacency are less prone to this failure.

Limitations of Integration Methods

Platform-Specific Biases

Each spatial transcriptomics platform has unique biases that cannot be fully removed by computational integration. These biases include differences in capture efficiency across genes, spatial resolution, and tissue penetration depth. A study introducing cycleHCR, a method that integrates multicycle DNA barcoding with hybridization chain reaction, noted that limited color channels in fluorescence microscopy have long constrained spatial analysis in biological specimens [8]. Platform-specific biases must be considered when interpreting integrated results.

Resolution Mismatches

Integration across platforms with different spatial resolutions requires assumptions about how measurements at different scales correspond. Aggregating high-resolution data to match low-resolution data loses information, while deconvolving low-resolution data to match high-resolution data introduces uncertainty. These assumptions can affect downstream interpretations.

Reference Dependence

Methods that anchor spatial data to single-cell references depend on the quality and completeness of the reference. If the reference lacks relevant cell types or contains batch effects, the integration results will be compromised. The choice of reference dataset is therefore a critical decision.

Computational Scalability

Some integration methods, particularly deep learning approaches, require substantial computational resources. Large datasets with millions of cells or spots may be intractable for some methods. Researchers should assess computational feasibility before committing to a method.

Safety and Regulatory Context for Data Sharing

Genomic Data Sharing Policies

Spatial transcriptomics data derived from human samples are subject to genomic data sharing policies. The National Institutes of Health Genomic Data Sharing Policy provides the framework for sharing genomic data generated from NIH-funded research [3]. Researchers should be aware of the requirements for data deposition, access controls, and privacy protections when working with human spatial transcriptomics data.

FAIR Data Principles

The FAIR Guiding Principles provide a framework for making data findable, accessible, interoperable, and reusable [4]. These principles are relevant to spatial transcriptomics integration because integrated datasets are valuable resources that should be shareable with the research community. Applying FAIR principles to integration outputs includes depositing processed data, documenting methods, and providing metadata that enables reuse.

Training Resources

Researchers new to spatial transcriptomics integration can access training resources from the European Bioinformatics Institute [1]. These resources provide instruction on data analysis methods and best practices for working with spatial transcriptomics data.

Professional Escalation Criteria

When to Seek Expert Consultation

Integration analyses can encounter problems that require specialized expertise. Consider consulting a bioinformatics core facility or computational biology expert when batch effects persist after applying standard methods, when integration results conflict with known biology, when working with novel platforms or data types, or when the computational requirements exceed local capacity.

When to Reconsider the Experimental Design

Some integration problems cannot be solved computationally. If batch effects are completely confounded with biological conditions, no integration method can reliably separate them. In this case, the experimental design should be reconsidered, and additional samples with balanced batch assignments should be collected.

When to Question the Reference Dataset

If integration results depend strongly on the choice of reference dataset, the reference may be inadequate. Consider whether the reference contains all expected cell types, whether it was generated from a comparable tissue, and whether it has been properly normalized. Alternative references or reference-free integration approaches may be needed.

Frequently Asked Questions

What is the difference between batch correction and data integration in spatial transcriptomics?

Batch correction specifically refers to removing technical variation between samples processed in different batches. Data integration is a broader concept that includes batch correction but also encompasses aligning datasets from different platforms, mapping spatial data to single-cell references, and combining multiple molecular modalities. Integration methods typically include batch correction as one component of a larger alignment procedure.

How do I choose between scRNA-seq anchoring and graph-based spatial learning methods?

The choice depends on your primary analytical goal. If you need to annotate spatial spots with known cell types from a single-cell reference, anchoring methods are appropriate. If you need to identify spatial domains or integrate multiple tissue sections, graph-based methods that use spatial adjacency are often superior. Some studies use both approaches in sequence, first anchoring to identify cell types and then using graph-based methods to refine spatial domains.

Can I integrate spatial transcriptomics data from different technology platforms?

Yes, integration across platforms is possible but requires careful attention to platform-specific biases. The gene overlap between platforms must be assessed, and normalization must account for differences in capture efficiency and resolution. Methods that use highly variable genes for integration are better suited to cross-platform integration because these genes are more likely to be consistently measured across technologies [10].

What is the role of spatial context in integration?

Spatial context provides additional information beyond gene expression that can improve integration. Spatially adjacent measurements are more likely to belong to the same tissue domain, and this information can be used to regularize the integration. Graph-based methods that incorporate spatial adjacency have demonstrated improved clustering accuracy and better delineation of fine-grained tissue structures compared to methods that ignore spatial context [19].

How do I validate that my integration removed batch effects without removing biological signal?

