A spatial multi-omics integration workflow combines molecular modalities through coordinated quality control, spatial alignment, feature harmonization, joint modeling, validation, and downstream interpretation. The goal is not simply to place transcriptomic, proteomic, epigenomic, or metabolomic measurements into the same matrix. A useful integration must preserve tissue structure while distinguishing shared biological signals from modality-specific information.

This distinction matters because each spatial modality measures a different layer of biology. RNA reflects transcriptional state, proteins are closer to functional molecular activity, chromatin accessibility and histone modifications capture regulatory potential, and metabolites reflect biochemical activity. Their feature spaces, dynamic ranges, sparsity, spatial resolutions, and noise structures are not equivalent.

For researchers new to this area, our broader spatial multi-omics integration overview provides background on the biological rationale. This article focuses instead on the practical integration path: what should happen from raw multimodal data to a reproducible biological interpretation?

Key Takeaways

• Spatial multi-omics integration is a multi-stage workflow, not a single algorithm.
• Same-section, serial-section, and unpaired modalities require different alignment assumptions.
• Each modality should pass its own QC before joint integration.
• Matrix factorization, probabilistic modeling, graph-based approaches, and optimal transport solve different problems.
• A good integrated representation should preserve shared biological structure without allowing one modality to overwhelm the others.
• Validation should include spatial reproducibility, known biology, modality balance, and sensitivity to analytical choices.
• Method selection should follow the data relationship and biological endpoint rather than the popularity of a particular algorithm.

Quick Definition
Spatial multi-omics integration is the computational process of aligning and jointly modeling two or more spatially resolved molecular modalities while retaining their tissue coordinates and modality-specific biological information.

Figure 1. End-to-end spatial multi-omics integration workflow from modality-specific data acquisition to downstream biological analysis.

Why Integration Is Difficult

Spatial transcriptomics, spatial proteomics, spatial epigenomics, and spatial metabolomics all describe the same tissue from different molecular perspectives, but their data structures can be very different.

Transcriptomic data are commonly represented as genes by cells or spatial locations. Protein measurements contain a smaller set of molecular features with different dynamic ranges. Epigenomic assays may contain highly sparse accessibility or histone-mark signals. Metabolomic measurements use mass-to-charge or annotated metabolite features rather than gene identifiers.

The spatial unit can also differ. One modality may resolve individual cells or subcellular features, while another represents multiple cells within a larger spatial unit. Image-derived histology adds another feature space entirely.

These differences create several integration problems:

• unequal feature dimensions;
• different levels of sparsity and missingness;
• modality-specific background noise;
• unequal spatial resolution;
• batch and section effects;
• incomplete correspondence between spatial coordinates;
• large differences in measurement scale.

This is why direct feature concatenation is rarely an adequate default. If one modality contains many more features or larger numerical variation, it may dominate the integrated representation even when that dominance has no biological meaning.

A multimodal spatial study should therefore treat data integration as a sequence of controlled decisions. The first decision is not which algorithm to run. It is what relationship actually exists between the datasets being integrated.

Do You Need Spatial Multi-Omics Integration?

Not every spatial biology question requires multiple molecular modalities. Adding another assay is most useful when the additional molecular layer addresses a question that cannot be answered confidently from the first modality alone.

Spatial multi-omics may be worth the added analytical complexity when a study needs to connect chromatin accessibility with transcription, compare RNA and protein states, relate metabolic activity to tissue architecture, or determine whether a spatial domain is supported by more than one molecular layer.

It can also help distinguish changes caused by cell composition from changes in molecular regulation within the same cell population.

By contrast, if the primary endpoint is simply to locate genes, define broad expression domains, or compare gene-level spatial patterns, a well-designed single-modality experiment may already answer the core question. A second modality should be added because it changes the biological interpretation, not simply because multi-omics integration is technically possible.

Researchers selecting the molecular layers for a project can explore CD Genomics spatial transcriptomics services alongside other spatial modalities before defining an integration strategy.

Step 1: Define Integration Geometry

Before preprocessing the data, determine how the modalities correspond in physical space. This "integration geometry" controls what types of assumptions the computational model can reasonably make.

Same-Section Measurements

The most direct situation occurs when multiple molecular layers are measured from the same tissue section and can be associated with the same cells, spots, or spatial coordinates.

Examples include paired RNA and protein measurements or spatial transcriptomic and epigenomic measurements obtained through compatible co-assay designs.

