Spatial Multi-Omics Analysis Algorithms: From MultiGATE to stClinic

Inquiry

Spatial multi-omics datasets promise something researchers have wanted for decades: the ability to interpret cell state and regulation in the context of tissue structure, rather than in a dissociated average. The challenge is that spatial transcriptomics, imaging proteomics (e.g., IMC/CODEX), and chromatin accessibility often have different noise profiles, dynamic ranges, sparsity, and spatial resolutions. That's why "feature concatenation + clustering" rarely holds up in spatial multi-omics.

This article spotlights four peer-reviewed computational frameworks—MultiGATE, SMODEL (dual-graph regularized ensemble learning), scMODAL, and stClinic—and focuses on when to use each method, what you can reasonably report, and what pitfalls commonly break reproducibility.

Quick Method Picker: Which Algorithm Should You Use?

• Need a unified embedding that respects tissue neighborhoods → MultiGATE
• Need robust spatial domain detection across multiple modalities → SMODEL
• Need biologically constrained single-cell multi-omics alignment → scMODAL
• Need multi-slice niche discovery with continuity across sections → stClinic
  Two failure points to watch: (1) graph construction + modality scaling and (2) multi-slice registration consistency. This is a research-use discussion of computational analysis and reporting—not clinical decision-making.

If you want a broader, non-algorithm-specific overview of integration concepts and modality landscapes, see Frontiers in Spatial Multi-Omics Integration: From Data to Mechanism.

At-a-Glance Comparison of Spatial Multi-Omics Integration Algorithms

This table is designed to be easy to scan, share, and reference in project planning or peer review.

Quick Glossary for Spatial Multi-Omics Modeling

• Spatial graph: a graph where nodes are spots/cells and edges connect physical neighbors (kNN, radius, Delaunay, etc.).
• Omics similarity graph: a graph where edges connect spots/cells with similar molecular profiles (RNA, protein, ATAC), regardless of physical distance.
• Latent embedding: a learned low-dimensional representation intended to capture shared structure across modalities.
• Spatial domain: a region of coherent molecular/spatial structure (often used for tissue partitioning).
• Tissue niche: a reproducible microenvironmental unit defined by coordinated function + spatial structure (often spanning multiple slices or conditions).
• Feature links: explicit biological relationships used as constraints (e.g., ATAC peaks → genes, proteins ↔ genes).

Figure 2. A practical decision tree for selecting an algorithm based on the primary analysis goal in spatial multi-omics projects.

Algorithm Spotlight: MultiGATE for Graph Attention Autoencoder Spatial Multi-Omics Integration

What It's For

MultiGATE is designed for integrative analysis of spatial multi-omics data, where you want a shared representation that respects both molecular features and spatial topology. It is especially relevant when neighborhood context (local microenvironment) is part of the signal.

Core Idea in One Sentence

MultiGATE builds a spatial neighbor graph and uses a graph attention autoencoder to learn a joint latent space that fuses multiple omics modalities while weighting informative neighbors.

Figure 3. Overview of MultiGATE: a two-level graph attention autoencoder that integrates spatial topology with multi-omics feature connectivity.

Inputs You Need

MultiGATE is often positioned to integrate combinations of:

• spatial transcriptomics (spots or spatially resolved single cells),
• optional scRNA-seq references,
• imaging proteomics (e.g., IMC/CODEX-like protein measurements),
• chromatin accessibility or epigenomic layers (ATAC/CUT&Tag-like signals).

If you're deciding between sequencing- and imaging-based approaches before integration, see Spatial Transcriptomics Platforms: Sequencing vs Imaging and How to Choose.

What You Get

• A joint embedding that can support clustering, domain-like structure, and downstream analyses.
• Model-derived signals that can support hypothesis generation around neighborhood influence or candidate regulatory patterns.

When To Use

Use MultiGATE when your core question resembles:

• "graph attention autoencoder for spatial multi-omics integration"
• "how to integrate spatial transcriptomics with imaging proteomics (IMC/CODEX)"
• "spatial multi-omics latent embedding that preserves tissue topology"

Practical Tips and Common Pitfalls

1. Graph construction is not a detail—it's part of the model.
   A kNN graph that is too dense can oversmooth boundaries; too sparse can fragment true domains. A practical habit is to run a sensitivity check on at least two neighbor settings (e.g., k=6 vs k=12) and report whether major conclusions hold.
2. Normalize modality influence before integration.
   Imaging proteins and ATAC-like features can have different sparsity and dynamic range than RNA. If one modality dominates gradients, your embedding may reflect measurement scale rather than biology. Standardize each modality (or use method-supported weighting) and validate that marker patterns remain plausible.
3. Treat inferred regulatory links as candidates, not conclusions.
   If you use any regulatory inference outputs, frame them as hypothesis-generating and prioritize orthogonal validation (e.g., perturbation, reporter assays, or spatially localized evidence such as RNAscope/IF for key nodes).

