SpaMOAL: A spatial multi-omics graph contrastive learning method for spatial domains identification

SpaMOAL is a graph-based contrastive learning method that integrates spatial coordinates, histological images, and multi-omics molecular profiles to accurately identify spatial tissue domains, outperforming existing approaches in benchmarking studies.

The Gist

Imagine you are trying to understand the layout of a bustling, ancient city. You have three different maps of this city, but they are all drawn in different languages and styles:

1. The Census Map: Lists every person's job and hobbies (Molecular data like RNA).
2. The Property Deed Map: Shows who owns the land and what rules apply to the soil (Epigenetic data like ATAC).
3. The Satellite Photo: Shows the actual buildings, streets, and parks from above (Histology images).

The Problem:
Scientists have been trying to figure out which parts of the city are "residential," which are "industrial," and which are "commercial" (identifying spatial domains in tissue). But looking at just one map is confusing.

• If you only look at the Census, you might group people by job, but they could be living in totally different neighborhoods.
• If you only look at the Satellite Photo, you see the buildings, but you don't know what's happening inside them.
• Old computer programs tried to combine these maps, but they often got confused, mixed up the languages, or ignored the street layout, resulting in a messy, inaccurate city plan.

The Solution: SpaMOAL
The authors of this paper created a new tool called SpaMOAL. Think of SpaMOAL as a super-smart, bilingual city planner who uses a special technique called "Graph Contrastive Learning."

Here is how it works, using a simple analogy:

1. The Neighborhood Watch (Graph Construction)

Instead of looking at the whole city at once, SpaMOAL looks at small neighborhoods. It draws a web connecting every house to its 6 closest neighbors. This ensures that the tool understands that "what happens next door matters."

2. The "Shared vs. Private" Translator

This is the magic trick. SpaMOAL realizes that while the Census, Property Deeds, and Satellite Photos are different, they all describe the same city.

• The Shared Language: It finds the common truths. (e.g., "This area is definitely a park," regardless of whether you look at the trees in the photo or the birds in the census).
• The Private Details: It also keeps the unique details. (e.g., The Census might tell you about a specific festival happening only in the residential zone, which the satellite photo can't see).

SpaMOAL separates these two types of information so it doesn't get confused, then stitches them back together into one perfect, unified map.

3. The "Self-Correction" Loop (Contrastive Learning)

How does the tool know it's right? It uses a game of "Spot the Difference."

• It asks: "Are these two neighbors similar?" If they are in the same neighborhood, it pulls them closer together in its mental map.
• It asks: "Are these two neighbors different?" If one is a factory and the other is a bakery, it pushes them far apart.
• It does this over and over, constantly correcting its own mistakes until the map is crystal clear.

Why is this a big deal?

The researchers tested SpaMOAL on real biological data, like:

• Developing Mouse Brains: They watched a brain grow from a tiny embryo to a complex organ. SpaMOAL could see the distinct layers and regions forming perfectly, whereas other tools saw a blurry mess.
• Human Breast Cancer: In cancer, the immune system tries to build "fortresses" (called Tertiary Lymphoid Structures) to fight the tumor. SpaMOAL could pinpoint exactly where these fortresses were and even tell the difference between the "active soldiers" inside and the "scouts" on the outside. Other tools missed these subtle details.

The Bottom Line

Before SpaMOAL, trying to understand complex tissues was like trying to solve a jigsaw puzzle while wearing blindfolds and only having pieces from three different boxes. SpaMOAL takes off the blindfolds, sorts the pieces by color and shape, and assembles the picture so clearly that we can finally see the true architecture of life.

It helps doctors and scientists understand how tissues are built, how they break down in disease, and potentially how to fix them.

Technical Summary

1. Problem Statement

The emergence of spatial multi-omics technologies (e.g., spatial transcriptomics, proteomics, and epigenomics) allows for the simultaneous profiling of multiple molecular layers within their native tissue context. However, analyzing these data presents significant computational challenges:

• Heterogeneous Integration: Different omics modalities (RNA, ATAC, protein) and histological images exist in distinct feature spaces. Simple concatenation or prior-knowledge-based conversion often leads to information loss or fails to capture complex non-linear relationships.
• Spatial and Morphological Context: Existing methods often neglect high-resolution spatial coordinates or fail to effectively integrate morphological information from histology images (e.g., H&E stains), which are crucial for defining tissue architecture.
• Limitations of Current Tools:
  • Single-cell multi-omics tools (e.g., Seurat WNN, MOFA+) lack spatial coordinate modeling.
  • Spatial transcriptomics clustering tools (e.g., SpaGCN, MUSE) generally cannot handle multi-omics data beyond transcriptomics.
  • Recent multi-omics methods (e.g., SpatialGlue, MISO) have specific gaps: SpatialGlue does not support histology images, while MISO lacks explicit spatial coordinate modeling and requires manual specification of low-quality modalities.

2. Methodology: SpaMOAL

SpaMOAL (Spatially Multi-Omics graph contrAstive Learning) is a graph-based contrastive learning framework designed to learn clustering-friendly representations by integrating spatial coordinates, molecular profiles, and morphological images.