Because the measurements share a spatial frame, the computational task can focus more heavily on feature integration. However, same-section measurement does not eliminate resolution mismatch, missing features, or modality-specific noise.

For projects specifically combining expression and protein information, CD Genomics provides an integrated spatial transcriptomics and proteomics solution for research applications.

Serial or Adjacent Sections

Many multi-omics studies measure different modalities on neighboring tissue sections. For example, RNA may be profiled on one section and metabolites, proteins, or chromatin features on another.

These datasets are biologically related but are not physically identical.

A vessel, tumor boundary, cortical layer, or immune aggregate can shift in size and position between sections. Registration therefore becomes part of the biological analysis, not merely an image-processing step.

The integration model should not assume that coordinate (x, y) in one section represents exactly the same biological material at (x, y) in another.

Unpaired Modalities

The third situation is unpaired integration.

A spatial transcriptomic dataset may be combined with a dissociated single-cell reference, a separate epigenomic experiment, or another sample that does not share exact spatial coordinates.

These workflows require computational mapping rather than direct spatial correspondence.

The important question is therefore:

Are the modalities paired at the molecule, cell, spot, tissue-region, section, or sample level?

That answer should be documented before selecting an integration algorithm.

Are Your Datasets Integration-Ready?

Existing datasets should be reviewed before assuming that computational integration is feasible. Two datasets can originate from the same project and still lack the spatial, biological, or metadata correspondence needed for a meaningful joint analysis.

Useful questions include:

• Were the modalities generated from the same specimen, adjacent sections, or different specimens?
• Are original spatial coordinates available?
• Are tissue images or anatomical landmarks available for registration?
• Do the datasets cover comparable tissue regions?
• What is the analytical resolution of each modality: cell, spot, pixel, bin, or region?
• Are biological replicates available?
• Are raw or sufficiently processed feature matrices available?
• Are sample IDs, batch information, tissue annotations, and experimental groups consistent across datasets?
• Do reference genomes, gene annotations, and feature identifiers use compatible versions?

Missing information does not always prevent integration, but it changes which analytical strategies are defensible. Reviewing these inputs before modeling can prevent a sophisticated algorithm from being applied to datasets that lack meaningful biological correspondence.

Figure 2. Same-section, serial-section and unpaired designs require different assumptions for spatial multi-omics integration.

Step 2: Preprocess and Align

Multimodal analysis should begin with modality-specific preprocessing. Combining two low-quality datasets does not create a high-confidence integrated dataset.

QC Each Modality Separately

The first review should ask whether every modality independently supports the intended biological comparison.

Useful QC categories include:

• detected feature counts;
• signal and background distributions;
• sparsity and missingness;
• spatial coordinate completeness;
• tissue or image quality;
• feature annotation rate;
• replicate consistency;
• batch or section effects.

The relevant metrics depend on the assay. RNA count matrices, antibody-based protein measurements, chromatin-accessibility profiles, and mass-spectrometry imaging data should not be judged using the same numerical criteria.

Spatial epigenomic studies, for example, require attention to accessibility or histone-mark signal quality in addition to coordinate structure. Researchers planning this layer can review available spatial epigenomics services.

Normalize Without Removing Biology

Normalization is needed because modalities operate on different scales. Batch correction may also be required when samples, sections, processing batches, or acquisition runs introduce technical structure.

However, stronger correction is not automatically better.

If a disease-associated compartment, anatomical region, or cell state is present mainly in one batch, aggressive correction can mistakenly remove real biology together with the unwanted technical signal.

The correction strategy should therefore be evaluated against known tissue structure and biological markers.

Register Spatial Coordinates

Same-section datasets may require coordinate transformation or segmentation matching.

Serial sections require more substantial registration. Anatomical landmarks, morphology, tissue boundaries, and molecular features can help estimate correspondence between sections.

Alignment uncertainty should be retained in the interpretation. A computationally aligned region is an estimate of correspondence, not proof that every cell has an exact counterpart in the adjacent section.

Workflow at a Glance
Modality-specific QC → normalization → batch assessment → spatial registration → common analytical units → joint modeling

Step 3: Harmonize Analytical Units

Even after spatial alignment, multimodal datasets may still describe tissue at incompatible analytical scales.

Cell, Spot, Pixel, or Region?

Before joint modeling, define the unit being integrated.

Possible units include:

• segmented cells;
• spatial capture spots;
• pixels;
• bins;
• pseudo-spots;
• tissue regions.

This choice affects every downstream result.

If protein data are measured at single-cell resolution but transcriptomic data represent mixtures of several cells, direct one-to-one integration may create artificial precision.

