Mosaic integration of spatial multi-omics with SpaMosaic

The paper introduces SpaMosaic, a scalable tool leveraging contrastive learning and graph neural networks to integrate heterogeneous spatial multi-omics datasets into a unified latent space for robust domain identification, batch correction, and accurate imputation of missing modalities.

The Gist

Imagine you are trying to solve a massive, intricate jigsaw puzzle of a living city (your body's tissue), but you don't have all the pieces from a single box. Instead, you have several boxes from different manufacturers.

• Box A has pieces showing the traffic flow (RNA/gene expression).
• Box B has pieces showing the power grid (chromatin accessibility/epigenetics).
• Box C has pieces showing the water pipes (histone modifications).
• Box D has pieces showing the buildings (proteins).

The problem? These boxes were made at different times, with different colors, and different scales. Some pieces are missing entirely from certain boxes. Trying to glue them together manually is a nightmare; the colors don't match, and the shapes are slightly off.

Enter SpaMosaic.

Think of SpaMosaic as a super-intelligent, magical "Puzzle Master" that can take these mismatched boxes and weave them into one perfect, seamless map of the city.

Here is how it works, broken down into simple concepts:

1. The "Mosaic" Problem

In the past, scientists could only look at one type of data at a time (just the traffic, or just the power grid). Newer technologies let us see two or three things at once, but it's expensive and hard to get everything on the exact same slice of tissue.

So, researchers end up with a "mosaic" of data:

• Section 1: Has Traffic + Power.
• Section 2: Has only Traffic.
• Section 3: Has only Power.

They need to combine these to see the whole picture. But because the sections were taken at different times or with different machines, the data is "noisy" and doesn't line up perfectly.

2. The Magic Tool: SpaMosaic

SpaMosaic is a computer program designed to fix this. It uses two main tricks:

• The "Translator" (Contrastive Learning): Imagine you have two people speaking different languages describing the same neighborhood. One says "The big red house," the other says "The crimson dwelling." SpaMosaic acts as a translator. It learns that "red house" and "crimson dwelling" are the same thing, even if they look different on paper. It forces the different data types to speak the same "language" so they can be compared.
• The "Mapmaker" (Graph Neural Networks): Instead of just looking at the data points in isolation, SpaMosaic looks at the neighborhood. It knows that in a city, the house next door is usually similar to the current house. It builds a "web" (a graph) connecting spots that are physically close to each other. This helps it smooth out the noise. If one data point looks weird, the tool checks its neighbors to see what the "real" signal should be.

3. What Does It Actually Do?

A. It cleans up the mess (Batch Correction)
If you took a photo of a park in the morning and another at night, the lighting would be totally different. SpaMosaic adjusts the "lighting" so the park looks consistent, regardless of when the photo was taken. It removes the technical "glitches" caused by different machines or times.

B. It fills in the blanks (Imputation)
This is the coolest part. If Section 2 is missing the "Power Grid" data, SpaMosaic doesn't just leave a blank spot. It looks at the "Traffic" data in Section 2, compares it to the "Traffic + Power" data in Section 1, and guesses what the Power Grid likely looks like in Section 2.

• Analogy: If you know a house has a very busy street in front of it (Traffic), and you know from other houses that busy streets usually have high-voltage power lines (Power), SpaMosaic can infer that this house probably has high-voltage lines too, even if you never measured them.

C. It finds the neighborhoods (Spatial Domains)
By combining all this info, SpaMosaic can clearly define "neighborhoods" in the tissue. It can say, "This area is the liver," or "This area is the brain's memory center," with much higher accuracy than looking at just one type of data.

4. Why is this a Big Deal?

Before SpaMosaic, trying to combine these different data sources was like trying to build a house with bricks from different eras; the walls would be crooked, and the rooms wouldn't line up.

SpaMosaic allows scientists to:

1. Build a "Google Maps" of the body: Create a comprehensive atlas where you can see genes, proteins, and chemical switches all at once, even if they were measured separately.
2. Discover hidden rules: Because it can "impute" (guess) missing data, it can find connections between things that were never measured together. For example, it found that specific chemical switches (histones) control specific genes in the brain's "corpus callosum" (the bridge between brain halves), a connection that was hard to see before.
3. Handle huge data: It's fast enough to handle millions of data points, making it possible to map entire organs, not just tiny slices.

In Summary

SpaMosaic is the ultimate data detective. It takes scattered, messy clues from different experiments, cleans them up, translates them into a common language, and uses the spatial context (who lives next to whom) to fill in the missing pieces. The result is a crystal-clear, high-definition map of our biology that helps us understand how our bodies work, how diseases start, and how to treat them.

Technical Summary

1. Problem Statement

Spatial multi-omics technologies (e.g., spatial ATAC-RNA-seq, CUT&Tag-RNA-seq) allow for the simultaneous profiling of multiple molecular layers (transcriptome, epigenome, proteome) within a tissue section. However, acquiring multiple omics from a single section is technically challenging, often involving trade-offs between resolution, sensitivity, and throughput. Consequently, most available data consists of single-omics profiles from different, non-overlapping tissue sections.

Current computational methods face significant limitations in this context:

• Single-cell mosaic integration tools (e.g., Cobolt, scMoMaT, MIDAS) ignore spatial context, leading to poor performance on spatial data.
• Spatial integration tools (e.g., BANKSY, SEDR, GraphST) are designed for "horizontal" integration (same modality, different batches) or "vertical" integration (same section, different modalities) but cannot handle the general case of spatial mosaic integration (different sections, different combinations of modalities, and varying resolutions).
• There is a lack of methods capable of constructing a unified, modality-agnostic latent space that simultaneously corrects batch effects, preserves spatial continuity, and imputes missing modalities across heterogeneous datasets.