Core Workflow

1. Data Preprocessing & Feature Extraction:
   • Omics Data: RNA, protein (ADT), and chromatin accessibility (ATAC) data are normalized and filtered. Highly Variable Genes/Features (HVGs) are selected.
   • Imaging Data: High-resolution H&E or immunofluorescence images are processed using pretrained CNNs (Inception_v3 or ResNet50) to extract morphological features for each spatial spot/cell.
2. Graph Construction:
   • An adjacency matrix is built based on Euclidean distances between spatial coordinates (k-nearest neighbors, default $k=6$).
   • Modality-specific spatial neighbor graphs are constructed where nodes represent spots/cells and edges represent spatial proximity.
3. Graph Contrastive Autoencoder Architecture:
   • Graph Encoders: Graph Convolutional Networks (GCNs) process each modality's graph to generate initial node representations.
   • Disentanglement: A Multi-Layer Perceptron (MLP) splits the representation into two components for each modality:
     • Shared Representations ($C$): Captures common biological signals across modalities.
     • Private Representations ($P$): Captures modality-specific variations.
4. Optimization Objectives:
   SpaMOAL jointly optimizes four loss functions to balance structural consistency, alignment, and discrimination:
   • Matching Loss ($\mathcal{L}_{mat}$): Aligns shared representations ($C$) across different modalities toward a unified latent space (using a shared variable $S$ derived via SVD) to ensure cross-modal coherence.
   • Correlation Loss ($\mathcal{L}_{cor}$): Enforces statistical independence between shared and private representations within each modality to reduce redundancy (minimizing Pearson correlation).
   • Reconstruction Loss ($\mathcal{L}_{rec}$): Ensures the latent representations can reconstruct original node features and preserve local neighborhood consistency in the spatial graph.
   • Contrastive Loss ($\mathcal{L}_{con}$): Enhances discriminative power in the private space. It pulls together spatially adjacent spots assigned to the same pseudo-cluster and pushes apart those in different clusters, refining microenvironment boundaries.
5. Final Output:
   • The shared variable $S$ and private representations $P$ are concatenated to form unified embeddings.
   • These embeddings are clustered (using Mclust, Leiden, or Louvain) to identify spatial domains.

3. Key Contributions

• Unified Multimodal Framework: SpaMOAL is one of the first methods to simultaneously integrate spatial transcriptomics, epigenomics, proteomics, and high-resolution histology images into a single graph contrastive learning framework.
• Disentangled Representation Learning: By explicitly separating shared and modality-specific components, the method captures both the consensus biological state and unique modality features, improving robustness against noisy or low-quality data modalities.
• Spatially Constrained Contrastive Learning: The method uniquely incorporates spatial adjacency into the contrastive loss, ensuring that learned clusters respect local tissue microenvironments and boundaries.
• Scalability: The algorithm is designed to be efficient, handling large-scale datasets with multiple modalities without the computational bottlenecks seen in competing methods.

4. Results

The authors benchmarked SpaMOAL against SpatialGlue and MISO across 11 datasets (17 samples), including simulated data, mouse brain development (MISAR-seq), mouse embryonic tissues (spatial ATAC-RNA-seq), and human breast cancer (spatial RNA-Protein-seq).

• Simulated Data: SpaMOAL consistently outperformed competitors in identifying ground-truth spatial domains across dual-omics (RNA-ATAC), triple-omics (RNA-ATAC-Protein), and image-integrated scenarios. It achieved higher Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), and V-measure scores.
• Mouse Brain Development (MISAR-seq):
  • SpaMOAL successfully resolved fine-grained structures (e.g., Cartilage, Tongue/muscle, Thalamus) across embryonic stages (E11–E18) that other methods fragmented or missed.
  • Integration of H&E images significantly improved the delineation of morphologically ambiguous tissues (e.g., separating hindbrain and midbrain).
• High-Resolution Mouse Brain (Spatial ATAC-RNA-seq):
  • SpaMOAL identified subtle laminar structures (e.g., ventrolateral thalamic nucleus) and marginal layers that were missed by unimodal or other multi-omics approaches.
  • It demonstrated superior spatial coherence, producing smoother domain boundaries compared to the fragmented outputs of MISO.
• Human Breast Cancer:
  • SpaMOAL accurately identified Tertiary Lymphoid Structures (TLS), distinguishing the B-cell core from the T-cell periphery.
  • It resolved functionally distinct subregions within basal-like tumors (proliferative vs. quiescent) by leveraging protein and morphological cues, even when transcriptional signals were weak.
• Efficiency: SpaMOAL demonstrated significantly lower running times compared to MISO (e.g., <1000s vs. >1800s on large datasets) while maintaining or exceeding accuracy.

5. Significance

• Biological Insight: SpaMOAL provides a powerful tool for deciphering complex tissue architectures in development and disease. Its ability to integrate morphology with multi-omics data allows for the discovery of spatial domains that are invisible to transcriptomics alone.
• Clinical Relevance: In cancer research, the ability to precisely map immune niches (like TLS) and tumor subpopulations with distinct proliferative potentials offers new avenues for prognostic assessment and therapeutic targeting.
• Methodological Advancement: By addressing the "modality gap" and "spatial gap" simultaneously, SpaMOAL sets a new standard for spatial multi-omics analysis, moving beyond simple data integration to true multimodal representation learning.
• Accessibility: The method is open-source, scalable, and robust to parameter variations, making it a practical solution for the growing volume of spatial multi-omics data generated by next-generation sequencing and imaging platforms.