TL;DR
Researchers have developed scSpace, a novel computational method that reconstructs the spatial organization of cells from single-cell RNA sequencing (scRNA-seq) data by leveraging existing spatial transcriptomics (ST) datasets. This allows for the identification of spatially distinct cell subpopulations, offering new insights into tissue architecture and disease mechanisms, particularly in complex conditions like cancer and COVID-19.
What Problem Does This Paper Address?
Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular diversity by providing high-resolution gene expression profiles for individual cells. However, a major drawback of scRNA-seq is the complete loss of spatial information during the tissue dissociation process. This spatial context is crucial, as cell location and neighborhood often dictate cell function and identity, especially in complex tissues like the brain, liver, and tumors. While emerging spatial transcriptomics technologies can capture spatial information, they often have limitations in unbiased transcriptome profiling of individual cells. Therefore, there's a critical need for computational methods that can integrate scRNA-seq data with spatial information from ST references to reconstruct cellular spatial organization and identify spatially heterogeneous cell subpopulations.
Methods
scSpace employs a three-component approach:
- Latent Biological Feature Representation Extraction: This step uses a transfer learning model called Transfer Component Analysis (TCA) to eliminate batch effects between scRNA-seq and ST data and extract shared biological features. TCA projects data into a Reproducing Kernel Hilbert Space (RKHS) and minimizes the Maximum Mean Discrepancy (MMD) between the distributions of the two data types.
- Spatial Reconstruction (Pseudo-space): A multi-layer perceptron (MLP) is trained using the shared latent features from ST data, with spatial coordinates as the outcome. This trained model is then applied to the latent features of scRNA-seq data to generate a "pseudo-space," effectively reconstructing the spatial arrangement of individual cells.
- Space-informed Clustering: This component identifies spatially variable cell subpopulations. It constructs a gene expression graph from scRNA-seq data and incorporates a spatial weight for each edge based on the cell distances in the generated pseudo-space. This weighted graph is then used for unsupervised clustering (e.g., using the Leiden algorithm), allowing the method to consider both gene expression and spatial proximity.
Datasets
The authors evaluated scSpace using:
- Simulated datasets: 140 simulated datasets were generated using the Splatter R package to mimic scRNA-seq and ST data with varying numbers of spatially heterogeneous cell populations.
- Biological datasets:
- Human dorsolateral prefrontal cortex (DLPFC) (10X Visium)
- Mouse primary visual cortex (V1) (STARmap)
- Human middle temporal gyrus (MTG) (snRNA-seq)
- Human embryonic heart (scRNA-seq and ST)
- Kidney (scRNA-seq and ST)
- Small intestinal villus (scRNA-seq and ST)
- Liver lobule (scRNA-seq and ST)
- Human skin squamous cell carcinoma (SCC) (scRNA-seq and ST)
- Human HER2 breast cancer (scRNA-seq and ST)
- Human lung (normal and COVID-19) (scRNA-seq and ST)
- Mouse cerebral cortex (scATAC-seq)
Key Insights and Ideas
- Bridging the scRNA-seq and ST gap: scSpace effectively bridges the information gap between scRNA-seq (lacking spatial data) and ST (lacking comprehensive single-cell resolution) by using ST as a reference to infer spatial locations for scRNA-seq data.
- "Pseudo-space" as a powerful concept: The creation of a "pseudo-space" allows for the visualization and analysis of cellular spatial relationships directly from scRNA-seq data, enabling the discovery of spatially organized subpopulations.
- Spatial context is critical for subpopulation identification: The paper strongly demonstrates that incorporating spatial information significantly improves the ability to distinguish cell subpopulations, especially those with subtle transcriptional differences but distinct spatial locations.
- Robustness across diverse tissues: scSpace shows promising performance across a wide range of complex tissues, including brain, heart, kidney, intestine, liver, cancer, and lung, highlighting its general applicability.
- Therapeutic marker discovery: The application of scSpace to melanoma and COVID-19 datasets revealed spatially localized cell subpopulations with potential roles in disease progression and therapeutic targets, underscoring its clinical relevance.
Key Results
- Spatial Reconstruction Accuracy: scSpace achieved a Pearson correlation coefficient (PCC) of pairwise distances between cells in the pseudo-space and original space above 0.9 on simulated datasets.
- Clustering Performance: On 140 simulated datasets, scSpace significantly outperformed other methods (Louvain, K-means, Hclust, SpaGCN, STAGATE, BayesSpace, DR-SC, scCoGAPS) in identifying spatially variable cell subpopulations, with consistently higher Adjusted Rand Index (ARI) scores.
- Brain Cortex Reconstruction: scSpace successfully reconstructed the hierarchical structure of human DLPFC layers and mouse V1 neocortex, preserving relative layer positions and neighbor relationships.
- Cardiomyocyte Subpopulation Identification: In embryonic heart data, scSpace accurately distinguished atrial, ventricular, and Myoz2-enriched cardiomyocytes, achieving ARI and NMI scores of 0.91 and 0.85, respectively, significantly outperforming Seurat.
- COVID-19 Myeloid Cell Invasion: scSpace captured the reduced distance between myeloid cells and alveolar/airway epithelial cells in COVID-19 patients, indicating immune infiltration, and identified a specific subpopulation (C4) with high oxidative phosphorylation and MHC class II pathway activity.
Limitations and Future Work
- Tissue Coherence Dependence: scSpace's performance can be influenced by the spatial coherence of the tissue. Tissues with low spatial coherence or significant heterogeneity between biological replicates (e.g., certain cancer patients) may lead to less accurate reconstruction.
- Spatial Reference Appropriateness: The choice and quality of the spatial reference dataset are crucial. The authors suggest assessing the appropriateness of the spatial reference and considering pre-processing steps for tissues with heterogeneous architectures.
- Extension to Other Omics: The authors explored scSpace with scATAC-seq data, demonstrating its potential applicability to other single-cell omics modalities, provided suitable spatial references are available. This suggests future work could focus on developing and validating scSpace for multi-modal spatial analysis.
- Computational Resources: While scSpace is efficient (processing 50,000 cells in ~20 minutes on a GPU), complex datasets or very large numbers of cells might still pose computational challenges.
Summary updated