A workflow may instead aggregate the higher-resolution modality, deconvolve the lower-resolution modality, or compare both at a shared regional scale.

Build Comparable Feature Representations

Each modality can be represented as a feature-by-location matrix:

• genes × locations;
• proteins × locations;
• accessibility features × locations;
• metabolites × locations;
• image features × locations.

The integration model then links these representations through shared samples, spatial neighborhoods, latent factors, biological priors, or learned mappings.

For metabolic studies, feature identification and spatial coordinate matching deserve particular attention because metabolite signals do not map to gene identifiers in the same way as transcriptomic or epigenomic features. CD Genomics offers spatial metabolomics services for research projects requiring this molecular layer.

The key principle is simple: alignment of images is not the same as harmonization of analytical units.

Step 4: Choose an Integration Strategy

There is no single spatial multi-omics algorithm that is appropriate for every project. Different model families answer different computational questions.

Researchers who need deeper algorithm-level comparisons can review our guide to spatial multi-omics analysis algorithms. Here, the more useful question is what each method family is designed to solve.

Figure 3. Spatial omics modalities require different integration strategies depending on data geometry and downstream research goals.

Matrix Factorization

Matrix factorization methods represent multiple omics datasets through a smaller number of shared latent factors.

MOFA+ is a representative framework. It decomposes multiple molecular matrices into latent factors that explain variation across modalities and can support variance decomposition.

MEFISTO extends this idea by incorporating temporal or spatial structure into the factor model.

These approaches are attractive when researchers want interpretable latent biological programs rather than a purely predictive embedding.

Their main limitation is that relatively simple latent-factor assumptions may not capture all nonlinear cross-modal relationships.

A useful question is:

Do I primarily need interpretable shared sources of variation?

If so, matrix factorization may be an appropriate starting point.

Probabilistic Models

Probabilistic approaches explicitly model uncertainty, technical variation, or the data-generation process.

totalVI jointly models RNA and protein measurements using a variational framework and is designed to separate biological and technical components, including protein background.

cell2location addresses a different task. It uses a hierarchical Bayesian framework and single-cell reference information to estimate the abundance of cell types in spatial locations.

RCTD also addresses spatial deconvolution through a likelihood-based framework.

These tools should not be treated as interchangeable. They belong to the same broad probabilistic family but solve different biological and analytical problems.

The relevant question is:

Do I need uncertainty-aware integration, reference-assisted mapping, background modeling, or cell-mixture decomposition?

Graph-Based Integration

Graph methods explicitly represent relationships between cells or spatial locations.

A spatial graph may connect physical neighbors, while a feature graph may connect locations with similar molecular profiles.

SpatialGlue uses spatial and feature graphs with attention-based integration to combine multiple modalities. In the benchmark results summarized in the source article, it showed strong performance across several spatial-domain metrics and datasets containing RNA, protein, or chromatin information.

MultiGATE uses graph representation learning to integrate spatial multi-omics and support regulatory inference.

SMART combines graph neural networks and metric learning for scalable spatial aggregation and cross-tissue integration.

MISO takes a related but distinct multimodal modeling approach, using interaction terms to model relationships among gene expression, protein, epigenetic, metabolomic, and histological information while accounting for modality quality.

Graph-based models are especially relevant when physical neighborhood structure is itself part of the biological hypothesis.

Ask:

Would the interpretation change if neighboring cells or locations were modeled explicitly?

If yes, graph-based integration deserves consideration.

Optimal Transport and Alignment

Some integration problems are fundamentally correspondence problems rather than latent-embedding problems.

Optimal transport provides a mathematical framework for mapping one distribution onto another while minimizing a defined cost.

PASTE uses a fused Gromov-Wasserstein strategy to consider both molecular similarity and spatial relationships when aligning tissue sections.

Related methods extend this idea to partially overlapping sections, multi-slice alignment, and three-dimensional reconstruction.

STitch3D jointly models multiple tissue sections to support reconstruction of spatial structure across the z-axis, while STAIR focuses on cross-section feature integration and alignment.

The key question here is:

Do the datasets first need to be mapped into a common spatial coordinate system?

If yes, alignment may need to happen before downstream multimodal fusion.

Step 5: Validate the Integration

A visually smooth embedding is not evidence that integration worked correctly.

Validation should ask whether the integrated model is stable, biologically plausible, and genuinely multimodal.

Check for Modality Dominance

Suppose one modality contains thousands of informative features and another contains only a much smaller targeted panel.

