SLGCA: spatial cross-level graph contrastive autoencoder for multislice spatial domain identification and microenvironment exploration

Xin Lu, Murong Zhou, Guohua Wang, Qiaoming Liu, Yuming ZhaoBriefings in Bioinformatics 2025 · DOI

TL;DR

This paper introduces SLGCA, a novel graph-based deep learning framework that accurately identifies spatial domains within tissue samples, even across multiple sections without prior alignment. By integrating local and global spatial information through a unique contrastive learning approach, SLGCA offers a more robust and comprehensive understanding of tissue microenvironments, crucial for disease research.

What Problem Does This Paper Address?

Understanding the spatial organization of cells within tissues is fundamental to deciphering biological processes and disease mechanisms. Spatial transcriptomics (ST) technologies provide this crucial information, but analyzing the resulting data presents significant challenges. Existing methods for identifying "spatial domains" (regions with distinct cellular compositions and gene expression patterns) often struggle to fully leverage spatial information, leading to inaccurate or fragmented results. Furthermore, integrating data from multiple tissue slices (multislice analysis) is essential for a complete picture but is often hampered by batch effects and the need for complex pre-alignment steps. This paper aims to address these limitations by developing a method that can accurately identify spatial domains within and across multiple tissue slices, effectively integrating diverse spatial information without requiring manual preprocessing.

Methods

SLGCA employs a graph convolutional autoencoder (GCN AE) framework. The core innovation lies in its cross-level graph contrastive learning strategy, which operates on two levels:

  • Local-level Contrastive Learning: This module focuses on learning from the immediate spatial neighborhood of cells. It treats spatially adjacent cells as positive pairs (encouraging their embeddings to be similar) and distant cells as negative pairs (driving their embeddings apart). This is achieved by creating two augmented views of the graph and enforcing similarity between corresponding nodes and their neighbors across these views.
  • Global-level Contrastive Learning: This module captures broader spatial relationships by aggregating local information into global embeddings. It then enforces similarity between corresponding feature dimensions across these global embeddings from different views, helping to suppress redundant correlations and improve the overall discriminative power of the learned representations.

The model is trained using a combined loss function that includes reconstruction of gene expression and adjacency matrices, along with the local and global contrastive losses. For multislice analysis, SLGCA constructs a block-diagonal adjacency matrix, allowing it to integrate data from different slices without explicit alignment.

Datasets

The study evaluated SLGCA on a variety of publicly available ST datasets, including:

  • Human dorsolateral prefrontal cortex (DLPFC) dataset (10x Visium platform)
  • Mouse anterior brain dataset (Allen Brain Atlas reference)
  • Mouse primary visual cortex (V1) dataset (STARmap platform)
  • Mouse somatosensory cortex dataset (osmFISH platform)
  • Mouse olfactory bulb dataset (Stereo-seq platform)
  • Human breast cancer dataset
  • Human liver cancer dataset

The paper also provides links to the original data and the source code for SLGCA on GitHub.

Key Insights and Ideas

  • Bridging Local and Global Spatial Information: SLGCA's dual-level contrastive learning is key. It's not just about what's next to a cell, but also how that cell fits into the larger tissue architecture. This holistic approach is crucial for accurate domain identification.
  • End-to-End Multislice Integration: The ability to integrate multiple tissue slices without manual alignment is a significant workflow improvement. This bypasses a major bottleneck in current ST analysis.
  • Uncovering Subtle Heterogeneity: The method successfully identified distinct subtypes of fibroblasts in liver cancer and nuanced molecular differences in breast cancer subtypes, demonstrating its power in exploring complex microenvironments.
  • Robustness Across Platforms: SLGCA performs well across diverse ST technologies (10x Visium, STARmap, osmFISH, Stereo-seq), highlighting its generalizability.

Key Results

  • SLGCA consistently outperformed 10 state-of-the-art methods in spatial domain identification accuracy across multiple datasets, achieving the highest median ARI and NMI scores.
  • On the DLPFC dataset, SLGCA achieved a median ARI of 0.65 and NMI of 0.72.
  • For multislice analysis, SLGCA demonstrated effective integration without prior alignment, achieving an ILISI score of 1.64 on two vertical slices and 1.13 on horizontal slices.
  • In breast cancer analysis, SLGCA accurately distinguished between ductal/lobular carcinoma in situ (DCIS/LCIS) and invasive ductal carcinoma (IDC), revealing distinct molecular features.
  • In liver cancer, SLGCA identified two distinct fibroblast subtypes with unique functional enrichments, providing new insights into the tumor microenvironment.

Limitations and Future Work

  • Computational Cost: The paper acknowledges that incorporating spatial neighborhood information in local contrastive learning increases computational and memory costs, especially for large-scale datasets. Future work could focus on optimizing these aspects.
  • Integration of Multi-omics: With the rise of multi-omics technologies, there's a growing need for analytical frameworks that can integrate multiple data layers within the same spatial context. The authors suggest this will be a key area for future research.

Summary updated

BibTeX
@article{lu2025slgca,
  author = {Xin Lu and Murong Zhou and Guohua Wang and Qiaoming Liu and Yuming Zhao},
  title = {{SLGCA: spatial cross-level graph contrastive autoencoder for multislice spatial domain identification and microenvironment exploration}},
  journal = {Briefings in Bioinformatics},
  year = {2025},
  doi = {10.1093/bib/bbaf574},
}
Figure