Minimum Reporting Checklist

• Inputs and preprocessing: modalities, filtering, normalization/scaling per modality.
• Graph construction: neighbor rule and parameters (k/radius), distance metric, boundary handling.
• Stability evidence: seed repeats and sensitivity to neighbor parameters; brief summary of what stays consistent.

Algorithm Spotlight: SMODEL and Dual-Graph Regularization for Spatial Domain Detection

What It's For

The dual-graph ensemble approach (named SMODEL in the referenced paper) targets a practical pain point: in spatial multi-omics, direct feature concatenation often fails because one modality can dominate and because spatial continuity must be respected without erasing modality-specific structure. SMODEL is best thought of as a robust spatial domain detection framework in multi-modal settings.

Core Idea in One Sentence

SMODEL builds two complementary graph constraints—a spatial adjacency graph and omics similarity graphs—then uses regularized ensemble learning to produce stable domain assignments.

Inputs You Need

• Multi-modal measurements per spot/cell (e.g., RNA + protein + chromatin accessibility).
• Spatial coordinates (to define adjacency).
• Optional: multiple learners/feature extractors per modality.

What You Get

• Spatial domain labels designed to be more robust when modalities disagree, noise is high, or one modality has lower quality.
• A framework where "separate first, integrate later" helps avoid early fusion bias.

When To Use

Use this method when your core query is:

• "dual-graph regularization spatial domain detection"
• "ensemble learning for spatial multi-omics domains"
• "robust spatial domain identification across RNA and protein modalities"

Practical Tips and Common Pitfalls

1. Domain detection depends on sampling more than people admit.
   If your tissue section under-samples a key structure, no algorithm can recover it reliably. Before tuning graphs, confirm you have enough coverage of expected compartments and consistent section orientation across replicates.
2. Avoid optimizing to a single "pretty map."
   Tuning neighbor graphs until boundaries look visually pleasing is common and hard to defend. Predefine graph rules (k/radius) and judge results using reproducibility (agreement across runs, slices, or replicate sections).
3. Dual-graph regularization balances constraints—it does not eliminate ambiguity.
   Some tissues contain sharp transitions in one modality and gradients in another. In those cases, report domains as structured hypotheses and validate using orthogonal markers or morphology.

Minimum Reporting Checklist

• Inputs by modality: coverage, sparsity, and any modality-specific preprocessing.
• Graph details: spatial adjacency (k/radius) plus omics similarity graph construction (distance metric).
• Robustness: agreement across seeds and neighbor settings; domain correspondence to at least one external signal.

Algorithm Spotlight: scMODAL and Feature Links for Multi-Omics Alignment

What It's For

scMODAL focuses on single-cell multi-omics alignment and addresses a core question: how do we align RNA, ATAC, protein, or other modalities into a shared representation without erasing biologically meaningful modality-specific variation?

Core Idea in One Sentence

scMODAL uses multi-encoders mapped into a shared latent space and introduces feature links (biological priors) so that features with known relationships align more coherently.

Inputs You Need

• scRNA + scATAC (multiome) or CITE-seq-type inputs (RNA + protein).
• Feature link definitions, such as peak-to-gene relationships and protein ↔ gene mappings.

For practical co-analysis considerations when combining transcriptomic and chromatin signals, read Spatial Transcriptomics and Spatial Epigenomics Co-analysis: Tools, Workflow, and Case Studies.

What You Get

• A unified embedding supporting joint clustering and cell-state alignment.
• Better interpretability when feature links reflect real biology rather than purely statistical proximity.