A joint representation may appear multimodal while being driven almost entirely by the higher-dimensional dataset.

Useful checks include:

• comparing embeddings from individual modalities with the joint result;
• quantifying modality contributions where the method supports it;
• repeating the analysis after reweighting or removing one modality;
• checking whether important features from smaller modalities remain biologically interpretable.

Test Spatial Reproducibility

Spatial domains or molecular niches should be evaluated across:

• biological replicates;
• adjacent or replicate sections;
• random seeds;
• graph parameters;
• normalization choices;
• registration settings.

A result that disappears after a small parameter change should not be interpreted with the same confidence as one that recurs across analytical settings and biological specimens.

Preserve Known Biology

Integration should not erase known markers, anatomical boundaries, or expected molecular relationships.

For example, a model that produces excellent mathematical mixing but destroys well-established tissue compartments may be optimizing the wrong objective.

Compare With Morphology

Histology provides an independent spatial reference.

Integrated molecular domains do not need to reproduce morphology exactly, but major compartment boundaries should be evaluated against tissue architecture when biologically appropriate.

Integration Quality Checklist

• Did each modality pass independent QC?
• Is one modality dominating the result?
• Are spatial domains reproducible?
• Are known biological markers preserved?
• Are results stable to reasonable analytical choices?
• Do major patterns agree with tissue morphology where expected?

Downstream Biological Analysis

Integration is valuable only when it improves biological interpretation.

Spatial Domain Identification

Joint modeling can identify tissue compartments supported by more than one molecular layer.

A transcriptomic boundary may coincide with protein abundance, chromatin accessibility, metabolic state, or histological structure, strengthening the interpretation of that domain.

Cell-Type Deconvolution and Mapping

Spatial locations containing mixtures of cells can be combined with reference datasets to estimate cell-type composition.

This can help separate changes caused by cell abundance from changes occurring within a cell population.

Cross-Modal Regulatory Inference

Epigenomic and transcriptomic data can be examined together to identify relationships between regulatory state and gene expression.

RNA and protein can be compared to investigate transcriptional and post-transcriptional patterns.

These relationships should be treated as hypotheses. A learned association, attention weight, or graph edge does not by itself establish a causal regulatory mechanism.

Cell-Cell Communication

Spatial context can refine ligand-receptor or signaling analyses by testing whether putative interacting populations are physically positioned to communicate.

3D Tissue Reconstruction

Serial sections can be aligned to reconstruct molecular patterns through tissue depth.

This can reveal structures that are difficult to interpret from a single two-dimensional section.

Choosing the Right Strategy

Method selection should begin with the data relationship and biological endpoint, not with a list of popular tools.

No category guarantees better results in every dataset.

Matrix factorization may be preferable for interpretable factors but insufficient for complex nonlinear relationships. Graph models can use tissue topology but are sensitive to how spatial neighborhoods are constructed. Probabilistic models can represent uncertainty but require an appropriate generative assumption. Alignment frameworks solve spatial correspondence but do not automatically solve downstream biological integration.

A useful method-selection sequence is:

1. Define how modalities are paired.
2. Define the common analytical unit.
3. Identify the biological output that matters.
4. Determine whether spatial neighborhood structure must enter the model.
5. Decide whether uncertainty, deconvolution, or registration is the dominant technical challenge.
6. Compare candidate models using reproducibility and biological validation, not a single benchmark score.

Planning a multimodal spatial study? Define modality pairing, spatial scale, registration requirements, and expected downstream outputs before committing to the computational framework.

Reporting and Deliverables

A spatial multi-omics project should produce enough information to trace how raw measurements became biological conclusions.

What to Prepare Before Integration

Project evaluation is easier when the relationships among samples, sections, and modalities are documented before analysis begins.

Useful inputs may include:

• a list of available molecular modalities;
• sample IDs and biological groups;
• the relationship among tissue sections and specimens;
• raw or processed feature matrices;
• spatial coordinates;
• histology or assay-associated tissue images;
• cell, spot, region, or tissue annotations where available;
• reference single-cell datasets when relevant;
• batch and experimental metadata;
• genome and annotation versions;
• the primary biological question;
• the expected downstream analysis outputs.

These inputs help determine whether the main challenge is coordinate registration, resolution harmonization, deconvolution, cross-modal modeling, or another analytical step. They also reduce the risk of selecting a method before the structure of the data is understood.