2. Methodology: SpaMosaic

SpaMosaic is a deep learning framework designed to integrate heterogeneous spatial multi-omics data into a unified latent space. It employs a contrastive learning framework combined with Graph Neural Networks (GNNs).

Key Technical Components:

• Input Processing:
  • Takes mosaic feature-by-spot count matrices from multiple sections with varying modalities (RNA, ATAC, Histone modifications, Protein).
  • Performs dimensionality reduction (SVD) on highly variable features for each modality.
  • Applies batch correction (Harmony/Seurat) within each modality before integration.
• Heterogeneous Graph Construction:
  • Constructs a spot-spot graph for each modality containing two edge types:
    1. Spatial Edges: Connects nearest spatial neighbors within the same section.
    2. Omics Edges: Connects Mutual Nearest Neighbors (MNNs) across different sections based on batch-corrected feature similarity.
  • Robustness: Uses an Isolation Forest algorithm to filter out high-noise MNN pairs and employs asymmetric neighbor sizing to handle sections with vastly different spot counts.
• Graph Neural Network Encoder:
  • Uses a Weighted Light Graph Convolution Network (LGCN) to encode modality-specific information.
  • The graph is weighted to balance spatial proximity (more reliable) against omics feature similarity (noisier).
• Contrastive Learning Objective:
  • Positive Pairs: Embeddings of different modalities from the same spot (in multi-modal sections).
  • Negative Pairs: Embeddings from different spots (even within the same section/modality).
  • The objective minimizes the distance between positive pairs (aligning modalities) and maximizes the distance between negative pairs (preserving biological variation).
• Feature Reconstruction:
  • For spots with only a single modality, a reconstruction loss is added to ensure the latent embedding retains sufficient information to reconstruct the input features.
• Imputation:
  • Once the modality-agnostic latent space is learned, missing modalities for single-modality sections are imputed using a k-Nearest Neighbor (kNN) approach. The method averages the feature profiles of the nearest neighbors in the latent space from sections where that modality is present.

3. Key Contributions

1. First Spatial Mosaic Integration Framework: SpaMosaic is the first tool specifically designed to handle the "mosaic" scenario where sections have different combinations of omics modalities, different resolutions, and different technologies.
2. Modality-Agnostic Latent Space: It creates a unified embedding space that aligns diverse data types (RNA, ATAC, Histone marks, Protein) without requiring co-profiling on the same section.
3. Robust Imputation: It enables the accurate imputation of missing modalities, allowing researchers to infer relationships between omics layers (e.g., linking RNA to Histone modifications) even when they were never measured together.
4. Scalability: The method is computationally efficient, capable of integrating over 100 sections and processing single sections with >800,000 spots.

4. Results and Benchmarks

The authors benchmarked SpaMosaic against state-of-the-art methods (Cobolt, scMoMaT, StabMap, MIDAS, MultiVI, UINMF, CLUE) and spatial clustering tools (BANKSY, CellCharter) across diverse datasets:

• Simulated Data: SpaMosaic outperformed all competitors in recovering ground-truth spatial factors, reducing noise (lower PAS and CHAOS scores), and achieving superior batch mixing (high iLISI) and modality alignment (high F1 score for label transfer).
• Embryonic Mouse Brain (Misar-seq): Integrated RNA and ATAC across developmental stages (E13.5–E18.5). SpaMosaic successfully identified coherent brain regions (e.g., dorsal pallium, subpallium) with boundaries matching the Allen Brain Atlas, whereas other methods produced noisy or inconsistent clusters.
• Postnatal Mouse Brain (Histone + RNA): Integrated data with section-size imbalances and different histone marks (H3K4me3, H3K27ac). SpaMosaic accurately resolved cortical layers and sub-structures (e.g., corpus callosum, striatum) that competing methods failed to distinguish.
• Cross-Technology Integration: Successfully integrated data from Visium, Visium HD, Stereo-seq, and DBiT-seq with varying resolutions (2µm to 55µm). SpaMosaic maintained spatial coherence and biological consistency where other methods failed due to resolution mismatches.
• Imputation Validation:
  • Imputed histone modification profiles recapitulated expected correlations with RNA (e.g., H3K4me3 positively correlated with gene expression).
  • Discovery: Imputed data revealed more region-specific regulatory links than the measured ATAC data alone. For example, imputed H3K4me3/H3K27ac profiles identified genes involved in oligodendrocyte differentiation and myelination in the corpus callosum that were not detected in the sparse ATAC data.

5. Significance

• Unlocking Heterogeneous Data: SpaMosaic allows researchers to construct comprehensive multi-omics spatial atlases by leveraging the vast amount of existing single-omics spatial data, which is far more abundant than multi-omics data.
• Hypothesis Generation: By enabling the imputation of missing modalities, the tool facilitates the discovery of regulatory mechanisms (e.g., chromatin accessibility driving gene expression) without the need for expensive, technically difficult co-profiling experiments.
• Standardization: It provides a scalable, flexible framework for unifying the rapidly growing landscape of spatial omics data, paving the way for large-scale, multi-tissue, and multi-species spatial atlases.

In summary, SpaMosaic represents a significant advancement in spatial bioinformatics, solving the critical bottleneck of integrating fragmented, heterogeneous spatial datasets to reveal a complete, high-dimensional view of tissue biology.