When To Use

This is a strong candidate for queries like:

• "feature links multi-omics alignment"
• "deep learning framework for scRNA scATAC integration"
• "how to align CITE-seq RNA and protein without losing modality-specific signals"

Practical Tips and Common Pitfalls

1. Feature links are a dependency, not a garnish.
   Poor or incomplete mappings can bias results. Treat feature links as curated input artifacts worth versioning (genome build, annotation version, mapping logic).
2. Don't over-constrain rare states.
   Strong constraints can compress rare transitional states if links favor canonical programs. Consider conservative link weighting first, then evaluate whether rare states remain separable.
3. Validate alignment using biology, not only mixing metrics.
   "Good mixing" can still be wrong. Use known markers, motif activity consistency (for ATAC), or expected protein–RNA relationships as sanity checks.

Minimum Reporting Checklist

• Feature links described: source, mapping logic, and versioning (what exactly was linked).
• Alignment sanity checks: marker consistency and whether rare states persist.
• Sensitivity: link strength or link set perturbation does not flip major conclusions.

Algorithm Spotlight: stClinic and Dynamic Spatial Graph Tissue Niches Across Multiple Slices

What It's For

stClinic targets a common limitation in spatial analysis: many workflows assume a single section captures the relevant structure. In practice, key tissue features extend across depths. stClinic is designed for multi-slice niche discovery, aiming to move from slice-specific clusters to trackable niches.
Note: the phrase "clinically relevant" appears in the original paper title; the discussion here is research-focused.

Core Idea in One Sentence

stClinic models multi-slice data as dynamic spatial graphs, encoding within-slice adjacency and cross-slice continuity so that niches are defined by stability across sections and coordinated multi-omics signals.

Figure 4. Overview of stClinic: dynamic graph learning integrates multi-slice spatial data to identify stable niches and link niche patterns to outcomes.

Inputs You Need

• Multiple spatial sections (slices) with consistent annotation and metadata.
• Multi-omics features per spot/cell (at minimum transcriptomic signals; additional modalities when available).
• A slice correspondence approach (registration/alignment strategy).

What You Get

• Niche definitions that can be followed across slices.
• Niche-level signatures that can be more robust than single-slice clustering when structures are continuous.

When To Use

Use stClinic for questions like:

• "dynamic spatial graph tissue niches"
• "multi-slice spatial niche discovery workflow"
• "how to analyze spatial data across multiple sections without losing continuity"

Practical Tips and Common Pitfalls

1. Registration consistency is the price of admission.
   Small annotation drift can create "new niches" that are actually mis-registered structures. Document your registration method and verify alignment using landmarks and a small marker panel when feasible.
2. Cross-slice missingness must be modeled or acknowledged.
   Adjacent slices are not identical. Be explicit about how missing regions are handled, and avoid claiming continuity where tissue is absent.
3. Report niche stability, not just niche labels.
   A niche is most convincing when it recurs across adjacent slices and has coherent signatures. Include recurrence evidence and a brief stability summary.

Minimum Reporting Checklist

• Registration described: how slices were aligned and verified.
• Continuity evidence: niche recurrence across adjacent sections with coherent signatures.
• Robustness: sensitivity to cross-slice linking parameters and handling of missing regions.

How to Choose the Right Spatial Multi-Omics Algorithm

Use these decision rules to select a method without overfitting to a single dataset:

1. If your priority is a joint embedding that respects tissue topology: start with MultiGATE.
2. If your priority is robust spatial domain detection across noisy modalities: choose SMODEL.
3. If your priority is biologically constrained single-cell alignment: choose scMODAL.
4. If your priority is continuity and niche discovery across multiple slices: choose stClinic.

For a broader survey of computational categories (deconvolution, segmentation, graph learning, and general ML strategy) without repeating details here, see Computational Strategies and Machine Learning for Spatial Genomics Data.

Implementation Checklist for Reproducible Spatial Multi-Omics Analysis

1. Input QC by modality
   • Report sparsity and missingness per modality.
   • Confirm spot/cell coverage matches the structures you want to model.
2. Graph construction transparency
   • Specify neighbor rule (kNN, radius, Delaunay) and parameters.
   • Document boundary handling (tissue edges often create artifacts).
3. Normalization and modality balancing
   • Standardize features by modality or apply method-supported weighting.
   • Avoid early concatenation unless you demonstrate that one modality does not dominate.
4. Stability checks
   • Rerun with multiple seeds.
   • Test sensitivity to neighbor parameters.
   • If multi-slice: test sensitivity to registration or cross-slice linking thresholds.
5. Reporting outputs
   • Provide domain/niche maps plus a short table of top markers/pathways per region.
   • Clarify that inferred regulatory or interaction signals are hypothesis-generating unless validated.

