TL;DR
This paper introduces SpatialLeiden, a novel method that adapts the popular Leiden clustering algorithm to incorporate spatial information, making it "spatially aware." This allows for more accurate identification of cell types and tissue structures in spatial transcriptomics data, outperforming existing methods in both accuracy and computational efficiency.
What Problem Does This Paper Address?
Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity by allowing researchers to measure gene expression at the individual cell level. Clustering algorithms, particularly Leiden clustering, are crucial for grouping cells with similar expression profiles to identify distinct cell types and biological insights. However, traditional Leiden clustering is "non-spatial," meaning it only considers gene expression similarity and ignores the physical location of cells within a tissue. This is a significant limitation for spatial transcriptomics, where the spatial arrangement of cells is critical for understanding tissue organization, cell-cell interactions, and the formation of spatial domains (like anatomical regions or cell niches). Existing methods that try to incorporate spatial information often do so by using non-spatial clustering on spatially derived features, or by using specialized spatial clustering algorithms that may not be as performant or widely adopted as Leiden. This paper aims to bridge this gap by making Leiden itself spatially aware.
Methods
The core innovation of SpatialLeiden lies in its integration of spatial information directly into the Leiden clustering framework. This is achieved through a "Leiden multiplex" approach, which allows for multiple layers of information to be considered during clustering.
Here's a breakdown of their approach:
- Spatially Variable Genes (SVGs): Instead of relying on highly variable genes (HVGs) which are common in non-spatial scRNA-seq, SpatialLeiden prioritizes SVGs. These are genes whose expression patterns are not random but show spatial correlation within the tissue.
- Spatially-Aware Dimensionality Reduction (msPCA): To capture spatial relationships, they employ MULTISPATI-PCA (msPCA). This method takes into account the spatial arrangement of cells when reducing the dimensionality of the gene expression data, preserving spatial patterns that might be lost in traditional PCA.
- Leiden Multiplex Clustering (SpatialLeiden): The key step is using the Leiden multiplex functionality. This allows the algorithm to consider two "layers" simultaneously:
- Gene Expression Layer: This is the standard latent space derived from gene expression data (using PCA or msPCA).
- Spatial Layer: This layer encodes the spatial relationships between cells. This is achieved by defining a "spatial connectivity" graph where edges represent the proximity of cells, often based on Euclidean distance or a k-nearest neighbor approach.
By combining these layers, SpatialLeiden can cluster cells based on both their molecular similarity and their spatial proximity, effectively making Leiden "spatially aware."
Datasets
The paper evaluates SpatialLeiden using several publicly available spatial transcriptomics datasets:
- Mouse Dorsolateral Prefrontal Cortex (DLPFC): This is a widely used benchmark dataset from 10x Visium technology, consisting of spatial gene expression data and histology images from multiple mouse donors. It includes ground truth annotations of anatomical domains, allowing for quantitative evaluation.
- Stereo-Seq: Data from mouse embryos at various developmental stages.
- BaristaSeq: Mouse brain primary cortex data.
- MERFISH: Mouse brain hypothalamus preoptic area data.
- osmFISH: Mouse brain somatosensory cortex data.
- STARmap & STARmap*: Mouse brain medial prefrontal cortex and visual cortex data, respectively.
The authors also reference performance metrics from a previous benchmark study (Yuan et al. 2024) for comparison.
Key Insights and Ideas
- Leiden can be spatially aware: The most significant insight is that the widely used Leiden algorithm, when augmented with spatial information, can become a powerful tool for spatial analysis, challenging the notion that it's inherently non-spatial.
- Spatial embeddings are crucial: Integrating spatial relationships directly into the clustering process, rather than just using spatially derived features, leads to a more accurate representation of tissue structures.
- SVGs over HVGs: Prioritizing spatially variable genes over general highly variable genes is a more effective strategy for capturing biologically relevant spatial patterns.
- Computational efficiency: SpatialLeiden achieves state-of-the-art performance while being significantly faster than many other spatial clustering methods, making it more practical for large datasets.
Key Results
- Improved accuracy: SpatialLeiden demonstrated substantial improvements in representing tissue structures, such as the neocortex layering pattern, compared to non-spatial Leiden.
- Competitive performance: Quantitatively, SpatialLeiden achieved high scores on Adjusted Rand Index (ARI) and Normalised Mutual Information (NMI), outperforming some specialized spatial domain detection tools like SpaGCN and performing comparably to BayesSpace.
- Speed advantage: SpatialLeiden processed datasets significantly faster than other methods, often at a fraction of the processing time, while using less RAM. For example, non-Stereo-seq datasets were processed in under 2 minutes with less than 400 MB of RAM.
- Broad applicability: SpatialLeiden showed strong performance across various spatial transcriptomics technologies (Visium, Stereo-seq, MERFISH, osmFISH, STARmap) and tissue types, ranking as a top-performing method for several.
Limitations and Future Work
- Parameter tuning: The authors acknowledge that, like other multi-modal clustering approaches, careful consideration and tuning of parameters (e.g., neighbor definitions, weighting of modalities, resolution) are necessary for optimal performance. They provide guidance on how these parameters can be set.
- Topological layer definition: The choice of how to define the spatial neighborhood (e.g., grid, Delaunay triangulation, k-nearest neighbors) can impact results, and the paper suggests that k-nearest neighbors generally performed better for imaging-based methods.
- Future directions: The authors express confidence that SpatialLeiden will become a go-to method for spatial data analysis due to its efficiency, simplicity, and ease of integration into existing workflows. They also suggest that future studies comparing spatial clustering algorithms should clearly distinguish between spatial and non-spatial implementations of methods like Leiden.
Summary updated