QC and Preprocessing Records

Useful records include:

• modality-specific QC summaries;
• feature filtering;
• normalization methods;
• batch handling;
• spatial coordinate processing;
• registration steps;
• analytical-unit definitions.

Integrated Outputs

Depending on the research question, useful outputs may include:

• aligned spatial coordinates;
• integrated embeddings;
• spatial-domain maps;
• cell-type or cell-state maps;
• modality contribution summaries;
• differential molecular features;
• cross-modal associations;
• pathway enrichment;
• cell-cell communication results;
• regulatory hypotheses;
• multi-slice or 3D reconstructions.

CD Genomics supports research projects requiring spatial transcriptomics data analysis, including spatial mapping, annotation, comparative analysis, visualization, and downstream interpretation.

Reproducibility Evidence

A complete analysis record should also retain:

• software and method versions;
• parameter settings;
• graph definitions where applicable;
• registration assumptions;
• sensitivity tests;
• replicate comparisons;
• analysis-ready figures and tables.

Useful Deliverable Principle
A result is easier to interpret and reproduce when the final package shows not only the integrated map, but also how QC, alignment, feature processing, and model choices produced it.

Common Integration Failure Modes

Even technically sophisticated integration methods can fail when upstream assumptions are incorrect.

Resolution Mismatch

Two modalities measured from the same tissue do not necessarily describe the same biological unit.

Mapping subcellular measurements directly to multi-cell spatial spots can create false precision. The integration scale should therefore be selected explicitly.

Modality Imbalance

High-dimensional RNA data can dominate a smaller protein or metabolite feature set.

This can produce an integrated embedding that appears multimodal but mostly reproduces transcriptional structure.

Modality contribution or perturbation analyses can help identify this problem.

Registration Error

Adjacent sections are not identical sections.

Small registration errors can move a molecular signal across an anatomical boundary or falsely suggest that two modalities occupy the same region.

Over-Correction

Batch correction can remove technical variation, but it can also remove biological differences when batch and biological condition are partially confounded.

Integrated datasets should therefore be checked against known biology before and after correction.

When Modalities Disagree

Cross-modal disagreement is not automatically evidence that an experiment or integration has failed.

RNA, protein, chromatin, and metabolite signals reflect different biological layers and can respond on different timescales. Post-transcriptional regulation, protein stability, metabolic buffering, or delayed transcriptional responses can create genuine differences between modalities.

Discordance can also arise from technical causes, including unequal spatial resolution, section-to-section variation, registration error, low signal quality, missing features, or an integration model that overweights one dataset.

The practical goal is therefore to distinguish biological discordance from technical discordance. Review modality-specific QC, tissue morphology, spatial correspondence, replicate consistency, and known biological relationships before forcing disagreeing datasets into a common pattern.

Over-Interpreting Learned Relationships

Attention scores, latent-factor loadings, correlations, graph connections, and predicted cross-modal relationships are computational evidence.

They are not automatically mechanistic evidence.

Regulatory relationships or molecular interactions should be described as candidates unless supported by appropriate independent validation.

Frequently Asked Questions

From Integration to Interpretation

Spatial multi-omics integration should not be treated as an algorithm-selection exercise alone.

A defensible workflow moves through a clear sequence:

biological question → integration geometry → modality-specific QC → spatial alignment → analytical-unit harmonization → model selection → validation → biological interpretation

The most useful model is not necessarily the most complex one. It is the model that preserves relevant spatial structure, retains meaningful modality-specific information, withstands reasonable sensitivity testing, and produces outputs that directly answer the original research question.

Planning a research-use spatial multi-omics project? Evaluate modality compatibility, spatial alignment requirements, QC checkpoints, analytical units, and expected downstream outputs before finalizing the experimental and computational design.

Reference

1. Spatial multiplexing and omics. Nature Reviews Methods Primers. 2024.
2. Deciphering spatial domains from spatial multi-omics with SpatialGlue. Nature Methods. 2024.
3. Resolving tissue complexity by multimodal spatial omics modeling with MISO. Nature Methods. 2025.
4. MultiGATE: integrative analysis and regulatory inference in spatial multi-omics data via graph representation learning. Nature Communications. 2025.
5. Integrative deep learning of spatial multi-omics with SWITCH. Nature Computational Science. 2025.
6. SMART: spatial multi-omic aggregation using graph neural networks and metric learning. Nature Communications. 2026.
7. Optimal transport for single-cell and spatial omics. Nature Reviews Methods Primers. 2024.
8. Multitask benchmarking of single-cell multimodal omics integration methods. Nature Methods. 2025.