Validation requires multiple metrics. Batch effect removal can be measured by the entropy of batch labels within clusters and the silhouette score. Biological signal preservation can be measured by the recovery of known marker gene expression patterns, the concordance of spatial domains with anatomical annotations, and the reproducibility of cell type proportions across replicate samples. Report both types of metrics to demonstrate that integration achieved a balance.

What should I do if integration merges rare cell populations into dominant clusters?

Rare cell populations can be lost during integration because most methods rely on homophily-based cell-cell similarity, which can merge rare cells into dominant populations [13]. Consider using methods that preserve local structure or that use marker gene bridges to maintain connections between rare cells. Methods such as scFormer use Z-score-guided cell-gene heterogeneous graphs in which highly specific marker genes serve as informational bridges, embedding rare-cell features directly into the graph topology [13].

How does spatial transcriptomics integration differ from single-cell RNA-seq integration?

Spatial transcriptomics integration has the additional challenge of preserving spatial context. Unlike scRNA-seq integration, where cells are treated as independent measurements, spatial integration must account for the physical relationships between measurements. This requires methods that can use spatial adjacency information and that can handle the lower gene coverage typical of spatial platforms. Spatial integration also often involves mapping spatial data to single-cell references, which is not a standard task in scRNA-seq integration.

What are the main limitations of deep learning methods for spatial integration?

Deep learning methods require large training datasets and substantial computational resources. The learned representations are difficult to interpret, making validation challenging. Deep learning methods can overfit to technical artifacts if not properly regularized, and their performance may not generalize to new platforms or tissue types. Despite these limitations, deep learning-based spatial multi-omics integration algorithms have emerged rapidly and offer new approaches for constructing unified latent representations and achieving cross-modal fusion [11].

Related Bioinformatics Guides

• Modern Transcriptomics: From Bulk RNA-Seq to Single-Cell and Spatial Resolution
• Single-Cell RNA-Seq Normalization: Batch Effect Correction and Dimension Reduction (PCA, t-SNE, UMAP)
• Spatial Transcriptomics Alignment and Cellular Neighborhood Analysis
• Single-Cell Transcriptomics of Host-Pathogen Interactions During Viral Infection
• Foundation Models for Single-Cell Biology

References and Further Reading

• EMBL-EBI Training. European Bioinformatics Institute.
• NCBI Data Resources. National Center for Biotechnology Information.
• Genomic Data Sharing Policy. National Institutes of Health.
• The FAIR Guiding Principles. Scientific Data.
• Spatial transcriptomics of healthy and fibrotic human liver at single-cell resolution.. Nature communications, 2025.
• Single-cell integration and multi-modal profiling reveals phenotypes and spatial organization of neutrophils in colorectal cancer.. Cancer cell, 2026.
• Multi-omics integration and machine learning reveal gut-immune signatures in idiopathic pulmonary fibrosis: insights from bulk RNA-seq, single-cell profiles, spatial transcriptomics, and experimental validation.. Frontiers in immunology, 2026.
• Deep-tissue transcriptomics and subcellular imaging at high spatial resolution.. Science (New York, N.Y.), 2025.
• Spatial transcriptomics data and analytical methods: An updated perspective.. Drug discovery today, 2024.
• Integration tools for scRNA-seq data and spatial transcriptomics sequencing data.. Briefings in functional genomics, 2024.
• Computational methods for spatial multi-omics integration.. Biotechnology advances, 2026.
• Whole-embryo spatial transcriptomics at subcellular resolution from gastrulation to organogenesis.. Science (New York, N.Y.), 2026.
• Specificity-driven cell-gene graph learning identifies rare cell states in single-cell and spatial transcriptomic data.. 2026.
• Progress in the Application of Machine Learning in the Field of Single-Cell and Spatial Transcriptomics. 2026.
• Predicting and compensating for small-sample thermal information data in precision machine tools: A spatial-temporal interactive integration network and digital twin system approach. Applied Soft Computing, 2024.
• Spatial integration of multi-omics single-cell data with SIMO. Nature Communications, 2025.
• Data integration through brain atlasing: Human Brain Project tools and strategies. European psychiatry, 2018.
• Automatic Integration of Spatial Data into the Semantic Web. International Conference on Web Information Systems and Technologies, 2017.
• Spatially informed clustering, integration, and deconvolution of spatial transcriptomics with GraphST. bioRxiv, 2023.
• Spatial multi-omics: novel tools to study the complexity of cardiovascular diseases. Genome Medicine, 2024.
• Benchmarking spatial and single-cell transcriptomics integration methods for transcript distribution prediction and cell type deconvolution. Nature Methods, 2022.
• WEST is an ensemble method for spatial transcriptomics analysis. Cell Reports Methods, 2024.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.