For practical guidance on analysis reporting and QC patterns that translate well across spatial workflows, refer to Spatial Transcriptomics Data Analysis: Workflow & Tips.

Common Pitfalls and Practical Tips

• Predefine ROIs and graph rules before looking at expression maps.
  This avoids unintentionally tuning to the picture.
• Use at least one orthogonal validation signal.
  Even a small ISH/IF marker panel or morphology landmark validation can prevent over-interpretation of algorithmic partitions.
• Avoid over-claiming regulatory inference.
  Treat inferred networks as candidate hypotheses that require follow-up validation.
• Multi-slice analysis is a workflow problem, not just a model choice.
  Registration, metadata consistency, and missing regions determine whether multi-slice niches are meaningful.
• Log everything.
  Random seeds, graph parameters, feature link versions, and preprocessing choices should be treated as first-class artifacts.

Data and Code Traceability

For data and code availability, follow each paper's Supplementary Methods and repository links (when provided by the authors). If public deposition exists, use the accession identifiers listed in the paper and reanalyze data under the same ROI, graph construction, and QC definitions to avoid mixing incompatible units of comparison.
For a quick checklist on locating public spatial omics datasets and interpreting deposition metadata, see How to Find and Use Spatial Omics Datasets Efficiently.

FAQs

How Do I Integrate Spatial Transcriptomics With Imaging Proteomics (IMC or CODEX) Without One Modality Dominating?
Start by balancing modality scales (normalization/standardization per modality), then choose an integration method that respects spatial topology (often graph-based). Validate with marker concordance: RNA and protein should agree for well-characterized markers, but mismatches should be interpretable rather than forced to align.
What's the Best Way to Choose k for a Spatial Neighbor Graph in Spatial Multi-Omics Models?
There is no universal k. Choose a neighborhood size that matches the tissue scale of your question (cell neighborhoods vs broader compartments), then run sensitivity checks with at least one alternative k and report stability.
How Can I Tell If My Spatial Domains Are Robust Across Modalities and Batches?
Use stability evidence (agreement across seeds and neighbor settings), check consistency across replicates/sections, and confirm domains with an orthogonal signal (histology landmarks, IF/ISH, or known markers). Avoid optimizing parameters solely for smoother visuals.
What's the Difference Between Spatial Clusters, Spatial Domains, and Tissue Niches?
Clusters are algorithmic groupings in expression space. Domains are spatially coherent partitions intended to reflect tissue organization. Niches emphasize functionally coordinated microenvironments and are strengthened by recurrence and continuity (often across slices), not only local similarity.
How Do I Analyze Multi-Slice Spatial Data Without Losing Continuity Across Sections?
Prioritize consistent section orientation and registration, then use methods that explicitly encode cross-slice relationships (dynamic graphs or cross-slice linking). Report niche recurrence across adjacent slices and document how missing regions were handled.

Ready to Apply These Methods in Your Research Workflow

These four frameworks highlight a consistent trend in spatial multi-omics: performance and interpretability improve when models explicitly encode spatial topology, preserve modality-specific structure, and (for multi-slice studies) enforce continuity across sections. Just as importantly, reproducible outcomes depend on how you define graphs, balance modalities, document preprocessing, and validate results—not on the model name alone.
For research teams that want an execution partner for spatial profiling and downstream analysis reporting (research-use workflows only), CD Genomics supports projects through Spatial Omics Lab, including planning around data types, QC, and interpretable deliverables such as QC summaries, domain/niche maps, marker tables, sensitivity reports, and analysis-ready figures; see Spatial Transcriptomics Services.

References

1. Miao J, Li J, Xin J, et al. MultiGATE: integrative analysis and regulatory inference in spatial multi-omics data via graph representation learning. Nature Communications. 2025;16(1):9403.
2. Li Y, Cai G, Chen F, Wen K, Ou-Yang L. Unveiling spatial domains from spatial multi-omics data using dual-graph regularized ensemble learning (SMODEL). Communications Biology. 2025;8(1):945.
3. Wang G, Zhao J, Lin Y, Liu T, Zhao Y, Zhao H. scMODAL: a general deep learning framework for comprehensive single-cell multi-omics data alignment with feature links. Nature Communications. 2025;16(1):4994.
4. Zuo C, Xia J, Xu Y, et al. stClinic dissects clinically relevant niches by integrating spatial multi-slice multi-omics data in dynamic graphs. Nature Communications. 2025;16(1):5317.